Borrowed Trust

Abstract

Abstract

What native advertising is in 2026, who earns from it, and how much weight the evidence behind its sales pitch will bear. Research current to 27 September 2026, with the United States as the main market.

The question

Native advertising is sold as performance. This paper asks what the public record shows behind that pitch. How big is the market, and by whose fence? Who controls it? What does a native dollar pay for? Do readers know they are looking at an ad? Has anyone shown that native causes sales? And what are agents and AI answers changing? The scope is the open web. Social feeds, search, retail media and AI assistants are treated as adjacent systems that compete for the same budget.

How it was done

The work ran as twelve research workstreams, a vendor-universe sweep and nine vendor profiles. An AI research agent ran each one from a shared brief. A second AI agent, the verifier, then checked it, with orders to refute it. In all, 977 sources were opened and logged. The research ledger holds 642 claims, and each number in the text traces to a source. The paper scores 24 vendors on 12 dimensions. The scoring rules were written before the first scoring pass. A revision followed once the verifier records were read, and the Native Quotient summary was added after scoring, at the author's request. A separate review of an interim draft raised one hundred findings, and a second outside review of the rebuilt package raised 38 more. Each was checked against its cited source where one could be reached; one was rejected. Each change is in the corrections ledger in chapter 12. This is AI-assisted research with a published process. It has not been audited or peer reviewed. No vendor saw it before it was published.

What it found

1. The market has no measured size. eMarketer puts US native display at $147.98 billion in 2026. That figure counts social feeds. They took 74% to 84% of it in each vintage whose split could be read. The industry's own revenue accounts have no native line at all ◆ ◆. The two large US-listed open-web platforms reported $3.21 billion of revenue in 2025, about 1.6% of Meta's ad revenue ◆.
2. It is a supply-rights business. Taboola's ten largest advertisers are under 10% of its revenue. Yahoo and Microsoft make up about 34% of its revenue from advertisers on publisher properties [1]. Teads Holding Co. discloses no partner at 10% of its supply cost, so this finding rests on Taboola's record. Traffic acquisition cost is its recorded cost of supply, mostly payments to publishers. Its cost of guarantees to publishers was about 18% of that cost in 2024 and 15% in 2025, up from 9% to 10% in 2021 and 2022 ◆.
3. Many readers do not recognise it. In three studies of sponsored articles and in-feed ads, 7% to 37% of those tested saw them as advertising. In one of those studies, 81% did so for standard ads. In the FTC's lab, about 47% did, and 68% with its recommended labels. Only the FTC study tested a recommendation widget. Better labels helped but did not close the gap ◆ ◆. These are dated results for certain formats and samples. They are not a 2026 population rate.
4. Conversion optimisation does not change the billing basis. A conversion is the action an ad is meant to drive, such as a sale. In the per-click campaigns the Taboola, Outbrain and MGID help centres document, the platform's model decides which clicks to buy and the advertiser still pays for every click. The financial risk stays with the advertiser ◆. A target cost per acquisition is a bidding setting, not per-acquisition billing. Both large platforms' filings describe per-click, per-thousand and per-acquisition pricing, and Teads prices some campaigns on incremental action. Neither discloses the share of revenue under each.
5. The causal evidence is missing, not negative. In the sources screened by the cutoff, this review found no public study with enough published design and results to estimate what open-web recommendation and in-feed units add to sales. Vendor evidence compares those who saw the ads with those who did not. In Facebook conversion experiments, observational designs like that overstated lift by about three to thirteen times at the median. This review found no published estimate of that bias for native brand-lift surveys ◆ ◆. Randomised tests of native formats elsewhere, in Facebook campaigns and in native search ads, found positive effects. The gap is specific to the open-web widget.
6. The platforms are building beyond the widget. Taboola sells display, video and ads inside AI answers through Realize and DeeperDive, and Teads Holding Co.'s connected TV reached 13% of revenue in Q2 2026. Neither discloses how much growth those lines add. In that quarter, revenue billed through Yahoo grew faster than Taboola's total ◆. Several publisher traffic and ad-volume indicators are falling. They measure sessions, ad requests and impressions, not one page-view series, and a shrinking supply of widget slots is an inference from them ◆.
7. Buyers can inspect less than they can buy. On both large native self-serve platforms, site lists are documented only after delivery. Controls are exclusion-first ◆. A regional platform, Readpeak, shows a pre-campaign list can be done.
8. Agents reach native through the platforms' own doors. Neither agentic protocol stack lists a large native platform as an implementer on the member pages opened. Taboola instead runs its own server for AI agents to manage campaigns ◆.

What would change these findings

A holdout study of native on its own, with published design and results. A holdout keeps the ads from one group, so the gap shows what the ads caused. A platform that discloses the actual lag behind its reported conversions. A site list shown before the campaign at a large platform. Recommendation revenue reported on its own and shown to be growing. Chapter 11 gives each scenario a time window to watch and the evidence that would count against it.

Limits

The open-web native market has no measured size, and this paper does not invent one. Its economics are those of the recommendation and in-feed networks; sponsored editorial sold directly by publisher studios, creator and influencer content, and in-mail native were not measured. The buyer guidance is written for the United States, while the platforms earn most of their revenue elsewhere; Asia-Pacific markets and their rules were not covered. The financial comparisons rest on measures the companies define. The recognition studies are old. The vendor scores come from one AI-assisted scoring pass, with no second scorer. A company can score low for being private rather than weak. Access limits are logged for each source, and chapter 12 names the ones that matter.

How to read it

Chapters 1 and 2 define and size the market. Chapters 3 and 4 follow the money and the supply contracts. Chapters 5 to 7 test the evidence on readers, delivery and performance. Chapter 8 covers commerce and AI answers. Chapter 9 scores the vendors. Chapter 10 covers enterprise buying and the rules. Chapter 11 sets out scenarios with signs to watch. Chapter 12 is the method and the ledgers, and the appendix lists each download. Each chapter opens with a short summary. In the web edition, the Skim control cuts the paper down to those summaries, the figures and the instruments.

Interests to declare

The author has worked at Samba TV since May 2025. Samba TV is a TV measurement and ad data company. Before that he worked at Verve Group, where he ran the Moments.AI business. Earlier he worked at Hearts & Science and Lotame. He co-leads the Signals & Measurement working group of AgenticAdvertising.org. That body governs AdCP, a protocol discussed in chapter 11. Samba TV is listed among AdCP's members, and Teads names Samba TV among its measurement-partner integrations; Teads' connected-TV business, discussed in chapters 4 and 11, overlaps Samba TV's market. None of the author's current or former employers listed above is scored. Teads is scored and profiled; its commercial relationship with Samba TV is disclosed here. No one funded or commissioned the paper, and no vendor reviewed it.

Chapter 1

Defining the market

"Native" names how an ad looks and where it sits. Arguments over whether the category is growing, working or worth buying almost all turn on one point. That point is which other axes are silently included.

Each source this paper opened defines native advertising the same way in one sentence. In the tables, each one differs. That one sentence comes from the FTC. Native is content that bears a similarity to the news, feature articles, product reviews, entertainment and other material that surrounds it online [2]. IAB's Native Advertising Playbook 2.0 was published in May 2019. As far as this research found, nothing has replaced it. It sorts that into three types: in-feed and in-content units, content recommendation units, and branded or native content. It requires a disclosure on all three [3] ◆. OpenRTB Native Ads 1.2 is the programmatic standard that carries native in the bid stream. It defines native as a set of components: a title, an image, a description and a link. The publisher renders them in its own style [4] ◆. Those three are not the same. The FTC's definition turns on resemblance, IAB's on format and OpenRTB's on assets. An editorially produced sponsored article need not be built from ad components at all. This paper's definition fits all three. Native describes advertising designed to fit its surrounding content or experience. Programmatic native commonly achieves that through publisher-rendered assets. On all else, the sources disagree or say nothing. That is why the paper needs six axes, not one word.

Six axes, and which two the word covers

The format axis is what the unit looks like. It can be a recommendation unit, a block of suggested links at the end of an article. It can be an in-feed unit between posts. It can also be a sponsored article, a native video player or a product card. The placement surface is where the ad appears. That can be an open-web publisher page, a social feed or a search results page. It can also be a retailer's site, an AI assistant's answer or a newsletter. The word "native" describes those two axes and nothing else.

The targeting method is why a given reader sees the ad. It can rest on contextual signals from the page, or on audience or behavioural data. It can use retargeting from a past visit or lookalike modelling. Or it can rest on nothing at all. The buying route is how the ad is bought. It can be a self-serve platform or a managed service under an insertion order. It can be programmatic through a DSP. Since 2025, it can also be an agent that talks to a seller's agent. The optimisation objective is what the delivery system chases. That can be impressions, clicks, conversions, return on spend or an incremental outcome. A conversion is an action the advertiser wants, such as a sale. An incremental outcome is one that would not have happened without the ad. Last, the commercial model is who pays for what. It can be cost per click, cost per thousand or cost per action. It can also be a revenue share to the publisher, a minimum guarantee or a software subscription.

The axes apply to a placement or a line item. A campaign can mix formats and surfaces and use more than one route. So a campaign can carry more than one value on an axis. A single placement usually takes one value per axis, though some axes, targeting in particular, can hold more than one. Take a recommendation unit on a news site, bought self-serve and paid per click. It has a conversion objective and no targeting beyond the platform's own model. It is a native ad. So is a sponsored product card in a social feed. It is bought programmatically, paid per thousand impressions and uses audience targeting. Both are native in the broad sense, but they take different values on every axis the example names, format and surface included. A market estimate that adds them together is adding a click business to an impression business.

Native is not contextual, and the confusion has a cause

The sources most often confuse native with contextual targeting. The cause is historical. The recommendation platforms that dominate open-web native rank ads against the page and the reader. So their delivery is contextual in the loose sense. And vendors that sell contextual targeting as a data product often deliver it into native units. But the two sit on different axes. Contextual targeting is a way of choosing the reader. It can apply to a display banner, a video ad or a native card. It describes the basis for the choice, not a promise that no personal data is used. A system can read the page and also use a device ID. Native is a way of dressing the ad. It can be targeted by context, by behaviour or not at all. This series has a companion paper on contextual advertising. It treats contextual as a targeting method. Its transaction layer is a taxonomy ID [5]. This paper treats native as a format and placement. Its transaction layer is a component list. As described in chapter 6, the recommendation engine is where the two meet. It uses page context as one input among several to a click-and-conversion model. That engine is not contextual advertising in the data-product sense. It is a ranking system that happens to read the page.

Three scope groups, one counting rule

The core market is paid native placements. It also takes in the platforms or publisher businesses that sell, deliver or optimise them. It covers recommendation units, in-feed native on open-web pages and sponsored editorial. It also covers native video or commerce units where they are sold as native. Chapters 3 and 4 measure this market through filings. Chapter 9 scores it.

Budget substitutes and adjacent systems compete for the same buyer's money. Or they supply something the core depends on. They are social feeds, search, retail media, DSPs, content commerce and affiliate networks. A social in-feed unit is native by format, which is why some estimates count it. Here it sits in the adjacent group. That is because its buyer, auction, measurement and rules belong to the platform. They do not belong to the open-web native market. When a figure in this paper includes a social or retail-media number, the paper says so. It also says what share is native.

Emerging surfaces are places where the native definition may apply. But the market there cannot yet be measured. They are mainly sponsored answers or recommendations inside AI assistants and AI answer engines on publisher sites. Buying through agents is a buying route, not a surface. It sits on that axis. Chapter 8 sorts the surfaces into live products, trials and announcements, and keeps them apart. Chapter 11 returns to what they and the agents change.

The counting rule across all three is paid promotion only. So this paper does not count an organic recommendation from the publisher's own engine. Nor does it count an editorial link, or a product listing a retailer ranks by relevance. It also leaves out an affiliate link the publisher chose to insert. But that exclusion is about counting, not about disclosure. An affiliate link carries a material commercial connection. Chapter 8 sets out the disclosure rules that apply to it. Where a boundary is contested, the paper says which side it took and why. The three cases that matter most are set out below. Four more rulings are brief. Creator and influencer sponsored posts, and podcast host-reads, are native by resemblance but sold by creators and platforms; they are adjacent here and not sized. Connected-TV home-screen tiles and pause ads are adjacent formats of the TV interface. In-app native and newsletter or in-mail native are core surfaces, but the paper documents them only in passing.

Three numbers that are not the same number

Market size is what was spent on paid native placements in a given period and place. Addressable opportunity is what could be spent if each budget that could move into native did so. What one company expects to win is a third question. The closest public answer is its own revenue guidance. That is a management forecast, not an estimate of what can be captured. The playbook that seeded this paper uses all three. So do the vendor filings in chapter 4 and the forecasts in chapter 2. Sometimes all three appear in one sentence. Take Taboola's 2025 annual report. It claims a $55 billion opportunity for its performance platform. It bases the figure on internal and external industry data [1] ◆. That is an addressable claim, not a market size. And it covers performance advertising broadly, not native. The paper keeps market size, addressable opportunity and company guidance apart. Each time it uses a number, it names which of the three it is.

How the definition changes the answer

Chapter 2 makes the result concrete. The same word gives estimates that differ by an order of magnitude. The gap depends on whether social in-feed is inside the fence. Chapter 3 shows that the firms called native platforms now sell more and more placements that are not native at all. Chapter 5 looks at whether readers can tell the unit from the content. That is the property that defines the format. In several studies of certain formats and disclosures, their ability to do so was low. Chapter 6 shows that the platforms now compete on the optimisation objective, not the format. And chapter 11 shows that the two listed native companies are moving along different axes at once. One is moving toward display and connected TV. The other is being pulled by the page-view economics of the format it is named after. None of those findings shows up if native is treated as one thing.

Chapter 2

Follow the spend

In the sources this project reviewed, only one public spend estimate carries the word "native". In the past, most of it was social feeds. Everything else has to be rebuilt from the platforms' own accounts.

The research brief asked how much reported native growth comes from a wider category, from adding social, or from price or mix. It set that against growth in open-web native itself. An honest answer means taking each estimate apart. This chapter does that before it quotes any of them.

What the word covers in each source

IAB's Native Advertising Playbook 2.0 cut native to three types in May 2019. They are in-feed and in-content units, content recommendation units, and branded or native content. The 2013 playbook had six types [3, 6] ◆. Content recommendation units sit in the blocks of suggested links on publisher pages. The new playbook removed paid search, "even though search ads technically meet the IAB definition of native". It folded promoted listings on commerce sites into the in-feed type ◆. It also dropped "buying and selling" and "measurement" as tests of whether an ad is native. Its reason was that programmatic buying no longer defines what is native [3] ◆. In the playbook, in-feed native covers three surfaces: content feeds, product feeds and social feeds [3] ◆. The playbook is broad by design. The sources that measure spend are narrower. Each one narrows it in its own way.

IAB and PwC's report on US internet ad revenue for 2025 puts the total at $294.6 billion, up 13.9%. It does not report native. Native is buried inside display, at $81.6 billion. Display "includes banner, rich media, sponsorship and native revenues" [7, 8] ◆. Social media, at $117.7 billion and up 32.6%, is reported next to the formats, not as one of them. The five formats already add up to the total. So social revenue sits inside display and video [7] ◆. The figures are US-based earned revenue. They come from a PwC survey of sellers. Public data and estimates fill in for sellers that did not take part. PwC states it has not audited them [7] ◆.

EMARKETER's US "native display" is the only explicit native estimate in the source set reviewed. Other syndicated research may exist that this project could not open. Some may fold native into wider categories. Neither kind is compared here. Each edition of the EMARKETER forecast is a vintage. The December 2025 vintage is the latest this project could open. It forecasts $147.98 billion for 2026, up 13.1%. That implies about $130.8 billion for 2025. An AI-assisted EMARKETER FAQ lists in-feed units, sponsored content, content recommendation widgets, in-app rewarded video and sponsored outstream video as native formats, and excludes banners and pop-ups. The forecast's own published scope says only display ads that reflect the form and function of their surroundings, so whether it counts the video formats is not stated [9] ◆. Its separate digital video line counts outstream and in-feed video. So the two totals may overlap and should not be added [10] ◆. EMARKETER also says native's share of total US display has fallen "as connected TV and retail media expand". So under its own definition, native grows more slowly than display [9] ◆.

None of the agency forecasters publishes a native line. MAGNA's June 2025 global forecast has none ◆. Nor do WPP Media's June 2026 update and dentsu's December 2025 forecast ◆. IAB Europe's 2025 benchmark for 30 markets has none either ◆. All four report search, social, retail media and video [11, 12, 13, 14, 15, 16]. The Tech Lab's OpenRTB Native standards page still says the API covers the six types of the 2013 playbook. The 2019 playbook replaced them [17, 3] ◆.

The definitional gulf, in one ratio

EMARKETER's implied 2025 native display figure is about $130.8 billion. That is about 1.6 times IAB/PwC's whole 2025 display total of $81.6 billion. By IAB's own account, that total includes native. The two numbers do not conflict. They fence different fields. EMARKETER counts social in-feed as native, and its FAQ lists outstream video too. IAB counts social as a channel that cuts across formats. It counts in-feed video under video. A reader may quote "$148 billion of native advertising". If that reader does not add that most of it is Facebook, Instagram and TikTok feeds, they have changed the subject.

How much is social? The last published split is from the 2019 vintage. Social was 77% of native display in 2018, falling to 74% by 2020 ◆. Native was expected to be 95.6% of social display spend. For non-social display spend, the expected share was only 30.8% [18, 19, 20, 21] ◆. A search-indexed eMarketer page appears to put 2023 social native at $68.17 billion. That would be about 70% of that year's total. But the page could not be opened. So the number is a lead, and it is not used. In 2017 the social share was 84.2% [20]. Read plainly, "native display" as measured was mostly social-feed ads in each vintage that published a split. In 2018–2020, the non-social part was about a quarter of the total. That part also includes in-app ad space outside this paper's core. Those shares describe those vintages. They are not a measured 2026 split. They are not used here to estimate one. The public eMarketer chart from January 2023 fills most of the gap. It shows $47.3 billion for 2019, $58.7 billion for 2020, $80.6 billion for 2021, $87.0 billion for 2022 and $97.5 billion for 2023. In that chart, native was 60% to 63% of US display spending [22]. No public value for 2024 was found. Any current split sits behind a paywall that returned access errors to this project. The series is drawn as bars by vintage, not as one line. That is because eMarketer revises past years from one vintage to the next. The 2019 vintage revised 2018 up by about 7%. The 2023 vintage raised 2019 from $43.9 billion to $47.3 billion [18, 22] ◆.

Secondary sources drift from the primary. One trade report turned "native is 95.6% of social display" into "96% of native ad spending goes to social". That is the same two numbers the wrong way round. Another gives two different figures for the same MAGNA forecast [19]. Each number in this chapter comes from a source this project opened. The register records which.

What the industry's own accounts show

IAB/PwC's format series is a structured industry revenue survey, filled out with estimates. It is not an audit, but it is the most consistent view of where the money went. It is unkind to native's home. Display grew 9.8% in 2025, from $74.3 billion to $81.6 billion. That was the slowest of the four major named formats. The leftover "other" category grew about 6.9%. Video grew 25.4% to $78.0 billion, and search grew 11.0% to $114.2 billion. From 2021 to 2025 display's compound growth was about 9.5% a year. Video's was 18.5% [7] ◆. Social's 32.6% growth to $117.7 billion sits inside those formats. Commerce media, another line that cuts across formats, reached $63.4 billion, up 18% [7]. Display's growth rate cannot pick out recommendation widgets. Because social cuts across the formats, display itself includes social units. What the series shows is narrower. The format that holds open-web native is losing share to video.

The budget destinations, on their own terms

The brief also asked how native compares with Meta and other places budgets go. It asked for clearly defined terms. The terms are the problem, so the figure states them. Meta's 2025 ad revenue was $196.2 billion, up 22%. Impressions rose 12% and average price per ad rose 9%. Those compound to about 22% [23] ◆. Google's ad revenue was $294.7 billion. Search grew 13.4% to $224.5 billion, and YouTube grew 11.7%. Google Network is the line that earns money from other people's pages. It fell 1.9% to $29.8 billion, its third annual decline in a row. It is now 10.1% of Google's ad revenue. In 2023, it was 13.2% [24, 25] ◆. Network combines AdSense on the web, AdMob in apps and Ad Manager. The filing says the decline came from AdSense, partly offset by AdMob. So the web piece fell by more than 1.9% [26]. Amazon's advertising services add up to $68.6 billion across the four reported quarters of 2025 [27] ◆.

Against those, Taboola's $1,912 million of revenue and Teads Holding Co.'s $1,300.5 million are small. They must be compared with care. Both recognise most revenue gross, as principal. That is, they book what buyers pay as revenue. They record their payments to publishers as traffic acquisition cost. Google also reports its network revenue gross. The share it pays to partners sits in cost of revenues. Meta, by contrast, sells mostly its own ad space [28, 29, 24]. The two platforms reported $3,212.5 million of revenue between them. Their ex-TAC gross profit, as each firm defines it, totals $1,243.2 million. Revenue minus reported traffic acquisition cost totals $1,226.8 million. The gap is Taboola's non-cash add-back. On revenue, they are about 1.6% of Meta's ad revenue. On ex-TAC, they are about 0.6% [28, 29, 23] ◆. The world's largest open-web native firms are a rounding error in the budgets they compete for.

Google Network's decline is the most suggestive number here. But the paper does not lean on it. Network is not a native-only line. It includes app and display ads sold through AdSense and Ad Manager. The decline fits with budgets pooling on owned surfaces, but it does not prove that. Like the recommendation platforms, Network depends on third-party page views. Chapter 11 shows those falling.

Answering the growth question

On the evidence above, reported native growth is hard to trace to open-web native. The one native series counts in-feed units on all surfaces, social included. Three-quarters or more of it was social feeds in the vintages that published a split. The industry's own accounts put native inside display. Display includes social units, and it grew more slowly than video. Growth in open-web native itself shows up only in company filings. In 2025, Taboola's reported ex-TAC gross profit grew 6.9%. Teads Holding Co.'s revenue fell 11.7% that year on a pro forma basis. No source reviewed measures the open-web native market as such. So this paper does not quote a size for it. It treats any number that claims to be one as a choice about what to count. That choice still needs to be checked.

Chapter 3

Distribution and platform power

Open-web native is sold mainly through two large US-listed firms. Long, exclusive and guaranteed contracts lock up their supply. The second tier has been folding into fewer owners.

In native ads, the money is decided in distribution. So this chapter starts there, not with formats or targeting.

Two companies, two supply strategies

Taboola.com Ltd reported gross revenue of $1.9 billion for 2025. Its ex-TAC gross profit was $713.5 million [1, 32]. Ex-TAC means revenue after traffic acquisition cost. That cost is the money paid to publishers for the ad space. Ex-TAC is the number the firm manages to. Chapter 4 explains why you cannot compare it across firms. Taboola's supply strategy is written into its contracts. Historically, most of its publisher deals required exclusivity or other preferred-usage incentives. The average term at signing was more than two years [1] ◆. Three deals define its network.

The first is Yahoo. The deal was announced on 28 November 2022. It closed on 17 January 2023, and runs for thirty years. It gives Taboola the exclusive right to run native ads across Yahoo's digital properties. Yahoo got 24.99% of Taboola's shares and a board seat [33, 34] ◆. The contract has since been amended at least seven times. Amendments 4 to 7 took effect between February 2025 and March 2026; 5, 6 and 7, in turn, took effect in June 2025, August 2025 and March 2026. They added portal access through November 2027, and a direct insertion-order route for Yahoo. They also rewrote the monthly method for judging performance. The filing redacts the money terms of that method [35] ◆. A thirty-year exclusive needed four amendments in about a year. It is being renegotiated in public, with the numbers hidden.

The second is Microsoft. The tie dates from a January 2016 MSN deal. That deal covered 50 markets and 24 languages. In March 2025 Taboola announced ten years of serving MSN, Edge and Windows. It also announced a move into Outlook and Microsoft 365, but gave no term or renewal date. Nor did it say whether the deal is exclusive [36, 37]. Microsoft also sells native itself across Edge, Outlook and Bing [38]. In 2020 it signed MGID for native across its news network [39]. So "exclusive" in this market appears to cover the recommendation unit, not the page. That unit is the block of suggested links a platform runs on the page.

The third is Apple. In mid-2024, Taboola became an authorised seller of native ads in Apple News and Apple Stocks. It was still named as Apple's ad partner in February 2026, amid criticism of ad quality [40, 41, 42, 43]. What that deal covers, and where, is known only from press reports.

Teads Holding Co. is the other listed platform. It used to be called Outbrain Inc. Outbrain bought Teads from Altice. The deal was announced on 1 August 2024, at about $1 billion. It closed on 3 February 2025 on revised terms worth about $900 million. The parent renamed itself Teads Holding Co. in June 2025. It trades as TEAD [44, 45, 46] ◆. It describes three kinds of publisher contract: revenue share, programmatic bidding and guaranteed minimums. It says they cover about 10,000 media owners. Its top 20 partners, it says, have stayed seven years on average [46] ◆. Its 2025 revenue was $1.3 billion. It rose 46% because the purchase was folded into its accounts. On a like-for-like basis it fell 11.7% [46]. Chapter 4 takes that apart.

Supply is concentrated, and so is one demand channel

The clearest fact in Taboola's filing is the gap between its two sides. In 2025 Yahoo and Microsoft were its largest partners. No other partner reached 5% of revenue. Its ten largest advertisers together were under 10%, and none was above 3% [1] ◆ ◆. Outside Yahoo, demand is diffuse. Revenue billed through Yahoo was 10.5% of 2025 revenue and 16.6% in the second quarter of 2026, more than the ten largest advertisers combined ◆. In the fourth quarter of 2025 Taboola had about 2,200 "Scaled Advertisers". The term means those that spent more than $100,000 over the last four quarters. In that quarter they made up 84% of revenue. Each brought in about $204,000 on average [1] ◆. That is a mid-market demand base tied to a highly concentrated supply base.

The strategic reading follows. Losing a major supply partner could hit revenue much harder than losing any one disclosed advertiser. How much harder would depend on substitution, margins, contract terms and transition periods. So the contracts that matter most are on the supply side. The next chapter covers their cost: the guarantees, and the equity given to Yahoo.

Taboola's count of digital property partners was "approximately 14,000" in its 2025 annual report. It was "approximately 12,000" in its second-quarter 2026 filing. Taboola does not define what changed [1, 47] ◆. Revenue grew over the same period. Taboola does not disclose whether the change means lost partners, a new way of counting or something else. Meanwhile it has been opening up ad space that is not native. On 26 February 2025, it launched Realize. It sells display and other ad slots across Taboola's network. Taboola describes that network as about 600 million daily active users. By October 2025 TIME, The Weather Channel, Gannett, Nexstar and Slate had opened display space to Realize advertisers [48, 49] ◆. In August 2026 Realize began running programmatic display on NBCNews.com and TODAY.com. Taboola called this a first [50]. In June 2026 it also opened DeeperDive to ads. DeeperDive is the generative AI answer engine it runs on publisher sites. Now it places ads inside AI results pages [51]. The largest native platform is moving past the recommendation unit, into display and AI answers. It does not disclose how much of its growth each line adds.

The second tier has fewer owners than it did two years ago

Between June 2024 and September 2026, the independent tier of open-web native fell into fewer hands ◆.

• Sharethrough merged with Equativ. The deal was announced on 12 June 2024. It made a firm of 720 or more staff in 18 countries. On 9 June 2025, Sharethrough began to operate under the Equativ brand. But a Sharethrough-branded site was still live at the cutoff [52, 53, 54] ◆.
• In 2017, Outbrain bought the native DSP Zemanta. A DSP, or demand-side platform, is the tool buyers use to buy ads. On 6 August 2024 Zemanta was renamed Outbrain DSP [55, 56].
• Life360 bought Nativo. It announced the deal on 10 November 2025, for about $120 million in cash and stock. It completed the purchase on 2 January 2026, as a filing of 5 January reported [57, 58, 59] ◆. Nativo's revenue and publisher count at the time were not disclosed.
• Dianomi plc is listed on AIM. It focuses on native for financial content. On 18 September 2026 it announced a cash offer from Taboola Europe Limited. Its board intends to recommend it unanimously. The offer is 64p plus up to 24p contingent per share. That is about £19 million, rising to £27 million. The contingent part is paid only if a set of Dianomi publishers move to Taboola's standard terms. It is scaled to those publishers' revenue. Holders of 75.3% of shares have committed to the offer. The deal is expected to close before the end of 2026. It is subject to shareholder approval. It also depends on the UK competition authority not opening an in-depth review [60, 61, 31] ◆. Dianomi's 2025 revenue was £27.4 million.

Since March 2021, TripleLift has been majority-owned by Vista Equity Partners. It changed CEO twice between July 2024 and February 2025. In July 2025 it made layoffs. Anonymous sources put the cuts in the mid-to-high double digits. The base was over 400 staff [62, 63, 64]. MGID, Revcontent, Readpeak and Kargo remain private. MGID claimed 185 billion monthly impressions in 2025. In 2020, it used the same figure. Its legal owner could not be established from any opened source [65, 39].

Two research leads for this chapter were wrong. They are recorded here as such. Sharethrough was not bought in 2023. The Equativ deal was announced in June 2024. Adyoulike was not bought by Opera. OpenWeb bought it in April 2022. The reported price was $100 million [66].

Buying routes: the major platforms checked support native, and the market split is unmeasured

Programmatic native runs on IAB Tech Lab's OpenRTB Native Ads specification. Native entered the bid stream with OpenRTB 2.3 in January 2015. In March 2017, the spec was last finalised as version 1.2. That version added third-party creative, event trackers and a privacy link [4, 67] ◆. Prebid.js uses it for header bidding. There, the OpenRTB approach replaces the legacy native format [68]. The major platforms checked support the format. They include DV360, Google Ad Manager, AdSense and Microsoft Advertising. The Trade Desk has supported it since March 2016. Its first native supply partner was Sharethrough. Amazon Publisher Services supports it too. It announced APS Native ads for Amazon advertiser demand on 21 May 2025. It did not say when or where they would be available, or whether other DSPs get access [69, 70, 71, 72, 73, 38, 74] ◆.

None of the platforms above says how much the format is used. No IAB figure for OpenRTB Native uptake was found. Prebid publishes none either. A Sharethrough executive said in 2016 that uptake of the protocol had been slower than expected [71]. Support is not the same as spend. Some vendors disclose fragments, such as Dianomi's programmatic revenue. But no market-wide measure was found of what share of open-web native is bought through open programmatic pipes.

The documented routes are the platforms' own. Taboola sells through Realize, both self-serve and managed service. Teads sells through the self-serve platform from its Outbrain side, its own DSP and managed service. MGID, Revcontent and Readpeak sell through self-serve platforms. Dianomi and Kargo sell mainly direct. A buyer who wants open-web recommendation inventory at scale must still go to a platform. It has to be the one that owns the contract with the publisher. That is what supply rights mean in practice. It is why the next chapter follows the money through those contracts.

Chapter 4

The economics of an impression and a click

About six in ten dollars of reported revenue are booked as traffic acquisition cost. The rest pays for the business. At the two large listed platforms, that business is moving in opposite directions.

Native platforms sell a click to an advertiser and buy a slot from a publisher. Everything about their economics follows from two things. One is the gap between those two prices. The other is what they must promise publishers to keep the slot.

Where a gross dollar goes

Both listed platforms book most revenue gross, as principal. Payments to publishers are booked as traffic acquisition cost, or TAC, inside cost of revenues [1] ◆. So the reported revenue line is roughly the advertiser dollar. TAC is an accounting expense. It is not a line of cash paid out in the same period. Taboola's TAC includes non-cash amortisation of the Yahoo agreement asset. It also includes payments for impressions bought on exchanges. In 2025, Taboola took in $1,912.0 million. It paid $1,214.9 million in TAC and spent $127.6 million on other cost of revenue. It reported GAAP gross profit of $569.5 million [1, 32] ◆. Its ex-TAC gross profit was $713.5 million, or 37.3% of revenue. Ex-TAC gross profit is, roughly, revenue minus TAC. It is the non-GAAP measure Taboola manages to. Adjusted EBITDA was $215.5 million, 30.2% of ex-TAC ◆. GAAP net income was $42.3 million, after a loss of $3.8 million in 2024.

Teads Holding Co. took in $1,300.5 million. It paid $770.8 million in TAC and $100.6 million in other cost of revenue. It reported gross profit of $429.1 million. Its ex-TAC gross profit was $529.7 million and adjusted EBITDA $93.4 million, 17.6% of ex-TAC. Its net loss was $517.1 million. That included a $352.1 million goodwill impairment and $28.9 million of acquisition costs. It also included $15.3 million of restructuring charges [75, 46] ◆.

Three cautions before comparing the two. First, each firm defines ex-TAC its own way. Taboola adds back non-cash amortisation of its Yahoo agreement asset, which sits inside TAC. That was $16.4 million in 2025. From the second quarter of 2026 it also adds back a $12.2 million write-off of publisher prepayments. Reported ex-TAC growth that quarter was 11.8%. Under the old method, with the write-off left in, it would have been about 4.7%. That gap shows how much the add-back choice moves the number. It does not measure underlying growth, because the write-off also hit the reported accounting result [47, 76] ◆. Teads adds back only other cost of revenue. Second, neither firm reports net revenue. Gross revenue less TAC is a derived proxy. For Taboola it differs from ex-TAC by the add-backs. Third, Teads Holding Co. includes video, connected TV and display. Taboola includes display sold through Realize. So neither figure is a total for the native market ◆.

Taboola's TAC has risen as a share of revenue in each year but one. It was 59.3% in 2022, 62.8% in 2023, 62.4% in 2024 and 63.5% in 2025 [1, 77] ◆. It is not a pure publisher payout ratio. It also covers ad space bought on exchanges, up-front and incentive payments, and non-cash amortisation. Since 2024, some amounts tied to Yahoo have been shown another way. That affects the comparison. The Outbrain business paid out more. The share it paid out fell from 76.3% to 73.5% between 2022 and 2024. Over the same years its revenue shrank from $992 million to $890 million [78] ◆. A shrinking recommendation business can still improve its unit margin. Outbrain did that in the three years before it bought Teads.

The price of exclusivity

Chapter 3 showed that supply rights are the moat. The filings show what they cost. Taboola discloses two ways it pays publishers. The most common is a revenue share. The other is a minimum guarantee. Under it, Taboola pays whichever is more: the revenue share, or an agreed amount per thousand page views. It settles each month. Guarantee contracts usually run two to five years. Large property contracts "in general, contain minimum guarantee requirements" [1, 47] ◆.

Taboola also discloses what those guarantees cost. The cost is what it owes under guarantees above what revenue-share terms would have required. That was about 9% of TAC in 2021 and 10% in 2022, about 18% in 2024 and about 15% in 2025. In the second quarter of 2026 it was about 13% [1, 77, 47] ◆. Applied to reported TAC, the 2025 share works out to something on the order of $180 million. Publishers got that money because a guarantee, not a revenue share, set the price. In January 2018 trade reporting said publishers were giving up $1–2 million annual guarantees for revenue shares, to gain flexibility [79]. The filings say guarantees came back. Their share of supply cost reached about 18% in 2024, roughly twice the 2021 level, then eased; with 2023 missing, the series cannot show whether 2024 was the peak. In January 2023, the Yahoo agreement closed. The rise followed that, but the filings do not say the deal caused it. The guarantee described is a rate per thousand page views. So it guarantees a price, not a total annual income.

Teads Holding Co. warns in its own filing that guarantees "require us to pay our media partner for the ad impressions we receive, regardless of whether the consumer engages with the ad or we are paid by the advertiser". It also warns that TAC "may not correlate with fluctuations in revenue" [46, 78] ◆. That is the risk transfer in one sentence. Under a guarantee, the platform bears the gap between what advertisers pay and what the slot costs. The publisher does not.

What Yahoo costs, and what it returns

The Yahoo agreement was paid for in equity. Taboola books it as an asset and amortises it inside TAC. At the end of 2025, that asset stood at $270.2 million. In the same year Taboola bought back 76.9 million of its own shares for about $255.4 million [80, 1, 47, 32] ◆. It discloses two related-party lines. The first is revenue billed through Yahoo, where Yahoo acts as the advertiser. That was $201.6 million in 2025, 10.5% of revenue, down from $233.6 million. In the first half of 2026 it rose to $148.9 million, 15.8% of revenue. The second is TAC paid to Yahoo for placements. That rose to $349.0 million in 2025 from $275.5 million. In the first half of 2026 it rose to $201.1 million [1, 47, 80] ◆.

Those two lines do not add up to a profit test. This paper does not run one. Revenue billed by Yahoo is not all that is earned on Yahoo supply. Some advertisers buy direct through Taboola and run on Yahoo pages. They are not in that line. What can be said is narrower. In its 2022 announcement, Taboola forecast about $1 billion of annual revenue from the deal. It also set out a path to $1 billion of ex-TAC gross profit by 2025 [80]. Total ex-TAC in 2025 was $713.5 million. The forecast was not met on its announced timeline. The amendments described in chapter 3 rewrote the terms for judging performance. The money terms were redacted.

In May 2026 Taboola's CEO said it paid over $1.5 billion to publishers in 2025 [81]. Audited 2025 TAC was $1,214.9 million. This paper cannot reconcile the two figures. It uses only the audited one ◆.

Post-acquisition losses, impairment and financing pressure

Teads Holding Co.'s 2025 headline growth of 46% is acquisition accounting. Pro forma figures treat the deal as if it had closed on 1 January 2024. On that basis revenue fell 11.7%, from $1,507.2 million to $1,330.9 million. The pro forma net loss was $545.5 million. Between 3 February and 31 December 2025, legacy Teads added $517.2 million of revenue and a $350.8 million net loss [46] ◆. The purchase was funded with $637.5 million of 10.000% senior secured notes due 2030, issued at a discount in February 2025. At the end of 2025, $628.2 million of principal was still owed. About $62.8 million of interest is due each year through 2029. At 30 June 2026 the balance sheet showed $614.5 million of debt: $607.4 million long-term and $7.1 million short-term. Against that the firm held $88.0 million of cash. Stockholders' equity was $7.4 million, down from $95.4 million at year end [46, 78, 82] ◆.

The 2026 quarters got worse. In the first quarter, revenue was $266.0 million, down 7%. Adjusted EBITDA was $0.8 million. In the second quarter, revenue was $284.6 million, down 17%. Ex-TAC fell 14%. Adjusted EBITDA was $7.0 million and the net loss $42.5 million. On 6 August 2026 the firm suspended guidance. That included its target of about $100 million of adjusted EBITDA for the year. It cited "the volatility of the Direct Response and SME business" facing "open-web headwinds" [83, 82, 75] ◆. That $7.0 million second-quarter figure had missed the firm's May guidance of $14 million to $22 million [83, 82]. The notes carry about $31.4 million of interest every six months. Half-year adjusted EBITDA of $7.7 million was roughly a quarter of that. Connected TV, the line the company highlights, grew 67% year on year. It made up 13% of revenue in the second quarter [83, 82, 46] ◆. In the second quarter, operating cash flow was positive and the firm held cash. But adjusted EBITDA is not the same as cash to pay interest. Judging liquidity needs data on working capital, debt service, capital spending and covenant terms. The results release does not set those out. The filings neither show a liquidity event nor rule one out. In August 2026 the company also received a second Nasdaq notice for trading below the $1.00 minimum bid, with until 8 February 2027 to comply, and sued Google over its ad-tech conduct [84, 85]. They do show that the Outbrain–Teads deal took a goodwill impairment within a year of closing. That impairment was larger than a third of the price. Financing pressure has followed.

The one clean set of books

Dianomi is small and listed in London. It focuses on financial and business content. Its accounts give the clearest public view of the share of revenue a native vendor pays publishers. That is because its cost of sales is almost all publisher revenue share. In 2025 it paid £19.97 million to publishers on revenue of £27.4 million. That is a payout of 72.9%, leaving a gross margin of 27.1%. Adjusted EBITDA was a loss of £0.3 million. It earned 78% of revenue in the United States [86] ◆. Two of its operating figures matter beyond its size. By its own total, impressions fell 14.1% in 2025. The company said this was partly due to "AI-generated and so-called zero-click summaries". Revenue per click rose 7.4% to 58p [86, 87] ◆. The same release gives rounded Apple and non-Apple impression figures. They add up to a slightly different total and imply a fall nearer 15%. This paper keeps Dianomi's stated figure and notes the gap. In Apple News, Dianomi earned 22.5p per click, against 110.5p on its open-web publishers. That is a ratio of about 4.9. The ratio is descriptive. Advertiser, audience, geography, creative and buying mix all differ between the two placements. So it does not show that the placement alone causes the gap.

The arbitrage loop, and who pays for it

Recommendation widgets are the blocks of sponsored links that sit on publisher pages. They are documented as one channel that funds made-for-advertising sites. These are pages built to host ads rather than to be read. The joint trade-body definition of September 2023 makes "a high percentage of paid traffic sourcing" a core trait. The ANA's report names content recommendation platforms, along with social networks, as the channels that source those visits [88, 89] ◆. It works as a loop. An arbitrage operator buys clicks from a widget. It sends readers to an ad-heavy page. Programmatic brand budgets pay for the ads there. The operator earns more from them than the clicks cost.

The best-documented case is Forbes. In March 2024, Adalytics showed three things about the www3.forbes.com subdomain. Per Similarweb estimates, it drew more than 70% of its readers from paid display ads on Taboola, Outbrain and other paid traffic sources. In one observed session, a 52-slide slideshow served more than 200 ads, against three to ten on a normal main-site article. And it had run ads since at least May 2017. Bid requests routed through a Prebid server cut the URL short, so buyers saw www.forbes.com. The verification tags present on the main site were not seen on the subdomain. Forbes shut it on 2 April 2024 after a Wall Street Journal inquiry. It called the subdomain an insignificant part of its business [92, 93] ◆ ◆.

Who pays the cost can now be answered. The reader pays in ad load: 201 impressions in a session is the product. The brand advertiser on the programmatic side pays in wasted spend. The ANA's 2023 study covered 21 marketers. It found that made-for-advertising sites took 21% of their impressions and 15% of their spend. Their CPMs, the price per thousand impressions, were 25% below other sites. And they passed standard quality screens for viewability, invalid traffic and brand safety [89] ◆ ◆. The widget platform is paid on the way in. The exchange is paid on the way out. The publisher whose masthead lends the page its credibility is paid too. Each party was paid, which may explain how the subdomain ran for seven years. The record does not show why Forbes kept it.

Whether that loop is shrinking is taken up in chapter 10, because the answer depends on who is measuring. Some marketers opted into the ANA's log-level benchmarks. Among them, the median share of spend on made-for-advertising sites fell from 10% in 2023 to 0.39% in the third quarter of 2025. On the ANA's separate cost-waterfall measure it was 1.1% in the first quarter of 2026, a quarter in which the panel grew from 54 to 86 marketers and the ANA named "AI slop" as an emerging sub-type [94, 95, 96, 97, 98] ◆ ◆. That result covers only the marketers who took part. The top quartile of that same group still had up to 27% of web spend on such sites in the third quarter of 2025 [97] ◆. No market-wide figure after mid-2023 could be accessed.

Chapter 5

The reader and the creative

Several studies found that most readers did not know certain native formats were ads. The studies ran between 2014 and 2018. This review found no representative US replication since, and two recent leads arrived too late to screen.

Native advertising is defined by one trait. It looks like the content around it. This chapter asks what that does to the reader. It uses original studies and systematic reviews. Work paid for by a vendor counts as what the vendor reports.

How many readers know it is an ad

The anchor study is Wojdynski and Evans. It was published in 2016 with fieldwork in 2015. Readers saw a sponsored article on a page styled like a newspaper. The study tested twelve disclosures. Readers then got open-ended questions, with no answers to pick from. Across all of them, 17 of 242 adult readers knew it was an ad. That is 7%. A follow-up eye-tracking study used 60 students. With the clearest wording, "sponsored by Dell", 18% knew it was an ad [99] ◆. In a YouGov sample of 800 US adults, 9% spotted the ad. The sample was matched to the population. It was fielded in early 2017 and published in 2020 [100, 101] ◆.

Hyman and colleagues used a different measure. They showed 896 panel members 16 real native ads. The ads came from nine to ten publisher and social platforms. They asked a forced-choice question: paid, unpaid or don't know. Recognition was 37% for native, against 81% for conventional ads. Per ad, it ran from 21% to 72%, and 15 of 16 were below half. The 72% item was a Facebook ad, outside this paper's core scope [102] ◆. The measure explains most of the gap between 7% and 37%. A forced choice gives credit more easily than open-ended coding. So any headline recognition rate must say which one it used.

Labels matter, within limits. Shown on their own, labels with "paid" were read as ads by 83–89% of those asked. Labels with "sponsored" got 76–79%. Industry coinages such as "Brand Voice" or "Presented By" got 57–64%. In context, "Brand Voice" content on Forbes was recognised by 21–38%. Adding a prominent "Paid Ad" bar to two real native ads raised recognition from 40% to 56%. On one of them, the bar replaced the small "Sponsor Content" label. A third of readers still believed the content was unpaid [102] ◆ ◆. Wording drove recognition in Wojdynski and Evans too. The labels "advertisement" and "sponsored content" got 12–13%, against 2–3% for "brand voice" and "presented by" [99] ◆.

Where the label goes is disputed. Wojdynski and Evans found a mid-article disclosure raised the odds of recognition five times. That was against the top-of-page spot the Federal Trade Commission recommends. And 90% of readers' eyes landed on a mid-article label, against 40% for a top one [99] ◆. The FTC ran its own lab study, with 48 people. The fixes it recommends raised recognition by 21 percentage points across its search and native versions combined. That took it from about 47% to 68%. For the native versions alone, the gain was 23 points. A recommendation widget is the box of suggested links often shown under a story. The FTC noted that labels in the top-right corner of a widget were rarely looked at. It also noted that a significant share still did not know the ads were ads [103] ◆ ◆. So advice on where labels go rests on a few small, old studies. And they disagree.

What recognition does to the brand and the publisher

Spotting the ad goes with harsher judgements. But these particular comparisons show links, not proof of cause. Chance did not decide who spotted the ad. So the gaps may partly reflect who spots ads. In Wojdynski and Evans, readers who knew the sponsored piece was an ad found it less credible. They liked the sponsor less, rated the story lower and were less keen to share it. In the second study, only the credibility penalty held up [99] ◆. The publisher is the site that runs the ad. In the YouGov sample, spotting the ad lowered how credible readers found the publisher. It also lowered how they felt about it. The result was the same for legacy titles and digital-only ones [100] ◆.

The largest pooled evidence is a 2020 meta-analysis by Eisend, van Reijmersdal, Boerman and Tarrahi. It covers 61 papers, 57 datasets, 473 effect sizes and 278,791 people. The formats run from product placement to blogs to native. Disclosing sponsorship helps people spot the content as an ad (r = .255). It helps them see that the content aims to persuade (r = .257). It lowers brand attitude, or how much they like the brand (r = −.108). It lowers credibility (r = −.132). It shows no significant average link to what they say they plan to do (r = −.023 across 137,601 people). Results vary a lot from study to study [108] ◆. The word "advertising" in a label raised both recognition and credibility. Disclosures placed after the content did the most harm to the brand [108] ◆. Two things follow. First, disclosure goes with lower brand attitude and credibility. Its effect on actual behaviour is not proven either way. A stated plan is not a purchase. And a non-significant average is not proof of no effect. Second, there is no pooled estimate for native alone, because the meta-analysis mixes formats.

The mechanism runs both ways. In the meta-analysis's path model, grasping the intent to persuade drives the attitude penalty. A 2023 experiment adds the opposite path. Disclosure can make the ad seem more transparent. That lowers persuasion knowledge, the reader's sense of being sold to. It also raises brand attitude and purchase intent [108, 112, 109] ◆. A companion banner for the same brand raised recognition as much as a text disclosure did. Bad reactions were muted when the sponsorship seemed transparent [113] ◆. "Disclosure hurts performance" is not a general law. Which path wins depends on how the ad is made.

Clicks, attention and trust are three different things

Two randomised studies matter for buyers. Randomised means chance decided who saw which version. Sahni and Nair ran field experiments on native ads in mobile search, with more than 200,000 users. They found no evidence that typical disclosure formats fooled users. But that is search, not the open-web widget [114]. Aribarg and Schwartz ran the only randomised test located that pits in-feed native against display on a news page. They held page position fixed on a news site. An in-feed native unit got a higher click-through rate than a display ad, because it looked like editorial content. The display ad drew more visual attention and higher brand recognition. With it, readers also rated the website more trustworthy [115] ◆. The authors put the click edge down to looking like editorial content. The same look may also lower recognition. That is a reasonable hypothesis. If true, that edge would be the FTC's concern and the vendor's selling point at once. But the study does not test that path. Its effect sizes also sit behind a paywall this project could not open.

Attention evidence points the same way in social feeds. A 2025 German eye-tracking study used a mock Instagram feed. A fixation is a pause of the eyes on one spot. Organic, unpaid posts drew 25% more fixations than sponsored ones, and about 250 milliseconds more dwell. Once readers fixated on a disclosure or a call-to-action, dwell time usually fell. They were then asked how they told an ad from a post. They cited product visibility most often (178 mentions) and the disclosure label second (95). Then came logos (81) and buttons (77) [116] ◆. The first draft of this paper said readers ignored the label. The source says it was the second most-used cue. The FTC's lab found that better disclosures cut looking time by 21%. It warned that looking time is independent of understanding. So attention metrics cannot stand in for recognition [103].

Vendors have paid for attention and trust claims too. Funding is a conflict that lowers their weight. But the deciding problem with the two found here is that their methods are not disclosed. Outbrain paid for a Lumen study of 900 consumers in three European countries. It reported that native ads on premium news were "+31%" more likely to be trusted than social ads, and "+16%" more likely to be clicked. It disclosed no baselines, definitions or method. The report itself could not be retrieved [117] ◆. It compares surfaces, news against social, not formats on the same page. So it neither confirms nor contradicts the field experiment. A 2015 vendor survey by Contently found 48% of readers felt deceived on realising content was sponsored. And 62% said a news site loses credibility when it runs native ads. The survey also found higher trust in the sponsor when readers judged the content high quality [118, 102] ◆.

What the units actually carry

For recommendation widgets, consumer harm is not an abstract question. Researchers have measured what the ads contain. In January 2020, a crawl covered 6,498 mainstream news sites and 1,055 misinformation sites. The team coded 2,419 ads on 300 sampled sites. By the authors' codebook, 44.6% were "problematic". That category spans content farms and supplements. It also spans insurance and mortgage advertorials, investment pitches, sponsored search and misleading polls. Problematic in this sense is not the same as fraudulent. The sample is also a dated snapshot, not a full count. Of native ads, 87% were problematic, against 20% of display ads. Taboola served 61.1% of all problematic ads. Of Taboola's coded ads, 85.7% were problematic. For ads Google served, the figure was 16.9% [119] ◆. On this measure, mainstream news and misinformation sites showed no significant difference. In a 2020 survey, raters judged 500 real web ads. Most raters called 20.6% of the ads clickbait and 11.2% deceptive. The four lowest-rated groups of ads were 43–72% native and recommendation ads [120] ◆.

Political clickbait rides the same units. Researchers crawled 1.4 million ads on 745 US news sites around the 2020 election. Of the political ads, 52% promoted "political news and media" articles. The recommendation network Zergnet served 79% of those. Taboola came next at 10%, then Revcontent at 5.7% and Content.ad at 1.8%. A smaller follow-up crawl of the same 745 sites ran in October and November 2024, over five weeks. It covered 15,110 ads. Again, it found clickbait political news ads common [121, 122] ◆. Regulators met the pattern even earlier. In April 2011 the FTC sued ten operations that ran fake news sites. The sites used made-up reporters to sell weight-loss products. The sites drew traffic with "attention-grabbing ads on search engines and high volume websites". Together they had paid more than $10 million to advertise [123] ◆.

Two caveats carry weight. First, the 2020 crawl is a snapshot. It comes from before the industry's made-for-advertising crackdown. Chapters 4 and 10 describe it. A February 2026 preprint crawled 450 news sites from three locations over five months. It pins problematic advertorial ads mainly on Taboola and Outbrain. That suggests the pattern went on. But it is a preprint, and its coding differs [124]. Second, "problematic" is a broad codebook. The evidence shows a narrower point than a verdict on the category. Compare the recommendation unit with display on the same pages. The unit has carried a materially higher share of low-quality demand. That was true on one platform above all. One independent 2026 preprint, with a different codebook, suggests it still is. This review found no peer-reviewed replication in the sources screened by the cutoff.

What the rules require

The FTC's December 2015 guidance sets a performance standard, not a formula. The test is whether consumers know the native ad is an ad. It recommends "Ad", "Advertisement", "Paid Advertisement" or "Sponsored Advertising Content". It calls "Promoted" and "Promoted Stories" at best unclear. It wants the disclosure in front of or above the headline. It also rejects one disclosure for a mixed group of paid and unpaid items in a recommendation widget [2] ◆. A policy statement on enforcement came with it. It cites consumers reaching fake news sites "by clicking on ads presented as attention-getting news headlines, which frequently appeared on legitimate news websites". It also makes ad agencies and affiliate networks liable [125, 2] ◆. Chapter 10 sets out the European and UK rules. It also covers the 2023–2024 changes to rules on endorsements and reviews.

Two points close the chapter. The FTC's advice is the top of the item. Wojdynski and Evans found that spot least effective. And the FTC's standard turns on whether consumers recognise the ad. But it sets no numeric pass mark. Compliance depends on each disclosure's wording, placement and overall net impression. The studies above show that few readers spotted the ads under several common disclosure practices of their time. They cannot settle whether any one unit running in 2026 is legal.

Chapter 6

What the machines optimise

In the per-click campaigns the help centres document, the commercial unit is still a click. The objective has moved to conversions. That shift hands the platform, not the advertiser, the decision about which clicks to buy.

Chapters 3 and 4 followed the contracts and the money. This chapter opens the box between them. Inside is the software that picks which ad a reader sees and what the advertiser pays for it.

Six stages, documented by the vendors themselves

The recommendation engine behind a native widget does the same job as any large ad system. The vendors have published enough to describe it. Taboola's 2025 annual report says how its system ranks each candidate. It combines three inputs into a relative value. The first is the predicted chance that a user interacts. The second is the predicted chance of a conversion after a click or view. A conversion is the action the advertiser actually wants. The third input is the bid. Its auctions run CPC and CPM pricing models at the same time [1] ◆. The same filing reports 500,000 recommendation-related requests a second. That figure has stayed the same in each annual report since fiscal 2022 [77]. The filing also reports over 600 million people reached a day. Its peak figure is up to 1.2 trillion recommendations a month. And it cites four data centres that process over 170 terabytes a day to train its models [1, 126] ◆.

The papers fill in the stages. In 2022 Taboola said it served more than a billion requests a day. Each day it deployed hundreds of models that predict clicks, on separate traffic segments. Each teacher model trained from scratch for about two and a half hours. It used around 100 million impressions from the prior fortnight. Student models were updated in small steps. Each was fine-tuned on about 12 million fresh samples in twelve minutes. They were redeployed every four hours [127] ◆. Taboola's 2018 paper was on deep density networks. It set the response budget at under 50 milliseconds. It defined the goal as revenue per thousand impressions at a fixed CPC. It split the algorithm into two modules. Exploitation uses what the model already knows. Exploration tests new options. The split lets it handle tens of thousands of new candidates a day [128] ◆. Zemanta is Outbrain's demand-side platform (DSP), the software that bids for ad space for buyers. It reported more than a million bid requests a second. Its maximum response time was 100 milliseconds. Its click models computed over 600 million predictions a second on CPUs [129, 130] ◆. Yahoo's native marketplace used a feature-enhanced collaborative-filtering model. It predicted the chance of both a click and a conversion for its auctions [131, 132] ◆.

Read together, the documents describe six stages. The first three are candidate retrieval, a click model and a conversion model. The fourth ranks candidates by expected value, with the bid. The fifth picks the creative after the auction. The sixth is an exploration module fed by incremental training. For Taboola, a current filing plus its papers show that design. For Zemanta and Yahoo, papers dated 2019 to 2022 show it [1, 127, 128, 129, 130, 131, 133, 134] ◆. A paper shows what was deployed when it was published, not in 2026. The figure marks which is which.

Every lift number in this chapter was measured by the vendor

The technical papers are specific about what each part is worth. They are also all self-reported. Distillation is the teacher and student setup above. Taboola's pipeline lifted revenue per thousand impressions by 0.53% to 0.85%. That was across four traffic segments in a six-month test. A naive warm-start on fresh data cut revenue. So fresher is not always better [127] ◆. Taboola's exploration module found 6.5% more new targets and 2.1% more advertisers. The cost was 0.05% of revenue. The deep model lifted revenue 2.9% over a tuned regression [128] ◆. Zemanta documented a feedback loop. A DSP only sees outcomes for impressions it wins. So it under-buys what it under-observes. Zemanta countered it by exploring where the model was least sure. That raised revenue 5.0% and click-through 9.6% against control. Random exploration also raised revenue, by 3.2%, but it made the model worse. So part of the gain was simply bidding more [130] ◆.

Yahoo's post-auction stages report the largest numbers. Carousel asset optimisation gave +8.6% click-through and +4.3% revenue. Dynamic creative optimisation based on conversions gave a 53.5% lift in conversion rate. But the control served title-and-image pairs at random. So the added value over click-based selection is unknown [133, 134] ◆. The same authors credit a later near-40% rise in carousel revenue to more traffic. That traffic came from extra advertiser demand. They take that as a sign advertisers were happy with the product. It is not a per-impression lift from the algorithm. It is a useful warning not to read product growth as optimisation. Clicks themselves are also noisy. Verizon Media studied accidental clicks with a short dwell time. It found they mislabel training data. Re-weighting them lifted revenue 1.18% [135] ◆.

The pattern is the point. Each performance claim for these systems comes from a live test. The vendor's staff ran it, and the vendor's paper reported it. No independent party has replicated any of them. None has an audited measurement. That does not make the numbers wrong. It makes them what the source hierarchy says they are: evidence of what the vendor reports.

From clicks to conversions, and who holds the risk

The commercial shift of the last three years is in the goal, not the unit. On Taboola, Outbrain, MGID and Yahoo, the advertiser still pays per click. But the platform now offers to turn your goal into per-click bids. The goal can be a cost per acquisition (CPA) or a return on ad spend (ROAS). The platform sets the bids with its own conversion model ◆. Taboola's Realize offers three bid strategies. Maximize Conversions is automated and is the one Taboola recommends. A target CPA is optional. Enhanced CPC raises a baseline bid on high-value impressions and lowers it on low-value ones. Fixed Bid is the third. Taboola's documentation says the entered bid "will always be the price paid in auctions" [136] ◆. Outbrain's Conversion Bid Strategy offers Max Conversions, Target CPA, Target ROAS and a semi-manual mode. Each is described as adjusting the campaign CPC [137, 138, 139] ◆. Its Engagement Bid Strategy aims at on-site metrics from Google Analytics. The options are clicks, pages per session, session duration or bounce rate. It may raise the CPC to three times the entered bid [140, 141] ◆. In June 2025, MGID announced CPA Tune. It keeps CPC pricing. But its algorithm "decides which clicks to buy based on predicted conversion probability" [142, 143, 144] ◆.

That last phrase describes a shift in who decides, not in who carries the financial risk. Under fixed CPC, you chose the bid. Under a CPA goal on a CPC unit, the platform's model picks which clicks are worth buying. But you still pay for each click bought, even those that never convert. MGID says so directly. The target is "a directional goal rather than a guaranteed cost per conversion", and "you continue to pay per click" [143]. Financial risk would move to the platform only under a different contract. It would have to guarantee the outcome or charge per conversion. The self-serve product documentation reviewed offers neither. But both large platforms' filings describe campaigns priced per acquisition, and Teads books some revenue on cost per incremental action. Neither discloses what share of revenue these carry or how incrementality is measured [46, 1]. You gain convenience and give up two things. One is control over what is bought. The other is the data that would show whether the platform's model is right. Fixed Bid and manual CPC are still on offer, so the choice is real. But the vendor guides steer buyers toward automation.

The learning rules published with these products are the practical limit. They also converge. Realize's Maximize Conversions has a learning phase of two to five days. Before you can set a target CPA, it needs at least 50 conversions. They must come in seven days in a row. It recommends a daily cap of ten to fifteen times the expected CPA. The 50-conversion rule applies only to setting a target. You can use Maximize Conversions without meeting it. Some campaigns lack the Taboola pixel or server-to-server tracking. For them, the strategy aims at page views and impressions instead [145] ◆. Outbrain's Target CPA aims for seven to ten conversions a day. It wants daily spend of at least seven times the target. It takes up to a week to learn [137, 139]. MGID's learning takes seven to fourteen days or ten to thirty conversions. It advises daily budgets of five to seven times the target. Its exploration phase bids higher until the first conversion, or until three times the target has been spent [142, 143, 144]. The rules come in three kinds. Taboola's threshold for a target is a hard gate. Outbrain's daily volumes and the budget multiples are advice. MGID's period is learning guidance. Together they mean a low-volume campaign learns slowly. It may never qualify for target-based bidding. Take an advertiser with three conversions a day. It gets 21 in a week, below Taboola's 50-in-seven-days threshold. It can run Maximize Conversions, but it cannot set a target cost per acquisition. How many advertisers sit below these lines is not published. The documented workaround is to aim at an upper-funnel event. That is a step that comes before the conversion. It eases the limit by optimising to a proxy ◆.

Two details from the documentation deserve a buyer's attention. First, take a "Maximize Conversions" campaign with no conversion tracking. It is optimising to page views. So the label and the objective can differ. Second, Realize's help centre states the price rule by strategy and route. Enhanced CPC cuts a winning bid to the second-highest bid. Creative Expedition charges the second price, even when it raises bids up to seven times. Fixed Bid charges the entered bid. Realize also sells ad space programmatically. That space clears at first price in private marketplaces and at second price on the open exchange [147, 148, 149, 136]. No primary source reviewed states one rule for all on-network inventory. For programmatic supply, the OpenRTB protocol signals the rule on each request [150] ◆. The mechanism may differ by product, route and date. The public documentation does not say. That silence is itself the finding. A buyer paying per click cannot see from it how the price was set.

Generative creative, and what it has been shown to do

Creative automation is the part of the stack vendors talk about most and document least. Taboola's GenAI Ad Maker makes images and titles. It uses outside models plus its own layer trained on network trends. Taboola announced Realize+ on 23 April 2026 after a beta. It was entering a second expansion phase in the second quarter. It adds a decision engine that moves budget in real time. It also adds an element generator that makes and revises ads and targeting. The same release opened the platform to a first Claude Skill for campaign setup and optimisation [151, 152, 1, 153] ◆. MGID has offered image generation, from text or from other images, since February 2024. It has also offered generated titles and descriptions in 29 languages since then [154]. MGID also scores a creative from poor to excellent before launch. The score uses its title, image and targeting. No accuracy figures are published [155] ◆. Making ads, scoring them before launch and setting bids are different jobs. The vendors document each on its own. Teads markets a predictive pre-test built on eye-tracking and EEG data. Its product page lists "4 billion signals a minute" [156, 157, 158] ◆.

The one outcome number comes from a Taboola press release. It covers a university study of Realize ads. The release's own headline is that AI ads match human creative. In the raw data, AI-made ads had a 0.76% click-through rate. Human-made ads from the same advertiser, campaign and day had 0.65%. The data spanned hundreds of thousands of ads and 500 million impressions. Under the study's tightest statistical controls, the two did about the same. The release gives no number for conversions. It says only that AI visuals did not reduce downstream conversion performance [159] ◆. The working paper behind it, by Exner, Hartmann, Netzer and Zhang, can be identified. The version this project opened, dated January 2025, has three points of note. First, ads with AI-generated images beat human-made ones on click-through only when they do not look AI-made. Second, it finds no significant effect of AI images on conversion rate for ads with a purchase objective. Third, it cautions that its data cannot isolate conversion effects [160]. Its sample totals differ from those in the SSRN abstract and in Taboola's release. So each figure is quoted with its version. The first draft of this paper read the raw gap as the result. The fact-check pass corrected it. Clicks in native units include accidental ones. So a click-through gap would not be a value gap anyway. Here is the summary that survives. Generative creative is shipped and widely used. A study on the vendor's platform found it no worse than human creative on clicks. Under some conditions it did better. Its effect on conversions is not established.

Speed, and what "real time" means here

The sources give latency figures for different stages. Latency here means response time. Taboola's engine responded inside 50 milliseconds in 2018. Its 2025 engineering blog puts the whole widget response at "hundreds of milliseconds". The blog also notes strict limits on database queries. Zemanta's budget for a bid reply was 100 milliseconds [128, 161, 129]. Taboola's rendering engine cut how long its scripts block the page's main thread. The cut was 485 milliseconds, about 70%. It also improved interaction-to-next-paint on publisher pages. That metric tracks how fast a page reacts to input. The gain was 6–36% at the 75th percentile, across four publishers. It did so without hurting click-through or revenue [161, 126] ◆. No vendor publishes a current end-to-end budget. Model training happens offline. But in the systems the papers describe, some work still happens when the ad is served. That work is retrieval, feature computation and running the models. So the sources do not support treating the request as a simple lookup. For a buyer, the learning-phase rules above limit a campaign more directly. The millisecond figures do so less directly.

Chapter 7

What performance evidence proves

No public incrementality study of open-web native, with its design and results, was found. One practitioner account says a Taboola geo test was run and used to calibrate a mix model. It does not publish the test's own result. What the vendors publish instead uses designs that randomised tests show can overstate.

Chapter 6 showed the machines chasing conversions. This chapter asks what evidence would show they caused any. That is the question of incrementality: what an ad caused, beyond what would have happened anyway. The standard test uses a holdout, a group kept from seeing the ads to compare against. For open-web native, no such evidence has been published in a form a reader can check.

Two facts every incrementality claim has to survive

First, ad effects are small next to the noise in outcomes. So even large randomised tests are not precise. One study looked at 25 field tests run on Yahoo for US retailers and brokerages. They covered millions of customers and $2.8 million of digital spend. The median confidence interval on return on investment was more than 100 percentage points wide. The authors calculate that a test that tells you something can need more than ten million person-weeks [162] ◆. "Experiment-grade" does not mean precise. Randomising supports an unbiased estimate when the test gets five things right. They are assignment, exposure, outcome measurement, attrition and spillover. A geo test runs ads in some areas and not in others. It needs similar areas too, and enough power to detect an effect. The label alone ensures neither.

Second, observational methods do not match the randomised answer. That holds even with great data. These methods skip random assignment. Gordon and colleagues used 15 Facebook tests. The tests held 500 million user-experiment observations. They found such methods "often fail to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables" [163] ◆. Their 2023 follow-up covered 663 treatment-control pairs from 563 Facebook experiments. It used more than 5,000 user-level features. Its result is the figure that matters most in this chapter.

Take one method, double machine learning. For upper, middle and lower funnel outcomes, it gave median lifts of 83%, 58% and 24%. The randomised lifts were 29%, 18% and 5%. So the ratios of medians were 2.9, 3.2 and 4.8. Stratified propensity matching was worse still, at 6.0, 9.8 and 12.8. A ratio of medians is not the median of the per-test ratios. These are also Facebook results. They are not a correction factor for any other platform. The authors conclude they "are unable to reliably estimate an ad campaign's causal effect" without experiments [164] ◆. The reason got its name a decade earlier: activity bias. Those who see an ad on a given day do more of everything online that day. So comparing those who saw an ad with those who did not overstates the effect. So does comparing before and after. In one Yahoo test, comparing before and after would have overstated the effect by 350%. The true effect was small and insignificant [165] ◆.

Neither fact is about native. Neither gives a bias multiplier for Taboola, Teads or MGID. What carries over is the problem of telling cause from selection. The selection that creates the bias is an algorithm choosing who sees the ad. That is what chapter 6 describes.

What native vendors publish

This project could open three sources on whether open-web native works. The first is Taboola's 2023 meta-analysis of Kantar brand-lift studies. They span Europe, Latin America and the United States. These surveys test whether ads shifted views of a brand. Taboola reports an average lift of 15% in favourability, against a 4.2% Kantar norm. The average lift in consideration was 6.7%, against 4.6%. For message association it was 87%, against 10.1%. Pixel and cookie tracking marked who saw the ads and who did not. Taboola gives no study count, sample sizes or intervals. The units are not stated [166] ◆. Most of the studies included video. The same page says the favourability lift was nearly the same as Kantar's video-only norm, at about 15%. The fourfold headline compares against an all-digital norm. The second is Outbrain's 2018 "True Engagement" case. Outbrain commissioned Nielsen to survey 1,573 users, for an auto client in Germany. It reports +23% brand lift. The groups compared were "those who clicked on the page and those who did not" [167] ◆. That is selection on the outcome. The third is Taboola's September 2026 Realize ID release. It claims "up to 2.4x increase in conversion efficiency". It does not define the metric. It gives no baseline, period, advertiser count or method. It uses the word incremental. But no holdout, test or measurement partner stands behind the figure [168, 169, 170] ◆.

These are evidence of what the vendors report. The IAB and IAB Europe issued guidance in November 2025. It defines incrementality as the causal impact of marketing against what would have happened without it. It states plainly that attribution and return on ad spend "show what happened, not whether marketing caused the result". Then it grades the methods. Experiments are strong. Model-based counterfactuals, which model what would have happened, are strong to moderate. Marketing mix models, which link sales to spend over time, are moderate to weak. Hybrid proxies are weak. These include baseline versus exposed analysis and new-to-brand share. So does incrementality as the platform reports it [171] ◆. The native vendor designs above fall in the weak class. Measured, a measurement vendor, markets a way to run incrementality tests on Taboola. Realize's own explainer describes geo-holdout and audience-split tests. But no randomised control was found in any Taboola or Teads product, study or case [169, 170].

Randomised evidence on native formats outside the open web points the other way from the vendors' silence. Facebook's in-feed ads are native by format, and the same large experiments found positive median lifts of 29%, 18% and 5% by funnel stage ◆. Sahni and Nair's randomised test of native ads in mobile restaurant search, with more than 200,000 users, found that the ads benefited advertisers, and that most of the extra conversions came from exposure, not from clicks on the ad [114]. That second point matters for chapter 6: per-click billing prices the click, while the effect in that test came from being seen. Neither study covers the open-web widget.

What the experiments that exist say about the mechanisms native relies on

Retargeting follows the visitor. It shows ads to users who have been to a site. Recommendation widgets carry it at scale. There are randomised tests of it. One ran with an online home-improvement store. Retargeting caused 14.6% more users to return within four weeks. A third of the first week's effect came on the first day. Second-week ads worked better when first-week ads had been shown [173] ◆. Against that, a field test at an online travel firm looked at dynamic ads. These show products a user viewed before. On average they worked less well than generic brand ads. They stopped doing worse only when browsing showed a reader's preferences had narrowed [174] ◆. In those settings, retargeting raised return visits. Showing the products a user had viewed was not better by default. Neither study measured incremental profit. Neither result extends to all retargeting.

Differences between users matter more than averages. Tests at eBay in 2012 found brand-keyword search had no short-term benefit they could measure. Non-brand keywords had negative average returns. New and infrequent users responded. Frequent users, who drove most of the spend, did not [175] ◆. A second study covered 288 US consumer-goods brands. Their TV elasticities were far below those in published studies. These measure how much sales move with ad spend. Marginal returns, the return on an extra dollar, were negative for more than 80% of brands. The authors list publication bias in earlier estimates as one possible source of managers' optimistic priors [176] ◆. The lesson that carries over is the direction, not the size. Published and vendor benchmark lifts are biased upward. And the users a system finds easiest to reach are often the ones who would have bought anyway.

Ghost ads show what a platform can do when it wants to measure cause and effect. Google's method finds the control-group users who would have seen the ad. For one retargeting campaign it measured +17.2% site visits and +10.5% purchases. It gives the same precision at a tenth of the cost of public service ad controls. It logs over 100 million predicted ghost ads a day [177] ◆. The platform has to build the method. This review found no version built by a native platform.

Windows, lag and the reported number

Attribution windows decide what a platform reports. They moved in 2026. A window is the time after a click or view in which a conversion counts for the ad. Google Ads' default click-through window is 30 days. Its default view-through window is one day. A view-through conversion comes from someone who saw the ad but did not click. These count only where at least half the ad was on screen for a second. They exclude users who interacted with any other ad. They sit outside the main conversions column. And they cannot be reported where cross-site cookies are blocked [178, 179] ◆. Meta deprecated the seven-day and 28-day view-through windows in its Ads Insights reporting API. The change took effect on 12 January 2026. The sources opened do not say campaign attribution settings changed that day [180] ◆. When a platform shortens its windows, attributed conversions fall. What the ads caused does not change. Meta's change says nothing about Taboola's reporting. Taboola's published defaults are a 30-day click window and a 24-hour view window [181]. The point is general. Take the September 2026 Realize ID release, which states no window or baseline. A claim like that cannot be checked against either kind of change.

Conversion lag is real. It differs by channel. Lag is the delay between the ad and the conversion. Retargeting gave a third of its first-week effect on day one [173]. Haus pooled hundreds of geo tests from the 2025 holiday period. It found 41% of media's incremental value came in the post-treatment window. That is the period after the test ended. But Haus is a measurement vendor, and the studies are its clients' [182] ◆. The MRC's Outcomes and Data Quality Standards, backed by the ANA, the 4A's and the ACA, cover outcome measurement across media. They require lookback windows to be disclosed before a campaign runs, and measurement providers to set limits on window length that data supports [183]. The IAB and MRC retail-media guidelines set the same kind of rule for retail media. The windows must be backed by data and disclosed before a campaign runs. Day-level lag data must be on hand [184] ◆. Several native vendors publish their attribution windows. Among them, Taboola's click window runs 1–30 days and its view window 1–24 hours. Revcontent uses a 30-day pixel window. Nativo's windows can be adjusted [181, 185, 186]. None of the sources opened for this paper shows the realised lag distribution behind reported conversions.

Marketing mix models do not settle it either

Buyers who cannot run tests turn to mix models. Both large platforms now publish tools to calibrate those models with tests. Calibrating means using test results to correct the model. Google's Meridian documentation states "there is no single formula to translate an experiment result into a prior". It warns that results from a different period may not transfer. It notes that a mix model's counterfactual is zero spend. Many tests instead measure reduced spend [187, 188] ◆. Meta's Robyn treats agreement with tests as a third optimisation goal. Its documentation cites an Analytic Edge whitepaper. The paper puts the average gap from ground truth at 25% for models that were not calibrated [189] ◆. Meta's October 2025 guidance cites two sources. In a study, calibrating with tests changed measured return by 29%. In a survey, 55% of those asked often get clashing results from different measurement tools [190] ◆. The IAB's December 2025 guide tells marketers to use tests to calibrate models. It says to match the same outcome, region and window. It also says to repeat a test before treating its result as an anchor [191] ◆. Take a mix model built on sales measured outside the ad platforms. It does not depend on any platform's window. Window choices matter only where attributed conversions feed the model. They can feed it as inputs or as calibration targets.

The industry itself is not happy. Three in four marketers say their attribution, incrementality and mix-model approaches lack the speed, accuracy or trust they need. That is the finding of the IAB's State of Data 2026 report [192] ◆. An often repeated figure says 71% of advertisers rank incrementality as the top retail-media metric. It could not be traced to the ANA release it is credited to. That release instead reports 71% naming sales conversion as their biggest goal [193] ◆. The figure is not used here.

What this means for the questions the paper set out

The incrementality question asked when native creates new demand. It also asked when native harvests intent, or takes credit for conversions that would have happened anyway. On the public record, the answer is unknown outside the platforms and their clients. That is because no test that would answer it has been published for this channel with its design and results. The question on AI and agentic buying asked what truly autonomous buying can do beyond today's optimisation. Whatever it can do, the evidence for it will have the same problem until it includes a holdout. An algorithm that picks the readers most likely to convert is what selection means. The Facebook tests show how badly selection can inflate observational lift. They do not tell us by how much on any other platform.

In practice, this leads to the measurement decision tool in chapter 10. The buyer's question decides the test. "Did this campaign cause sales" needs a randomised or geo holdout. It also needs enough volume to see a small effect. "Which of two creatives performed better" needs only a split. "Is the platform's reported CPA true" needs the window, the conversion definition and the lag data. The MRC's outcomes standard and the retail-media guidelines both call for the window to be disclosed before the campaign, and the lag data to back it. Native vendors document the first two in varying detail. None of the sources opened discloses the third.

Chapter 8

Commerce and new discovery surfaces

Each native vendor points to commerce. None of them reports it as a line of its own. The systems next door are better documented. They show attributed sales, not incremental ones. That is, sales an ad gets credit for, not sales the ad caused.

This paper grew out of the Performance & Native playbook. The playbook argues that native moves in steps. First comes placement, then recommendation, then commerce, then outcome infrastructure. This chapter tests the commerce step. It checks it against filings, product docs and publisher accounts. Then it looks at the newest discovery surface, the AI assistant.

What the native vendors sell as commerce

Taboola's commerce story is a data story. Its 2025 annual report points to Connexity. Through it, publishers get access to over 500 million structured product listings. These earn money per click or per action. The report pitches the pairing of reader data with purchase data as an asset [1] ◆. In 2021, Connexity cost about $800 million. It had itself bought Skimlinks the year before. At the time of the deal it had $176 million of revenue. It had 1,600 direct merchants. Its ex-TAC gross profit was $78 million [195] ◆. Ex-TAC means gross profit after traffic acquisition cost. That is the money paid out to publishers for their traffic. Turnkey Commerce launched with TIME in February 2023. In March 2024, it was extended to the Associated Press. For a publisher's commerce section, it supplies the editors and the tech. It also supplies the affiliate links that make money. The terms are not disclosed [196, 197] ◆. Skimlinks says it turns links into affiliate links across 48,500 merchants. It says it serves 23 of the top 25 publishers worldwide [198] ◆. Its public FAQ says it shares revenue and takes "a small percentage" [199]. Its support documentation states a typical 75/25 revenue share: the publisher receives 75% of the commission the merchant pays, and Skimlinks keeps 25%. Individual terms may differ [200].

None of this shows up as a revenue line. In the readable parts of the 2025 annual report, e-commerce and Connexity revenue are not broken out [201, 1] ◆. If Connexity had not grown at all since 2020, it would be about 9% of Taboola's 2025 revenue. That number illustrates what the gap hides. It is not a measurement. The filing is clear on one thing: where the firm now stands. Realize launched in February 2025. The filing calls it a performance platform that "addresses the limitations of native advertising alone". Realize supports formats that reuse display and social creative. It claims a $55 billion opportunity, based on internal and external industry data [1] ◆. The largest native firm's 10-K no longer leads with native.

Teads Holding Co. says even less. Its 2025 annual report names retail media as a planned growth focus. It also names it as a risk to its growth plan. Connected TV, in-app and AI answer environments are named alongside it. But the report discloses no retail-media or commerce revenue, product or partner. Pentaleap is not mentioned. The first-quarter 2026 report does not mention retail media or commerce at all. It says Teads makes money on CPC, CPM, CPV and CPA, with no commerce line [46, 202] ◆. The commerce aim lives in a press release. On 24 July 2025, Pentaleap and Teads announced a real-time-bidding integration. It lets advertisers buy retailers' onsite sponsored-product ads through Teads Ad Manager. The release named no launch retailer. By the cutoff, no live retailer, volume or result had been published [203, 204] ◆.

What the adjacent systems show

Native vendors point to retail media as the budget next door. They usually attach a forecast. Criteo, which calls its program the world's largest independent retail-media API program, tells a flatter story. Its Retail Media segment made $263.9 million of revenue in 2025, up 2%. That was about 13.6% of the group total. It made $259.7 million of contribution ex-TAC, about 22.1% of the group's. The fourth quarter was down 17% on scope changes with two clients. For 2026, the firm guides flat to up 2%. Criteo serves about 235 retailers. It defines offsite as ads "across the open internet". It does not disclose the split between onsite and offsite [205, 206] ◆. Its Commerce Grid supply platform lists native among the formats it supports. It claims 60 or more DSP integrations [207, 206] ◆.

Among the sources retrieved, Amazon's docs are the clearest. They show how retail purchase signals reach publisher pages on the open web. Amazon Publisher Cloud is built on AWS Clean Rooms. It lets publishers plan packages of ad space and enrich their signals. The goal is to unlock Amazon DSP demand. The data never leaves Amazon. Dotdash Meredith is quoted mapping "intent-to-buy signals" across more than 1.5 million articles [208, 209] ◆. In August 2023, Amazon said Sponsored Products would run on Pinterest, BuzzFeed, Hearst Newspapers, Raptive and Ziff Davis sites. They would run automatically, on advertisers' current cost-per-click bids [210]. They are served "Beyond Amazon" on third-party sites and priced per click. In mid-2025, two Amazon ad-tech vendors reported a new setting for off-Amazon campaigns. They said its default was "Maximize reach" [211, 212, 213] ◆. Suppose the vendors have the default right. Then a rise in ads shown off Amazon, in native-like slots, may not mean more advertiser demand. It may reflect a supply-side default. No Amazon docs on the 2025 setting were found.

Amazon Marketing Cloud is free for advertisers who qualify. It lets them query the platform's data in aggregate. The rules say plainly that uploaded signals cannot be exported [214] ◆. A clean room is a shared space where firms can match data without handing it over. Access to one does not make an analysis causal. Take an exposed-versus-unexposed comparison: those who saw the ad against those who did not. Run it inside a clean room and it is still just that. What a clean room does is allow a proper design, where the platform supports one. None of the nine native vendors profiled in chapter 9 documents a clean room in the public material this review opened.

Content-to-commerce at the publisher

Publisher filings show what commerce is worth to the firms whose pages carry native units. The answer varies by publisher. The reasons have nothing to do with ad tech.

People Inc. was renamed from IAC in June 2026. It grew affiliate commerce 11% in the second quarter of 2026. Meanwhile its core sessions fell 22%, "due primarily to the impact of the growing prominence of Google AI Overviews" [215, 216] ◆. BuzzFeed's affiliate commerce fell. It tied the fourth-quarter drop to "changes in supplemental bonus structures from our partners". About 28% of its 2025 revenue came from Amazon [217, 218] ◆. Future's affiliate revenue rose 10% in the first half. It fell 22% in the second. For the full year, Future cites a 13% fall in unique page views, weaker consumer confidence, currency moves and closures. Future reports that Google Search's share of its audience fell from 67% in 2019 to 27%. It also reports AI Overviews reaching half of its key terms [219] ◆. Three publishers, three directions, on similar traffic shocks. What they share is how the money is booked. Each of these lines is booked as attributed commissions. One partner can change its commission or bonus. That can move the line more than any change in how the placement performs ◆ ◆.

Conversion data in, aggregates out

The playbook's commerce thesis needs data to flow both ways. Conversion data, the record of who bought or signed up, must flow into the native platform. Learning must flow back out to the advertiser. The docs show the first path is narrow. For the second, the public documentation reviewed describes aggregate reporting. It does not establish an advertiser right to export the platform's learned models or reusable audience data. Contracts may differ.

The docs that come from Outbrain list the ways in. One is a browser pixel with 24-hour first-party matching. Another is offline conversion files. They are emailed to a service account. They are keyed to the click ID, the tag the platform gives each click. Only conversions within 72 hours of the click are used to optimise campaigns. Since November 2025, there is also a free Shopify app [220, 221, 222] ◆. The system also takes a server-to-server postback, keyed to its click ID. A postback is a message sent straight from the advertiser's server to the platform. Order value, currency and order ID are optional fields [223]. These are legacy Outbrain settings. They are not a window for every current Teads product. Teads also publishes a server-side conversion API template for Google Tag Manager. It supports consent mode. That page does not document its endpoint, fields or deduplication [224]. Taboola's server-to-server path is a postback keyed to its click ID. It carries an event name, plus optional revenue, currency, quantity and order ID [225]. Taboola's own docs say server-side tracking covers attributed conversions only. That limits audience building and optimisation, compared with the pixel [226, 227, 228] ◆. Meta's Conversions API is the reference design. It has server events, event deduplication, hashed customer IDs and offline events. These work for both measurement and audiences. By contrast, the native vendors' public docs mainly show a postback with a click ID and an event [229, 230, 226, 220] ◆. On the public docs, the gap with a conversions API is identity. The native postbacks carry a click ID and value fields. They carry no hashed customer IDs. Parity is not shown. What integrations offer under contract may differ.

What comes back out is aggregate reporting. Taboola's Backstage API returns campaign summaries by day, campaign, site, country, platform and item. It takes uploads of first-party audiences, by email or device ID. Outbrain's Amplify API exposes campaign and performance data. It is "available for a select number of partners by request". None of the public docs read offer a way to export audiences, click-level logs or model learnings [231, 232, 233, 234, 235] ◆. That is not proof that no customer can get them. An advertiser keeps its own click IDs and conversion records. Contracts can grant more than help centres describe. On the public record, though, the model's learning stays with the platform. That is the dependence the research brief asked about. The advertiser's conversions train the platform's model. What the docs show coming back is a summary report.

Affiliate incentives and the rules that bind them

Content-to-commerce runs on affiliate commissions. The disclosure rules tightened in 2023 and 2024. The revised FTC Endorsement Guides took effect on 26 July 2023. Under them, a disclosure must be hard to miss and easy to understand. In interactive media it must be unavoidable. The FTC struck the word "small" from its affiliate-link example. So any affiliate pay needs disclosure. Staff guidance says "affiliate link" alone is not enough [236, 237] ◆. The FTC's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials came next. Adopted 5–0 in August 2024, it took effect on 21 October 2024. It bans fake or false reviews and testimonials. It bans buying reviews that express a particular sentiment. It also bans certain insider reviews that are not disclosed. It bans review sites a firm controls but presents as independent. And it bans suppressing reviews. Knowing breaches carry civil penalties. The rule does not ban AI-assisted writing as such, or all incentives. Incentivised reviews stay lawful if the reward does not hinge on sentiment [238, 239] ◆. A publisher commerce section that ranks products by commission, not merit, is now exposed on two fronts.

The newest surface: ads inside the answer

Paid placement inside AI assistants fits the native definition most clearly. That is because the ad sits beside or inside the answer. In 2026, several outcomes arrived. They do not point one way.

OpenAI announced plans for ads in ChatGPT on 16 January 2026 [240]. It began testing them on 9 February. The test covered logged-in US adults on the Free and Go tiers. Plus, Pro, Business, Enterprise and Education were excluded from the initial test. The ads were labelled and sat below the answer, apart from it. OpenAI said that ads do not affect answers [241, 240]. They gained a self-serve Ads Manager in May. By September they were reported in more than 40 countries. An OpenAI statement was reported on 1 September 2026. It claimed a $1 billion annualised revenue run rate, reached in under 200 days. It also claimed tens of thousands of advertisers. It listed product feeds, location targeting, custom audiences and outcome-optimised bidding [242, 243] ◆. A run rate scales recent revenue up to a full year. It is not what was earned over a year. On 16 September OpenAI said it was testing "Sponsored Agents" with select US advertisers. After clicking an ad, a user can start a clearly labelled chat with an agent the business sponsors. OpenAI says that chat is kept apart from ChatGPT's own answers and from the user's original conversation [244]. Its help centre calls this a limited alpha and was not taking requests for early access. No technical documentation for advertisers was published [245]. Trade press named Wayfair as a pilot advertiser [246]. The revenue and advertiser figures are the seller's own, and they reach this paper through reporting. OpenAI's own pages were opened in the second review round, after an earlier access error; they describe the products but say nothing about how well they work. Perplexity had shown sponsored follow-up questions in the United States from November 2024. They were labelled and sat beside answers [247]. The firm phased the programme out in 2025. In February 2026 it said the programme would not return. In its words: "A user needs to believe this is the best possible answer, to keep using the product and be willing to pay for it" [248, 249] ◆. One assistant found an ad business. Another concluded that ads and trust do not mix. No third party has measured either outcome. Research on how well users spot ads inside AI answers is thin. A 2024 study of users found they often missed product placements in generated search answers [250]. Work on machine detection shows sentence-transformer classifiers catch generated native ads with precision and recall above 0.9. The language models themselves struggle at the task [251] ◆. No representative study was found.

The native vendors are moving into the same surface from the publisher side. In June 2026 Taboola opened DeeperDive as an ad engine. DeeperDive is the generative AI answer engine Taboola runs on publisher sites. It puts ads into AI results pages. Taboola also offered the same ad model to other AI firms. The scale figures are Taboola's own. It claims tens of millions of answers a month, for more than 7 million users [51]. Teads has had an SDK in public beta since November 2025. It places native units inside chat interfaces. An MCP integration is listed as coming soon [252]. In March 2026 Dianomi announced a tie-up with Dappier. The plan is to embed a chat-style answer engine, with brand-agent ad units, on publisher sites. By mid-2026 it was still in talks with publishers [253]. Microsoft says ads in Copilot drew 73% higher click-through than traditional search between February and May 2025. That is a first-party figure from the vendor [254]. Amazon made sponsored-product prompts generally available in the United States on 25 March 2026. Clicking one can open a chat in its shopping assistant, Rufus. It was renamed Alexa for Shopping in May 2026 [255, 256]. None of this means all assistant answers carry ads. It means the economics of the recommendation unit are being rebuilt inside answer pages. That unit is the block of suggested links under an article.

Google is moving in small steps. After May 2025, ads in AI Overviews reached desktop and entered AI Mode. In May 2026 the firm announced formats "instantly tailored to a person's unique query". It also expanded the Direct Offers pilot it had introduced in January 2026. It gave no uptake or revenue figures [257, 258] ◆. The commerce rail is separate from the ads. OpenAI and Stripe co-maintain the Agentic Commerce Protocol, which is labelled beta. It powers Instant Checkout. That launched in September 2025. It came with the statement that results "are not sponsored, and ranked on relevance alone" [259, 260, 261, 262] ◆. That statement came before ChatGPT Ads. OpenAI's 2026 commerce docs refer to separate rules for ads and checkout. Whether the two stay apart is open at the cutoff.

Chapter 9

The vendor field: who can document what

24 vendors scored on 12 dimensions. One summary index, built the way the author's Contextual Quotient is built. Nine deep profiles, each with a SWOT. And the wider universe they came from. The matrix measures what a vendor can document in public. It does not measure how well its ads perform.

How to read the matrix

Each cell scores one capability from 0 to 5, capped by the evidence. A 5 means a filing, an official document or an independent observer shows in detail how the capability works. A 3 means it is real but thinly evidenced, narrow or in beta. A 0 appears only where a source documents that it is absent. If the record gives no basis to score, the cell says n/e. That is not a zero. If the dimension does not fit the business, the cell says n/a. One case is conversion data at a pure exchange, where it sits in the buyer's own buying tool. Conversion data is the record of sales, sign-ups and other actions that follow an ad.

Each scored cell carries the reason for its score, an availability label, an evidence grade, a date and its sources. The web edition opens them from the cell. The long-format CSV in the appendix lists them all. Vendor marketing can earn at most a 3, and vendor documentation at most a 4. So a large help centre does not buy a high score. The coverage column shows how many of the twelve dimensions could be scored. That keeps a thin record in view.

A quotient, and why not a sum

The matrix has one summary column, the Native Quotient. It uses the method the author built for the Contextual Quotient in his contextual paper, so the two indices share a construction and a way of reading them. Their scores and bands are relative to different capabilities and samples and are not interchangeable. Sharing a construction does not validate the method.

Adding up the twelve scores would be wrong. A sum counts each n/e cell as zero. So it punishes a company for a thin public record, not a weak product. It would also reward breadth over depth. Averaging only the scored cells is the opposite mistake. Take a vendor documented on four dimensions, all of them strong. It would beat a vendor documented on all twelve. Missing evidence would become an advantage.

NQ does three things instead.

1. It groups the dimensions. Depth is what the engine does with an advertiser's goal. It covers optimisation, conversion data, creative tools, commerce and workflow. Reach is where it can act and through which routes. It covers supply, formats, interoperability and geography. Standing is whether buyers and publishers can check and trust it. It covers measurement, supply transparency and publisher controls. Each group is the mean of the scored cells inside it. So an n/e cell is skipped, not counted as zero.
2. It weights the groups 45/25/30. Depth counts most. That is because the buyer's question in this paper is whether a platform can pursue an outcome. Standing comes next, because the paper's central gap is proof. The weights are the Contextual Quotient's. They are a design choice, kept so the two indices read the same way. They were not re-derived for native.
3. It shrinks the result toward the field mean. A vendor with few scored dimensions is pulled toward the field average of 55. The pull grows as the record thins. This is a rule the author chose, not a statistical estimate: the twelve dimensions are different capabilities, not repeated measures of one thing. The strength is four extra field-average observations, a third of the twelve dimensions. A vendor with all twelve scored keeps three quarters of the weight on its own record; one with four scored keeps half. Below six scored dimensions the record is too thin to band at all, so those rows get a score but no band.

The bands are set from the field mean. The field mean is the score of a vendor with exactly average evidence. Because every score is pulled toward it, even a perfect record does not reach 100: at the current field mean, a vendor scored 5 on all twelve dimensions would get NQ 89. Front rank starts at 67, Strong at 56 and Mid at 47. Scores below that are Narrow. The three bars under each score in the matrix are Depth, Reach and Standing, in that order.

Shrinkage damps a sparse record. It does not make missing evidence harmless. The build tests this directly: it turns each of the 168 scored cells into n/e, one at a time, and recomputes. Leaving out a low score can lift a vendor by more than the extra shrinkage takes away. If MGID's commerce capabilities score of 1 were unknown, its NQ would rise from 62 to 66. Across the field, one missing cell moves a score by up to 6 points up or 6 down, and 13 vendors have at least one cell that moves them by three or more. That is why n/e stays n/e and is never treated as zero, and why the coverage behind each score is shown next to it.

What survives every weighting

The top capability score goes to the platform moving fastest beyond the recommendation unit. That unit, or widget, is the box of suggested links often shown under a story. Taboola has the highest Depth and Reach in the field. Most of those scores come from Realize. It brings conversion bidding, server-side conversion feeds, and native video and carousel formats. Its display formats are not counted as native. Chapter 3 shows the company branching out from the widget. It is moving into display, and into AI answers through DeeperDive. Chapter 11 shows its revenue guidance falling while profit guidance rose. So the strongest documented native capability belongs to the company building most beyond the widget. Its filings do not say how much of its growth those new lines carry.

Standing is where the field is thin. Measurement support never scores above 3. It is unscored for two-thirds of vendors. Supply transparency is scored for fewer than half. Revcontent has the lowest Standing of the profiled vendors. It names no third-party verifier. The platform measures its own results. Dianomi shows how much a thin reading costs. Its first scores were low. They rose once the verifier read its help centre in full. There the verifier found conversion tracking, DoubleVerify and IAS integrations and publisher-level reporting. Readpeak scores well on Standing for a small company. That is because it lists all its publishers before a campaign runs. The two large platforms do not.

The exchanges hand the outcome to the buyer. TripleLift and Sharethrough score on supply, interoperability and transparency. But conversion data and outcome bidding sit in the buyer's DSP, the tool the buyer uses to buy ads. So those cells are unscored or low. TripleLift's Depth comes mainly from commerce units and its deal tool. That describes their model. It is not a defect in it.

The large platforms' native products are thinly documented as native. Google, Microsoft, Meta, Amazon and AppLovin each have three or four scored dimensions, too few to band. That is because their public material describes native as a format inside a much larger product. Their NQ sits near the field mean because of how the score is built. It says little about them.

Deep profiles and SWOT

Nine companies were profiled in depth. Two are the listed, scaled platforms: Taboola and Teads Holding Co. Two are the independent networks MGID and Revcontent. Dianomi is the finance specialist under offer from Taboola. Readpeak is a European regional platform. Nativo is now part of Life360. The last two are the exchanges with native roots: TripleLift and Sharethrough. Sharethrough is now part of Equativ.

Each profile covers what the company sells, who buys it and how it earns. It covers what the company depends on. It lists each capability with its availability and evidence grade. It gives the performance evidence and who paid for it. It also sets out limits, risks and fit.

Each profile opens with a SWOT. Each line cites its sources and carries a tag. Record means a filing, contract, regulator document, study or independent report. Vendor documentation means the company's own help centre, specifications or product pages. Vendor claim means a performance or scale figure the company states. Inference means the author's reading of dated facts. Most opportunities and threats are inferences by nature. A gap in the public record shows up as a weakness only in the form "not documented publicly as of September 2026". It never appears as "cannot". The web edition lets a reader pick one quadrant and compare it across all nine.

The wider universe

The matrix is drawn from a wider list of 68 companies. For each one, a primary page was opened. The list covers the core market, nearby substitutes and new surfaces. It is a reasoned sample, not a census. Adult-only, gambling-only and ad-fraud networks were left out on purpose. Companies that surfaced only as names were listed but not scored. A company was scored if two tests held. First, it sells or delivers open-web native as a main or clearly separable business, or native is bought through it at scale. Second, at least four of the twelve dimensions had evidence, counting a dimension documented as not applicable.

Chapter 10

Enterprise adoption

Enterprise buyers can now see more of the programmatic supply chain than at any time since 2016. They can see almost none of the native one. What they can check, test and take with them decides if native gets a line on the plan.

This chapter turns the evidence into real choices. They are the ones a buyer, an agency and a publisher face. It treats the author's Performance & Native playbook as a set of claims to test. Those claims are about how enterprise buying works. It also re-checks the playbook's market facts against primary sources at the cutoff.

Who buys, and through what

At the cutoff, the record shows four enterprise buying routes. Managed service, run by the vendor's own sales and key-account teams. The vendor's self-serve buying tool. At Teads it sits under master service agreements with holding companies. At Taboola it is Realize, with an optional Claude Skill front end. DSP and programmatic links through OpenRTB Native. And, since June 2026, publisher-side native supply open to SSP demand through Teads' EngageOS with Magnite's demand server [46, 1, 153, 265, 266, 158, 267, 268] ◆. A DSP is the software buyers use to bid for ad space. An SSP is the tool sellers use to offer it. Taboola's and Teads' product pages state no spending floor. The ranges quoted for them come from third-party blogs. MGID's help centre sets a $100 minimum deposit for self-serve accounts. It recommends $650, and its blog gives a $50 minimum campaign budget [269, 270].

The two scaled platforms describe different customers. Taboola's annual report counts about 2,200 Scaled Advertisers, working with it directly or through agencies, and says a significant majority of its revenue came from those working with it directly rather than through an agency. Its ten largest clients made up under 10% of revenue [1] ◆. Teads describes master service agreements with agency groups for its buying tool. It also describes a team just for large global brands. And it cites about 50 joint business partnerships. They average $4 million of spend a year. In 2025, 93% of its revenue came direct from buyers on its own platforms [46] ◆. Its second-quarter 2026 results listed renewals with Stellantis, LVMH, Warner Bros. and Dyson. The same release suspended guidance, its public forecast. The reason was swings in direct-response and small-business demand [271] ◆. Teads reported the renewals next to weakness in direct response and small business. Its reports do not show enterprise revenue or profit on their own. No single "holding-company corridor" story fits both.

What the playbook got right, and what has moved

Twenty market and competitive facts in the playbook were re-checked against primary sources. Most structural claims hold. Outbrain closed its purchase of Teads on 3 February 2025 ◆. The firm became Teads Holding Co. in June [45, 273, 46] ◆. Omnicom closed its purchase of Interpublic on 26 November 2025 [274] ◆. GroupM became WPP Media on 28 May 2025 [275, 276] ◆. Life360 closed its purchase of Nativo on 2 January 2026 [277, 59] ◆. Reddit's Max campaigns entered beta in January 2026 ◆. They opened to all advertisers in September 2026 [278]. Taboola's Realize+ entered expanded beta in the second quarter ◆. The Trade Desk shipped Kokai Zuma on 27 August 2026 [279, 153, 280] ◆.

Several details are stale or unsupported. At the cutoff, Omnicom Media's roster is OMD, PHD, UM, Initiative and Hearts United. Hearts United combines Hearts & Science and Mediahub. It launched on 28 August 2026 with about $9.1 billion of 2025 billings. The playbook lists the two as separate brands. Omnicom's own site still shows Mediahub on its own [281, 282, 272] ◆. The playbook names Annalect and Acxiom as Omnicom's data backbone. Omnicom Media's site names Omni, Acxiom Real ID and Flywheel. It does not mention Annalect [282] ◆. Trade press says Omnicom retired the Annalect and Kinesso brands in mid-2026 [283]. The playbook calls Criteo the subject of acquisition speculation. Criteo's own releases show it moved its legal home from France to Luxembourg. The move was completed on 29 July 2026. The releases also show an approved move to a US legal home, expected in January 2027. They make no reference to a sale [284, 272] ◆. In July 2026, Bloomberg reported a takeover approach from Vista Equity Partners and Quinti Capital. Criteo declined to comment. So the playbook's "acquisition speculation" matches the public record [285]. CitrusAd now trades as Epsilon Retail Media [286, 287, 272] ◆. And the line "MFA's direct share has fallen sharply (ANA)" rests on two measures. They cannot be compared. Both are discussed below.

What a buyer can inspect

The ANA's programmatic transparency work is the main source on what a large buyer can see. It was built on log-level data, not on vendor reports. Log-level data is the record of each impression bought. Its 2023 study covered 21 marketers. It found made-for-advertising (MFA) sites took 21% of impressions and 15% of spend ◆. Their CPMs, or prices per thousand impressions, were 25% below other sites. And they passed standard screens for viewability, invalid traffic and brand safety [89] ◆. Some marketers then opted into the ANA's quarterly benchmarks. The ANA's December 2024 release said their average MFA spend fell to 6.2% in 2024. The ANA's own first-quarter 2025 report restates that value as 1.1%. It then sat at 0.4% to 0.6% of spend through 2025. It rose to 1.1% in the first quarter of 2026. That was roughly flat across its performance groups. That quarter the ANA named "AI slop" as an emerging sub-type [288, 289]. The ANA says its waterfall, average and median values cannot be compared [94, 95, 96, 97, 98] ◆ ◆. Private marketplaces are invite-only auctions. In the third quarter of 2025, 81.6% of participants' combined spend went through them. For the median buyer it was 88.8%. Their median domain count fell to 26,372 [97, 290] ◆.

Three things stop that series from meaning "solved". First, the participants, from 21 marketers in the 2023 study to 86 in the first quarter of 2026, chose to join the log-level benchmarks. So it is a result for that group. The top quartile still had up to 27.4% of web spend on MFA sites. That was in the third quarter of 2025 [97] ◆. Second, the market-wide series from Jounce Media stops in mid-2023, at about 30% of web auctions. No market-wide figure for 2025 or 2026 was accessible [89, 291] ◆. Third, none of it measures native supply. Recommendation platforms are walled-off buying routes. They do not pass through the DSP logs the ANA studied.

For native, transparency has to come from the platforms. Their own docs show what they offer. Both give reports by placement after the ads run. Outbrain and Teads offer publisher and section reports "once your campaign has gathered some data". By default, a campaign can exclude up to 30 publishers and 100 sections. Realize has a sites tab with per-site results and a setting to block sites. It has a tool called SpendGuard that caps spend by itself. It also has a pre-bid brand-safety setting whose supplier is not documented [292, 293, 294, 295, 296] ◆. Neither native self-serve platform documents a list of sites before the campaign. Outbrain DSP, though, accepts publisher include lists, which name the sites to buy [264]. A smaller regional platform shows it can be done. Readpeak's docs show a full named list of its publishers on offer before a campaign. They also show per-site bids and spend limits [297, 298]. The ANA's own advice is the reverse of what these controls support. It favours inclusion lists over exclusion lists. Exclusion lists, it says, are "largely ineffective" against domains that appear daily [89] ◆.

Taboola's answer to MFA exposure is a curated tier called Taboola Select. The tier is described as premium editorial inventory "free from MFA properties, as verified and ensured by Jounce Media". Taboola publishes no MFA claim for the rest of its network. Its own policy page defines MFA sites as ones that "use paid ads on major platforms to lure real humans". The ANA and Adalytics show its widgets playing that role [299, 300] ◆. Both vendors publish landing-page rules. Teads sets numeric caps. It allows no more than five ad placements and one video player per screen. It bars auto-refresh above three times a minute [301] ◆. Realize requires an "Advertorial" label, sponsor disclosure and contact details. It bans false-urgency timers [302, 303] ◆. Neither publishes figures on rejections, takedowns or audits. So having a rule does not show it is enforced.

What a buyer can test

Chapter 7 showed that no randomised incrementality study of open-web native is public. Incrementality means the extra results the ads caused. So the test has to be the buyer's. The decision tool below starts from the question a buyer is really asking. It matches that question to the test that answers it. It sets out what each test must assume, where it fails and what it still cannot say.

The contract matters as much as the test. Under the ANA's contract guidance, an agency must disclose when it acts as principal rather than agent. That is when it buys media itself and resells it. Affiliate definitions must reach the top of the holding company. Mark-ups on principal inventory must be capped. The advertiser must get access to transaction data. And auditor agreements must be attached as contract exhibits [305, 306] ◆. For a native buy, the matching asks are narrower and more specific. The first is the attribution window and the realised lag distribution behind reported conversions. A conversion is the action the platform counts as a result. The window is how long after an ad a conversion can still be credited to it. The lag distribution shows how long the reported conversions actually took. Both should be disclosed before the campaign, as the MRC's outcomes standard requires of measurement providers and the retail-media guidelines require of retail media [183, 184]. If the platform prices per acquisition, ask how an acquisition, or an incremental action, is measured. The second ask is the site list, before the ads run. The third is the definition of the conversion the platform optimised to. The docs show this can be a page view when tracking is absent [145]. The fourth is an export of the placement log. The windows and conversion definitions are documented, at least by Taboola [181]. Three items were not found in public documentation. One is the lag distribution. Another is a pre-delivery site list at the two large platforms. The last is a placement-log export. Buyers may be able to get them by contract.

Guarantees, packaging and the outcome promise

The playbook argues native must be sold as outcome infrastructure to reach enterprise budgets. The contracts looked at here run the other way. In both scaled platforms' filings, the only guarantee language is about publishers. The platform promises them a minimum per thousand impressions [46, 1, 266, 158, 267] ◆. The clearest guarantee to buyers found in the wider market is Vevo's "Attention Guaranteed" with Adelaide. Announced in January 2026, it promises a minimum attention score, not a sale. It discloses no make-good, meaning what the seller gives back if the promise is missed [307, 308] ◆. In May 2025, Teads' own blog described a pilot. It was for "Attention-to-Outcome Guarantees" built on Adelaide scores. No terms, markets or current status were published [309]. An attention guarantee is not an outcome guarantee. The playbook's phrase "attention-to-outcome guarantees" overstates what has been promised by contract in public.

The signal from buyers is mixed. A Taboola-sponsored survey covered 200 senior marketers at large US and UK firms. It found 80% would spend more on the open web if it had the automation of walled gardens. Those are the big closed platforms [153, 310] ◆. Stated intent is cheap. In the same quarter, Teads suspended guidance. The cause was volatile direct-response and small-business demand. The evidence supports two points. Enterprise budgets for native follow proof that the buyer's own measurement team will accept. And the two large native platforms do not document the tools such a team would need. Those tools are a holdout, a disclosed lag distribution and a pre-campaign inventory list. A holdout is a group kept from seeing the ads.

The publisher's decision

For the publisher, adoption is a decision about where the widget sits on the page. The recommendation widget is the block of paid and editorial links, often below a story. The cases below show financial incentives that can conflict with quality and with control over placement. The record shows publishers turning up clickbait in the units when they could [79] ◆. It shows them swapping $1–2 million guarantees for revenue shares. They did it to gain freedom over placement [79, 311] ◆. In one documented case, a publisher ran a paid-traffic subdomain for seven years. One observed session on it carried more than 200 ads [92, 93] ◆. Against that, the widget is one of the few income lines that need no sales team. Teads' EngageOS proposes one auction for ads and for links to the site's own stories. The aim is to optimise a whole session's yield. No adoption or lift figures have been disclosed yet [265, 271] ◆. Jounce Media has a "cheap reach" category for ad space under 15% viewability. It includes widgets below the fold, where the reader must scroll to see them. Jounce names Outbrain and Taboola among the operators. Buyers called it a necessary evil [312] ◆. A publisher choosing a native partner in 2026 faces a choice. One path is guaranteed pay, with exclusivity. The other is flexible pay, without it. Chapter 4 showed which way the money has been moving.

The rules, in three jurisdictions

The legal floor under all of this asks for much the same things. Where it differs is enforcement ◆. In the US, Section 5 of the FTC Act sets the test. An ad format is deceptive if it materially misleads reasonable consumers about the ad's commercial nature or source. Deception is judged on the ad's overall net impression ◆. The burden falls on "everyone who participates directly or indirectly in creating or presenting native ads". That includes publishers, agencies and affiliate networks [125, 313, 2] ◆. Lord & Taylor, settled in 2016, is the best-known FTC native case. It involved a paid magazine piece and about 50 undisclosed influencer posts. But it is not the only one. In February 2019 the FTC approved final orders against Creaxion and Inside Publications. Inside Publications had run paid ads disguised as articles in its magazine [314]. Enforcement since 2023 has gone to endorsements and reviews. The Consumer Reviews and Testimonials Rule was pleaded in a 2026 order. The FTC's Legal Library lists roughly ten review-related matters dated 2023 to 2026, by last-updated date; one, the Rytr order, was set aside in December 2025 [315, 316, 317, 318] ◆ ◆ ◆. This paper did not run a defined case search. So it makes no claim about enforcement trends. Section 5 of the FTC Act is the statute. The 2015 policy statement and guide, and the Endorsement Guides, are the agency's reading of it. The reviews rule is a trade regulation rule.

State privacy law governs behavioural targeting of native units. California's CCPA as amended covers "sharing" personal information for cross-context behavioural advertising. That means ads based on what a person does on other firms' sites and apps. A business that sells or shares personal information must honour valid opt-out preference signals. It generally must also post the "Do Not Sell or Share" link. It may leave the link off only if it processes the signals without friction and makes the related disclosures, as the regulations define. The regulations say plainly that the link is not an alternative to honouring the signal [319] ◆. Since the first regulations took effect in 2020, a compliant browser signal has counted as an opt-out of sale. Since 29 March 2023 that also covers sharing, for the browser and any linked profile. An amended package took effect on 1 January 2026 [320, 321, 319, 322, 323] ◆. From 1 January 2027, California's Opt Me Out Act applies. It makes it unlawful to maintain a browser without an opt-out signal that consumers can set [324] ◆. As of 24 November 2025, nineteen states had signed comprehensive privacy laws, by the IAPP tracker's convention. Of these, seventeen had an explicit right to opt out of targeted ads. Counts differ by date, and by whether narrower laws such as Florida's are included [325, 326] ◆. By the cutoff, at least two more states had signed. Oklahoma signed in March 2026, and Alabama in April 2026. Both laws take effect in 2027 [327, 328].

The EU bans undisclosed advertorials outright. An advertorial is an ad made to look like a story. Undisclosed advertorials are item 11 on Annex I of the Unfair Commercial Practices Directive. Everything on that list is unfair in all circumstances. Item 11a adds undisclosed paid search results or ranking [329] ◆. The Digital Services Act has set rules for online platforms since 17 February 2024. Micro and small firms are exempt. Platforms it covers must label ads clearly with who placed them and why the user sees them. They may not show ads based on profiling that uses special categories of personal data. Nor may they do so when they are aware, with reasonable certainty, that the user is a minor. Very large platforms must also keep a public archive of their ads [330, 331, 332] ◆. The AI Act's Article 50 transparency duties applied from 2 August 2026. They cover telling people when they deal with AI. They cover marking generated content so machines can read it. They also cover deepfakes. And they cover AI-generated text published to inform the public on matters of public interest. Such text is exempt if it has had human review or editorial control. Someone must also hold editorial responsibility for it. So they do not require the same label on all AI-assisted ad copy [333]. The AI Omnibus delayed the separate high-risk duties to December 2027 and August 2028. It also gave generative systems already on the market more time. They have until 2 December 2026 to meet the Article 50(2) duty. That is the duty to mark content so machines can read it. The other Article 50 duties applied from 2 August 2026 [334, 335]. No rule found expressly governs paid placements inside AI assistant answers. There, the closest authorities are the paid-search precedents and the platform labelling duties [336, 337, 338, 339, 340] ◆ ◆ ◆.

The UK's CAP Code requires marketing to be obviously identifiable. Advertorials must be headed as such ◆. Schedule 20 of the Digital Markets, Competition and Consumers Act has applied since 6 April 2025. It makes undisclosed paid editorial promotion and fake reviews unfair practices. It lets the competition authority enforce this directly [341, 342] ◆. The ASA is the UK ad regulator. In February 2026, it upheld a complaint against a medicated weight-loss advertorial on Mumsnet. The ad had only a small "Created by" credit ◆. In April 2026 the ASA upheld another, against a paid search ad. The ad presented a trader's site as an independent review body [343, 344, 345] ◆.

Industry guidance sits below all of this. IAB's Native Advertising Playbook 2.0 of May 2019 requires a disclosure on every native type. It defers to the FTC's 2015 documents. It has no legal force and sets no exact wording. So the "Promoted" labels that the FTC calls ambiguous stay in use [3] ◆. The tool below lists each source with its effective date and scope. Nothing in this section is legal advice.

Chapter 11

What changes next

The two large native players on the open web are moving in different directions. One is spreading beyond the recommendation unit, or widget. That is the box of suggested links on a publisher's page. The other is being squeezed by the math of a format sold against page views. Agents, protocols and answer engines change who buys and where the ad sits. They do not change whether the click was worth anything.

This chapter sorts three things. It splits what was announced from what was done. It splits a retired brand from a retired product. And it asks what agentic buying can do for native beyond the optimisation chapter 6 described. Agentic buying means software agents do the buying.

Two trajectories, one quarter

The second quarter of 2026 gives a useful contrast in reported results. It is not a natural experiment, though. Nothing assigned the two firms their different conditions. Teads Holding Co. reported revenue down 17% to $284.6 million. Ex-TAC fell 14%. Adjusted EBITDA fell 74% to $7.0 million, and it posted a $42.5 million net loss ◆. Then on 6 August it suspended all guidance. That included the roughly $100 million adjusted EBITDA target it had reaffirmed a quarter earlier. It cited "the volatility of the Direct Response and SME business" and "open-web headwinds" [271, 346, 29] ◆. Its restructuring in late 2025 cut staff by about 10%. It replaced its chief commercial officer, chief marketing officer and North America managing director [29, 346] ◆. The company casts its enterprise and connected TV business as growing. It names direct response and small business as the problem. That points at the performance and widget lineage, the part that was Outbrain. Its segment reports do not show that lineage's results on their own.

Taboola reported the opposite. Ex-TAC gross profit rose 11.8%. Adjusted EBITDA rose 22.8% and beat guidance. It raised full-year guidance for both [30, 347] ◆. In the same release its revenue came in below its own range for the second quarter. Its full-year revenue range was cut by $76–106 million at the endpoints. Management did not highlight the cut [348, 347, 30] ◆. There is a precedent. In August 2024 Taboola cut its revenue guidance by about $170 million but held ex-TAC. It explained that some Yahoo revenue "will now only be reflected in ex-TAC" [349, 350, 41, 351] ◆. Management did not tie the 2026 cut to presentation. On the earnings call it gave two causes. One was leaving publisher ties, largely in Greater China. The other was Google phasing out its Explore More feature [352]. Take the definition used before the second quarter. On that basis the raised ex-TAC range would be about $760 million to $771 million. That is flat to slightly down against May. Three things could each push revenue and ex-TAC apart. One is more revenue shown net. Another, described in chapter 4, is the new non-cash add-back. The last is a shift in demand or mix. Management disclosed no bridge between them, so the paper does not choose.

The two firms also differ in how they disclose the shock from AI search. Teads' 2025 annual report says generative AI answers in search "reduced [users'] need to click through to original publisher websites". It says the resulting fall in partner traffic "has affected, and could in the future have a significant effect on, our revenue" [46] ◆. Taboola's filings never use the phrase "AI Overviews". Its 2026 quarterly reports do not contain "generative AI". Its 2026 releases say nothing about search traffic [353, 30, 347, 348] ◆. One reading is that the exposure differs, since Yahoo and Microsoft are portals with direct traffic. The other is that their disclosure habits differ. The record cannot tell them apart.

The traffic shock, measured where it can be

The pressure on the widget sits at the top of the publisher's funnel. People Inc.'s core sessions fell 5% over the full year 2025. In the fourth quarter of 2025, they fell 13%. In the first quarter of 2026 they fell 17%. In the second quarter they fell 22%. The company tied the fall "primarily to the impact of the growing prominence of Google AI Overviews" [354, 355, 356, 215] ◆. Its digital revenue still grew 6% in the second quarter of 2026, driven by performance marketing and a Meta content licence. Advertising held flat, as higher programmatic rates offset lower volumes [215] ◆. Ozone's benchmark data, as reported by Digiday, covers the second quarter. It puts publisher ad requests down 32–37% year on year in the United States. In the United Kingdom they fell 39–41%. In the UK, eCPMs rose 30% in June and made up part of the loss. An eCPM is the price paid per thousand ads shown. US programmatic spend fell 44% over the first half [357] ◆. Those are the alliance's own benchmarks of premium inventory, not a neutral measure. They suggest that a large premium publisher can reprice. The paper infers that a long-tail widget host, without that pricing power, cannot do so as easily; no data here measures it. Dianomi blamed its 2025 fall in impressions, 14% by its own total, on "AI-powered content discovery tools". It then grew 2% in the first half of 2026, from a lowered base [358, 359] ◆.

These series do not measure the same thing. People Inc. reports sessions, Ozone reports ad requests and Dianomi reports impressions. A session can hold more than one page. A page can send more than one request. A request may go unfilled. The paper infers that fewer sessions and requests mean fewer widgets shown. That is because the widget is sold against page views. No series measures widget inventory directly.

Retirements: brands, products and the difference

The record of failures has to tell a retired name from a retired product. Most of the "exits" in this market are the former ◆ ◆ ◆.

This project could open just one documented case of a publisher dropping a widget. It is nine years old. In 2017, Outside magazine replaced Outbrain with in-house widgets. Outbrain was about 10% of its digital ad revenue. The in-house widgets earned higher click-through. The piece that reported this also noted that Outbrain "accounts for 30% of revenue for some publishers" [374] ◆. The paper infers that dependence on revenue at that scale is one reason more publishers have not followed. No population-level series of widget removals was found.

Agents and protocols: shipped, demonstrated, announced

The question on AI and agentic buying asked what truly autonomous buying can do beyond current optimisation. The record at the cutoff has three layers. The map above labels each item as live, beta, announced, retired or unconfirmed. It gives the date and says whether anyone outside the vendor has seen it.

The protocol layer is real. It is also split in two, and the two stacks do not refer to each other. AdCP is governed by AgenticAdvertising.org, which calls itself a pending trade association. On 23 September 2026 AdCP showed a stable release line at v3.1.24. It showed a release candidate for 3.2 the next day. Its home page showed a third version string. Eight tags were cut in sixteen days [375, 376, 377] ◆. IAB Tech Lab's effort is AAMP, short for Agentic Advertising Management Protocols. It was listed as released on 26 February 2026 at the umbrella level. But its real-time part was still in public comment. Its audiences part had no release listed [378, 379, 380] ◆. On 22 September 2026 Tech Lab announced AAMP 3.0. Its new OpenProposal spec is in public comment until 22 October [381]. Neither set of pages mentions the other ◆. Only one open source build of AdCP for sellers was opened, Prebid Sales Agent. It calls itself alpha [382, 383, 377, 384] ◆. The group claims the first media buy between agents ran on 16 October 2025 with LG Ads. It discloses no buyer, amount or third-party observer [385, 375] ◆. The native format's own programmatic contract is OpenRTB Dynamic Native Ads. It was finalised as version 1.2 in March 2017; the Tech Lab index lists July 2017. The agent specs, by contrast, change each week [378, 379, 376] ◆.

The large native platforms do not appear in any of it. Taboola, Teads and MGID are missing from AdCP's member logos. Nor are they on the Prebid Sales Agent's list of ad servers. Take Taboola's and Teads' results for the second quarter of 2026. Neither mentions AdCP, AAMP, MCP or buying between agents [375, 382, 30, 271, 386, 387] ◆. AdCP can describe native. Its 3.1.24 docs refer to native formats, and its catalogue lists native entries. But a proposal for a standard native format is put off to the 3.2 track. It waits on an audit against OpenRTB Native 1.2 [388, 389]. On 23 September 2026, the commercial platform built by AdCP's co-creator was rebranded from Scope3 to Apostra. It lists native among its channels. It describes live campaigns in six countries but names no customer and no dollar figure [377, 384, 390, 391] ◆ ◆. The vendor universe lists Kargo among sellers with native roots that are now pitched as high-impact. Kargo is among AdCP's self-reported members [384, 385]. But joining is not building on it. No live buying through it was found. Agents may reach native supply through a middleman before any large native vendor adopts a protocol.

The vendor layer is mostly assistants and automation within limits. The tasks worth testing are separate. They are planning, tool use, budget changes, placement changes, making creative and watching campaigns. A marketing label does not say which of these are handed over. Taboola's Abby, announced in October 2024, guides campaign setup by asking questions. It also edits creative. Its one performance claim is that early campaigns went live 75% faster than manual setup. It gives no sample or method [392, 393] ◆. Realize+ was announced on 23 April 2026 after a beta. It is marketed as "an agentic system that turns advertiser goals into outcomes". The pitch is that it keeps making and carrying out campaign decisions ◆. The same release opened the platform to a Claude Skill for setting up and tuning campaigns [153]. Taboola's filings disclose no Realize, Abby or AI metric of any kind [351, 386, 393, 30] ◆. MGID's MIA, launched in July 2026, reads live account data and explains problems. And "the decision to act remains yours" [394, 267] ◆. MGID has offered AI image and text tools since 2024 [154]. Teads' release for the second quarter names predictive AI and EngageOS for publishers. And it names a bundle for connected TV [271, 387] ◆. Its developer site carries an SDK for conversational ads in public beta. It lists MCP support as coming soon [252].

The platform layer is where autonomy is closest. It is still gated. Amazon's Ads Agent builds campaigns from a media plan and tunes them in plain language. But "campaigns only launch after you review and approve" [395, 396, 397] ◆. Google showed advertiser agents at its May 2026 event. Its beta for a YouTube business agent came in September 2026 ◆. Next to these sits a Demand Gen claim. It cites a 30% average rise in conversions, with no stated baseline [258, 398, 399] ◆. Meta's Advantage+ case studies report large lifts for selected advertisers. Two of them come from lift studies. None covers the whole platform [400] ◆. The Trade Desk's results mention Kokai only in words, with no figures [401] ◆. Most systems whose docs were opened keep some human approval step. Amazon has its review step. AdCP has its governance check and audit log. MGID has its "yours to decide", and Taboola its question-led setup [395, 377, 394, 392, 375] ◆. Realize+ and Apostra describe the system acting on its own within rules the buyer sets. That is a weaker gate than a sign-off on each campaign. An approval step does not rule out the system acting alone inside set limits. What neither has published is what happened when it ran.

On this record, what agentic buying changes for native is the buyer. Take an agent that can read a seller's supply, quality, pricing rules and how the seller defines outcomes. It will choose sellers that expose them in a form machines can read. AdCP defines tasks to report what ran and to send back results. Amazon opens its clean room to plain-language queries. Among the native vendors profiled in chapter 9, only Taboola documents an interface for agents in the material this review opened. It is an official Realize MCP server that can read and change campaigns. Taboola published it in July 2026 [402, 403]. None of those vendors documents a way for agents to set goals or report outcomes beyond its own API [377, 395, 386, 387, 267] ◆. Wrapping an existing reporting API for agents is a small job for engineers. As positioning, it is a big one. No protocol changes the evidence problem from chapter 7. An agent that optimises to the conversions a platform reports is doing attribution faster. The incrementality question passes through untouched. That question is whether the ads caused conversions that would not have happened anyway.

Scenarios, and what would tell you which one is running

Three scenarios fit the record. Each names what it rests on and the case against it. Each names indicators to monitor. These are signals, not decisive tests. Where a direct test exists, the box names it and says whether the data are disclosed.

The paper's read is that scenarios one and two are running at once, at different firms. Scenario three has not started in public. That fits what chapters 3 through 7 found. The moat is supply. The platforms are building beyond the widget, though neither discloses how much growth those lines add. And public causal evidence is missing. The paper cannot say why it is missing. Plausible reasons include these. Tests may have been run and kept private. Measuring effects this small is costly and noisy. The incentive to publish may be weak. And one format is hard to isolate inside a multi-format platform. The paper does not choose between them.

Failure modes for this paper's own outlook

Three things would change the conclusions. The first is a made-for-advertising series for the whole market for 2025 or 2026. It would show whether widget arbitrage has collapsed or grown. No source this project could open provides one. The second is an incrementality result for native alone. Its design and results would be public. It would move the evidence chapter from absence to a finding. The third is Taboola's next annual report. It will show whether the 2026 revenue cut was presentation or demand. It will also show whether its filings begin to say what Teads' already say about AI search. All three are dated events a reader can check. The handoff note lists them.

Chapter 12

Method, limits and corrections

How the paper was built and what the verification pass caught. What could not be reached, and each change made after the challenge passes. It is published so readers can check the paper rather than trust it.

What the paper set out to answer

The paper asks one question. What is native advertising in 2026, who earns from it, and what evidence backs what buyers are told about it? The brief split that into twelve sub-questions, one per workstream. They were fixed before research started:

1. What counts as native? How big is it, by each measurer's own definition?
2. Who sells and distributes it? How concentrated is supply?
3. What does a dollar of native revenue pay for? What do publishers keep?
4. Do readers recognise it as advertising? Does the creative or the disclosure change that?
5. How do the recommendation and delivery systems decide what to show?
6. Is there causal evidence that native spend adds sales or conversions?
7. How clean and transparent is the supply?
8. How do commerce links and first-party data change the unit?
9. How do large buyers buy and control it?
10. What have AI and agentic buying changed? What is still announced rather than deployed?
11. Which rules on disclosure, privacy and AI apply? How are they enforced?
12. Where has the model failed? What would show which way it is heading?

Chapters 1 to 11 answer them, roughly in that order. The vendor universe and the nine vendor profiles answer a thirteenth, practical question. Who could a US buyer actually use, and what can each one document?

How it was built

The research ran as parallel streams from one brief. Before any research began, the brief fixed the definitions, the three scope groups, the source hierarchy and the evidence classes. Twelve workstreams covered definitions and market measurement, market structure, financial and publisher economics, and consumer and creative evidence. They also covered delivery technology, causal effectiveness, supply quality, commerce and data, and enterprise buying. Then came AI and agentic execution, disclosure and governance, and failure modes. A thirteenth stream built the vendor universe but did not score it. Nine agents built vendor profiles for Taboola, Teads Holding Co., MGID, Revcontent, Dianomi, Readpeak, Nativo, TripleLift and Sharethrough. Each stream wrote a structured evidence file. It lists sources with their dates of publication, event and access. For each source it also gives geography, period, funding and access limits. It logs findings, with the supporting passage, scope, evidence class, confidence and counterevidence. And it holds data tables for each number that might be charted.

Each file then went to a separate verifier agent. The verifier had not seen the research, and it was told to refute it. It picked the claims that mattered and opened the cited source. It checked denominator, unit, geography, date, baseline and population. It looked for a contrary or more primary source. Then it gave a verdict: supported as stated, supported with qualification, contradicted or unresolved. Verifiers wrote their own files and did not edit the research. The verifiers checked up to seventeen claims per file. So a verifier re-checked 317 of 642 ledger claims. Claims outside that sample are marked unresolved in the ledger. A note says they rest on the source the research pass opened. The numerical and citation review traced each number that matters in the prose back to its source. That review is described below.

Synthesis, scoring and drafting came next. This work was AI-assisted and in the author's voice, for his review before publication. The paper's argument was not fixed in advance. The brief's candidate theses were tested against the evidence. The headline findings in the abstract are the ones that survived. Several of the author's own starting positions did not, and the handoff note lists them.

How sources were found, and what an absence means

There was no bibliographic database search. There was also no formal screening protocol of the kind a systematic review uses. Each research agent worked from the brief's leads and the source hierarchy. Filings, regulators and standards bodies came first. Then came peer-reviewed studies, official technical docs, measurement publishers and the trade press. Vendor material came last. The searches in each workstream mixed vendor and product names, format terms and the workstream's topic. The format terms were native, in-feed, content recommendation, sponsored content and advertorial. Topics included traffic acquisition cost, ad recognition, incrementality, made-for-advertising and agentic buying. Searches ran across the open web, SEC EDGAR, regulator sites, and company developer and help centres. They also used arXiv, DOI resolvers and scholarly metadata services. Nothing published after 27 September 2026 was used. A search result was a lead, never evidence. Each cited source was opened, and the record notes how much of it was read.

Sources with limited access were handled the same way throughout. If a source could be read only in part, the register labels it. The label says abstract, partial extract, paywalled excerpt, or access error with text reached by other means. Nothing is quoted from the part that was not read. A study read only as an abstract supplies no effect size beyond what the abstract states.

That method limits what an absence can mean. When this paper says a document, study or feature was not found, it means a narrow thing. It was not found in the sources this review opened, in public material, by the cutoff date. It does not mean the thing does not exist. And it says nothing about private contracts. Where a vendor might hold the item privately, the text says so.

What the process is not

Two agents reading the same source are not two independent measurements. A model that checks another model's work is not an external audit. When this paper says a claim was verified, it means a second agent opened the source and read the passage. When it says a figure was recomputed, it means a script recomputed it from the source rows. The formula is in the derived-checks file. No third party has reviewed the paper. No scored vendor has seen it. Nobody funded or commissioned it.

Scoring ran as an AI-assisted first pass under the author's rules, then a revision once the verifier records and the final reviews were read, and a re-audit of the format scores after the second outside review. The corrections ledger lists every score that changed. The rules file was written before the first pass, but it carries no independent timestamp. There is no second scorer. So there is no inter-rater reliability statistic, which measures how far two scorers agree. This paper's central complaint is that native vendors grade their own homework. For such a paper, that is the biggest gap left in the method. It is named here rather than left for a reader to find.

Sampling and scoring, stated plainly

The vendor universe is a reasoned sample, not a census. It started from named seeds. It added each native specialist found while researching the twelve workstreams. A primary page was opened for each row. Adult-only, gambling-only and pure ad-fraud networks were left out on purpose. Some vendors surfaced only as names, with no source that could be opened. They were listed but not scored. The matrix scores 24 vendors. Each sells or delivers open-web native as its main business or a clearly separable one. Or it is an intermediary through which native is bought at scale. Each could also be evidenced on at least four of the twelve dimensions, counting a dimension documented as not applicable. Budget comparators such as Meta's own feeds and TikTok are compared on revenue in chapter 2; Meta Audience Network is scored as a publisher-side intermediary. They are not scored, because their business model is not comparable on these dimensions.

The twelve dimensions, their definitions and their score anchors were fixed in data/scoring-rules.json before scoring began. Scores run 0 to 5. They are bounded by the evidence, not absolute. A 5 means the capability is documented in operational detail. The source is a filing, an official document or an independent observation. A 3 means it is real but thinly evidenced, narrow or in beta. A 1 means adjacent or incidental. A 0 is used only where a source documents absence. The label n/e means the record gave no basis to score. The label n/a means the dimension does not apply to the vendor's model. Each scored cell carries an availability label. It carries an evidence grade, from A (filing or official document) to E (none). It also carries a date, a written reason and at least one source. The build fails if any of those is missing. The same standard applies to each vendor. A large help centre earns grade C, not a higher score.

The first draft published no composite score. The reason was that a composite mixes capability with how much evidence there is, and rewards vendors that document more. The final paper carries one, the Native Quotient (NQ). It was added after scoring, at the author's instruction. It is built the way his Contextual Quotient is built, so the two share a construction and a way of reading them. Their scores and bands are relative to different capabilities and samples and are not interchangeable. The design answers part of the objection, not all of it. It has three groups: Depth, Reach and Standing. Each group is the mean of the dimensions scored inside it. So an n/e cell is skipped, not counted as zero. The groups combine at 45/25/30. The result is then shrunk toward the field mean in proportion to how few dimensions are scored. Shrinkage damps a thin record, but chapter 9 shows it does not make missing evidence neutral. The dimension scores were fixed before NQ existed, and none was changed to fit it. The format scores were later re-audited against their anchor after the second outside review; those changes followed the anchor, not the index. Bands are reported instead of ranks, with thresholds set from the field mean. The reweighting test in chapter 9 shows why. Under four other weightings, positions move by up to 5 places. Coverage is the number of dimensions scored out of twelve. It is shown for each row so that thin evidence stays visible. Of 288 cells, 168 carry a score, 83 are n/e and 37 are n/a.

What could not be reached

Access limits shaped this paper. The register records them per source. Of 977 sources, 90 are recorded as not read in full; for a further 137, added by verifiers, the extent read was not recorded. The ones that matter are these. EMARKETER's pages returned access errors to direct requests. Its executive summaries were read by other means. So the 2024 value and any split after 2019 are missing. Chapter 2 says so. OpenAI's and Perplexity's own announcement pages could not be opened. So the ChatGPT Ads and Perplexity facts rest on text reached by other means or on second-hand reports. They are labelled that way. EUR-Lex rejected requests during the research pass. The Digital Services Act articles were later read in the Official Journal text. The AI Act's Article 50 was read from a word-for-word unofficial copy. The AI Omnibus was read from Commission material. Several academic papers could be read only as abstracts. Several SEC filings were too long for the fetch tool. It returned the business and risk sections and only parts of the financial notes. So the reconciliations drawn from them are marked derived.

One tooling limit hit the later streams. The research agents shared a web-search quota, and it ran out part-way through the run. So the governance, failure-mode, universe and profile streams reached sources by direct address, not by search. So did all the verifiers. So discovery in those streams is narrower than in the earlier ones. The completeness review below treats that as a known bias rather than pretending it away. Two fetches of PDF documents returned summaries with tables that were not in the documents. Each figure from a PDF in this paper was re-extracted word for word by script. The incident is recorded so a reader knows the class of error the process had to guard against.

The challenge passes, and what they changed

After the verifier pass, an interim draft got a separate, detailed review with one hundred findings. Each finding was checked against its cited source where one could be reached. Most were accepted. A few were accepted in part, with reasons. The changes are in the corrections ledger, tagged with the review's finding number. That review is a critique of a draft, not an audit or a peer review. Five reviews were then run on the final text. Their findings, and what was done about each, are in the corrections ledger below, tagged F, and in the handoff note. A fact review checked each number that matters against the source record. It did the same for each date, entity, ownership claim, product status, technical claim, citation and superlative. An argument review attacked the central thesis. It asked whether the scope was chosen to favour it, and whether native and adjacent markets were lumped together. It asked whether a weak comparator flattered the result, whether correlation became cause, and whether counterexamples changed the reading. A scoring review challenged the sample, the dimensions, evidence consistency and the treatment of missing data. A completeness review looked for missing stakeholders, formats, geographies, failed businesses and contrary studies. A numerical and citation review reran each derived calculation and resolved each reference ID. It checked that each cited page supports the sentence it is attached to. It also matched each figure to the canonical dataset. A second outside review of the rebuilt package returned 38 findings, tagged R2 in the ledger. Where a finding made a factual claim, it was checked against the primary source before anything changed. One was rejected: Taboola files a 10-K, not a 20-F. The rest led to corrections in the text, the scoring, the NQ method, the print edition and the package. Two research leads in the brief itself were wrong, and are recorded. Sharethrough's merger with Equativ was announced in June 2024, not 2023, and Adyoulike was bought by OpenWeb, not Opera.