Blog · Engines

Ads in AI answers: what they buy and what they do not

Published 15 October 2026 · 10 min read

The short answer: Paid placement inside an AI answer buys a labelled slot beside the answer, not a place in it. OpenAI says ads sit below a response, clearly labelled, on separate systems from the chat model. Google says ads can run above, below or within an AI Overview, with no way to target or opt out of that placement, and reported as Top Ads. Neither says buying a slot makes you more likely to be cited organically, and nobody has published evidence that it does.

That last point is the one buyers keep asking about, and the one with the least published material behind it.

What OpenAI sells, in its own documentation

OpenAI set out five ads principles on 16 January 2026, one of which it calls answer independence: ads do not influence the answers ChatGPT gives, and they are separate and clearly labelled. Testing began in the United States on 9 February 2026, for logged-in adults on the Free and Go tiers, with Plus, Pro, Business, Enterprise and Education accounts outside it. By 23 September 2026 OpenAI said ChatGPT Ads ran in more than 60 countries.

The unit itself is conventional. OpenAI's advertiser help article lists six components: advertiser name, favicon, title, copy, landing page and an image asset. It sits below a ChatGPT response, labelled as sponsored and visually separated from it.

Buying works the way search buying works. The same article describes impression and valid-click billing, a Views objective bought on CPM and a Clicks objective bought on CPC, a maximum bid set at ad group level with a suggested starting point of $3 to $5 USD per click, and what OpenAI calls a relevance-weighted, second-price auction. Advertisers supply context hints describing topics where their product may be relevant, and OpenAI states plainly that these are not exact-match keywords and do not guarantee delivery in any particular conversation.

Reporting in the beta covers impressions, clicks, spend, click-through rate, average CPC, average CPM and conversions. What advertisers do not receive is anything about the conversation: OpenAI's user-facing help page says they get aggregated, non-identifying performance data and no access to chats, chat history or memories.

Placement is bounded by policy rather than by bid. The ad policies page, updated 10 September 2026, limits the initial test to consumer verticals such as lifestyle and household goods, local services, travel and experiences, and digital products or education, with financial services, healthcare and medicine, and legal services reviewed case by case and everything else disallowed at launch. Ads are also kept away from what OpenAI calls sensitive user contexts, including mental and personal health conversations.

One crawler reads your landing page, a different one reads your site

This is the detail that connects advertising to the rest of answer-engine work, and it is a line in OpenAI's bot documentation rather than a theory. OpenAI documents four user agents. OAI-AdsBot/1.0 is the fourth, and its job is narrow: it validates the safety of pages submitted as ads on ChatGPT, and OpenAI says it may also use landing page content to decide when an ad is most relevant. It visits only pages submitted as ads, what it collects is not used to train foundation models, and its IP ranges are published at openai.com/adsbot.json.

So your landing page is fetched and read for ad relevance by an agent that has nothing to do with the one building the search index behind organic answers. The advertising pipeline and the citation pipeline touch your site through different doors. Which doors are open, and what each controls, is the subject of the AI crawlers and user agents worth knowing.

Google: a placement you cannot choose, and cannot refuse

Google's advertiser help page on ads and AI Overviews is unusually blunt about control. Ads are eligible above, below or within AI Overviews. Within the overview, eligible formats are text and Shopping ads from existing Search, Shopping and Performance Max campaigns, and the system considers both the user query and the content of the overview. Above and below placements are live wherever AI Overviews are, which Google describes as 200-plus markets. Within-overview ads are narrower: English, mobile and desktop, in twelve countries including Australia, Canada, India, Nigeria, Singapore and the United States.

Two answers on that page belong in any planning meeting. You cannot directly target ad placements in AI Overviews. You cannot opt out of serving ads in AI Overviews. Sensitive verticals, among them adult, alcohol, gambling, finance, healthcare and politics, are excluded from the within-overview placement.

Measurement is thinner than the placement suggests. Google says ads in AI Overviews are reported as Top Ads, and that Google Ads does not currently offer segmented reporting when ads show within Search AI Overviews, so the surface is not separable in your ad reports. It is not separable on the organic side either, where Search Console's generative AI performance report has carried impressions and no clicks since its worldwide rollout on 31 August 2026. Both halves arrive aggregated, which is worth stating before anyone builds a model on top of them. The organic half is covered in Google AI Mode: what changes for a publisher.

For AI Mode, Google announced two Gemini-built formats on 20 May 2026, Conversational Discovery ads and Highlighted Answers. The second sounds like a citation and is not: when AI Mode lists recommendations, Google says eligible high-quality ads can appear on that list as a Highlighted Answer, with an independent AI explainer beside the creative, and both formats continue to be labelled Sponsored. Announced is the right word. We found no Google help page documenting eligibility, availability or reporting for either format.

Microsoft and Perplexity

Microsoft has built conversational formats inside Copilot. It began piloting Showroom ads with selected clients in April 2025, an experience a user can be invited into after showing purchase intent, and in April 2026 introduced Offer Highlights, which let brands promote product differentiators inside Copilot conversations for retail use cases in English-speaking markets. Microsoft states that organic responses are not influenced by sponsored content and that sponsored content is "distinctly marked".

Perplexity announced advertising on 30 July 2024 as part of its publishers' programme: sponsored related questions, with a revenue share to publishers whose content a paid interaction referenced, alongside launch partners including TIME, Der Spiegel and Fortune. We could not verify the programme's current status from a Perplexity source, so we are not describing one.

A paid slot is rented in the strictest sense. You control the creative, the landing page and the budget. Placement is decided in an auction you do not see, in a slot you cannot target by name on either Google or OpenAI. The cost is per impression or per click, and the placement stops with the invoice, same day.

An organic citation is not bought. You control the page, the markup and whether crawlers can reach it. Selection is decided by retrieval and by the model's choice of what to quote, and the cost is the work. It persists while the page remains the best available source and the engine keeps retrieving it, and it disappears without notice when either changes. How that selection runs on one engine is set out in how Google AI Overviews chooses its sources, and the broader version of the rented-and-owned split is in rented visibility vs owned visibility in AI search.

Does paying make citation more likely? No platform documentation we read says so, and every platform statement we read points the other way. OpenAI says advertisers have no ability to shape, rank or alter ChatGPT's responses. Google describes ad selection as reading the overview's content, an arrow pointing from the answer to the ad rather than from the ad to the answer. Microsoft says organic responses are not influenced by sponsored content. None of that is independent proof of absence, and none of these companies publishes the experiment that would settle it. Anyone selling a paid slot as a route to organic citation is making a claim their own supplier has not made.

What is known about whether the ads work

Less than the spend implies.

OpenAI has published business figures rather than advertiser outcomes. At the end of August 2026 it said ChatGPT Ads had reached $1 billion in annualised revenue run rate in under 200 days, with tens of thousands of advertisers running campaigns. That is a seller's revenue, not a buyer's return. Its March 2026 pilot update reported no impact on consumer trust metrics and low dismissal rates, with no sample, method or window attached.

Microsoft has published the closest thing to comparative performance data. A post dated 6 August 2025 reports 73 percent higher click-through rates and 16 percent stronger conversion rates for ads in Copilot, with customer journeys 33 percent shorter, sourced to Microsoft first-party data from its owned and operated search properties, global, February to May 2025. Read those as a vendor's self-report about its own inventory: no sample size, no control condition and no breakdown by vertical is published alongside them.

We traced no performance figure for ads inside AI Overviews or AI Mode to a Google page.

The only independent work we found is academic and studies users, not advertisers. "Ads that Talk Back", an arXiv preprint first posted in September 2024 and revised in October 2025, built a chatbot that embedded product advertising in its responses and ran a between-subjects study with 179 participants. Ad injection had only a slight effect on response quality. Participants struggled to detect the ads, and sometimes preferred the responses that contained hidden ones. That is a research prototype rather than a shipping product, and it says nothing about returns, but it is a reason to treat disclosure as a live question.

Honest limits

This post establishes what four companies publish about their own advertising products, and little beyond that. It does not establish that ads in AI answers perform better or worse than ads anywhere else, because the only comparative figures available are one vendor's first-party data about its own inventory. It also does not establish that paid placement has no effect on organic citation. It establishes that no platform claims such an effect, that three of them describe mechanisms separating the two, and that we found no study testing it either way. Those are different statements.

Two practical gaps. OpenAI's live help centre blocks automated fetching, so its two help pages here were read from Internet Archive snapshots dated 23 September and 4 October 2026. And the AI Mode formats are announced rather than documented, so anyone describing how to buy them is working from something other than Google's documentation.

FAQ

Can I pay to be cited in an AI answer? Not as far as any platform documents. What is on sale is a labelled slot beside or inside the answer. OpenAI says advertisers have no ability to shape, rank or alter ChatGPT's responses, Google says the AI Mode ad formats it has announced are labelled Sponsored, and Microsoft says organic responses are not influenced by sponsored content.

Does running ads make organic citation more likely? No platform documentation we read claims it does, and we found no published study testing it. Treat any vendor pitch built on that link as an unsupported claim until someone publishes the measurement.

Can I keep my ads out of Google's AI surfaces? No. Google's advertiser help page states that you cannot opt out of serving ads in AI Overviews, and that you cannot directly target that placement either. Eligibility follows from running Search, Shopping or Performance Max campaigns.

Can I report on AI answer placements separately? Not on Google. Ads in AI Overviews are reported as Top Ads, and Google says it does not currently offer segmented reporting for ads shown within Search AI Overviews. On the organic side, Search Console's generative AI performance report carries impressions and no clicks.

Sources

Every URL below was opened and checked against the sentence citing it on 15 October 2026. Where a page is noted as read from the Internet Archive, the live page refused automated fetching.

Related reading

Google AI Mode: what changes for a publisherHow Google AI Overviews chooses its sourcesRented visibility vs owned visibility in AI searchThe AI crawlers and user agents worth knowingWhy an AI answer changes between two identical runsHow ChatGPT decides what to cite, and how to get cited