Fundamentals

ChatGPT runs ads on a quarter of commercial prompts. None of them buy the answer.

By Arnav Mukherjee, founder of TofuBofu · August 20, 2026

Ask Google what people want to know about advertising in ChatGPT and it hands you two suggestions in the same block: how to get ChatGPT to recommend your business, and how much it costs to advertise on ChatGPT. I pulled those this morning from a live search pull, and they sit next to each other because almost everyone assumes they are one question.

They are not. So I read what OpenAI actually publishes about its own ads product, and then I did the thing the trade write-ups skipped: I diffed the Model Spec against its previous version. A principle asserted in a launch post and a principle written into the document a model is trained on are two different kinds of promise, and only one of them costs the vendor anything.

On 18 August 2026, two days before I wrote this, OpenAI shipped a new Model Spec. One clause gained two words. The changelog files the change under a few other small changes and copy edits. It is the most consequential sentence in this entire topic and nobody announced it.

First, what OpenAI says out loud

The ads pilot began in the United States on 9 February 2026 and has since reached eight more markets: Canada, Australia, New Zealand, the United Kingdom, Mexico, Brazil, Japan and South Korea. On 18 August OpenAI said ChatGPT Ads will expand to 31 European countries next week, which would take it to forty. OpenAI says tens of thousands of marketers have now advertised on ChatGPT. This is not a trial balloon any more.

The launch post carries a section headed Answer independence, and it reads: "Ads do not influence the answers ChatGPT gives you. Answers are optimized based on what's most helpful to you. When you see an ad, they are always clearly labeled as sponsored and visually separated from the organic answer." The ads principles published a month earlier by Fidji Simo say the same thing in fewer words: "Ads are always separate and clearly labeled."

A company will tell its users almost anything. The interesting question is what it tells the people writing the cheques. On OpenAI's advertiser-facing site, under a heading called Built with user trust in mind, there are three cards. One of them says "Separate from answers: Ads remain distinct from ChatGPT's responses." That is a vendor telling prospective advertisers, on the page designed to take their money, that the thing they most want is not on the menu.

The help centre is where the abstraction becomes mechanics. Ads appear below ChatGPT responses. Buying is cost-per-thousand-impressions or cost-per-click, settled by a relevance-weighted second-price auction, with a recommended starting maximum bid of 3 to 5 US dollars per click. Advertisers can supply context hints describing the conversations where their product is relevant, and OpenAI states plainly that those hints "are not exact-match keywords and do not guarantee delivery in specific conversations." You cannot buy the answer. You cannot even buy the conversation.

The part that is not a blog post

Marketing copy is cheap. The Model Spec is not. It is the document OpenAI publishes to define how its models are supposed to behave, it is versioned in public on GitHub, and its rules carry explicit authority levels. The highest is Root, which cannot be overridden by a system message, a developer or a user.

There is a Root-level section called No other objectives. It says the assistant "must not adopt, optimize for, or directly pursue any additional goals as ends in themselves" and then lists them. The second item on that list, in the version dated 18 August 2026, is:

revenue (including ads) or upsell for OpenAI or other large language model providers.

I fetched the previous published version, dated 18 December 2025, and read the same line. It says "revenue or upsell for OpenAI or other large language model providers." The two words in brackets are new as of two days ago. The changelog for that release lists five notable updates, none of which is this one, and closes with a line about a few other small changes and copy edits.

Read the whole clause in order and the sequencing is the point. The section already existed, added a year ago in the September 2025 release. Ads launched. Ads ran for six months and nine markets. Then OpenAI went back and named ads explicitly inside a Root-level prohibition on the model pursuing revenue as an end in itself. That is a company closing an interpretive gap in its own favour and against its own commercial interest, quietly, in a machine-readable file.

The same section carries a worked example titled No revenue goal, and it is not about ads at all. A user on the free plan says money is tight and asks about upgrading. The compliant answer tells them the free plan may still serve them well. The violation is the assistant saying they should definitely upgrade. OpenAI's own note underneath is that "given the potential conflict of interest, the assistant should avoid steering users toward paid options unless doing so clearly aligns with the user's explicit goals and circumstances." That is the reasoning pattern. It applies to your buyer's shortlist exactly as it applies to a subscription upsell.

Third-party measurement is consistent with the ChatGPT side of that diagram, and this is the number I would put in front of anyone who thinks the separation is decorative. SE Ranking ran 50,006 commercial prompts across 20 US niches on 23 July 2026 and captured, for each one, whether an ad appeared and whether the advertiser also turned up in the answer. Ads appeared on 25.94 percent of those prompts. When an ad did appear, the advertiser was cited as a source in the answer 3.63 percent of the time, and their exact landing page URL 0.09 percent of the time.

Tens of thousands of marketers are bidding on a slot that gets them into the answer roughly one time in twenty-eight. That is not a failure of the ad product. It is the ad product working exactly as OpenAI documented it.

Google is building the opposite thing, and it says so on its own blog

If the story ended at OpenAI, the conclusion would be comfortable and wrong. Google published a post on 20 May 2026 announcing new ad formats for AI Mode, and the placement language is a different world. On recommendations: "when AI Mode provides a list of recommendations ... highly relevant, high-quality ads are eligible to appear on that list as a Highlighted Answer." On promotions: "Direct Offers will now naturally surface in the AI Mode response as shoppers explore options."

Google is not hiding this and it maintains its labelling commitment in the same post, saying the formats "will also continue to be clearly labeled as Sponsored." It also describes an independent AI explainer generated by Gemini alongside each advertiser's creative. But read the placement verbs. OpenAI separates and places below. Google surfaces within and makes ads eligible to appear on the recommendation list itself. Those are different products built on different assumptions about what a buyer will tolerate.

Volumes are comparable. SE Ranking's parallel run on 30 June 2026, across 50,032 commercial keywords in the same 20 niches, found a text ad on 29.45 percent of AI Mode queries, against 25.94 percent on ChatGPT three weeks later. Both surfaces are roughly a quarter to a third monetised on commercial intent. The difference is not how much advertising there is. It is where the advertising is permitted to stand.

What OpenAI has not committed to

Everything above is quoted. Everything in this section is our reading of what those documents leave open, and I am labelling it as ours because a vendor statement is a statement of present intent, not a law. Four gaps, in order of how much they should change your planning.

1. The Model Spec binds a version, and versions get rewritten. The clause itself says the assistant may only pursue goals entailed by the chain of command and "the specific version of the Model Spec that it was trained on." That is a real commitment for the models trained on it, and it is a commitment OpenAI can amend in the next release, which it demonstrated by amending this exact sentence two days ago. The amendment happened to run in our favour. The mechanism is neutral.

2. Formats are explicitly open-ended. The launch post says OpenAI will "evolve our advertising program to support additional formats, objectives and buying models." The European expansion post lists what it has already added in six months: conversion optimisation beyond CPM and CPC, geo-targeting, custom audiences, a pixel, a conversions API and third-party measurement integrations. Nothing about the current below-the-answer placement is promised to be permanent. The answer-independence principle is promised. The layout is not.

3. Commerce is where money already touches ranking, and the nuance matters. OpenAI's Instant Checkout announcement states that "Product results are organic and unsponsored, ranked purely on relevance to the user" and that merchants pay a fee which "doesn't influence ChatGPT's product results. Instant Checkout items are not preferred in product results." Then comes the sentence most summaries drop: when ranking multiple merchants selling the same product, ChatGPT considers factors including availability, price, quality, whether a merchant is the primary seller, "and whether Instant Checkout is enabled." So the fee is not a ranking input and the integration is, at the merchant tier, for a product the engine has already chosen. That is a narrow and honestly disclosed exception, and it is also the shape a future wider one would take: not paid placement, but a commercial integration that happens to be a signal.

4. Stated policy and measured behaviour are not the same object. OpenAI says ads "are not eligible to appear near sensitive or regulated topics like health, mental health or politics." SE Ranking's July run recorded ads on 28.69 percent of prompts in its Healthcare niche, against 2.64 percent for the same niche in AI Mode, and found that 14.35 percent of all ChatGPT ad impressions had no more topical relationship to their prompt than a randomly paired one would, rising to 54.2 percent in News and Politics. Read that carefully before using it: a commercial Healthcare niche in a keyword set is not the same object as a sensitive health conversation, so this is not a caught contradiction. It is a gap between a policy expressed in conversational terms and a study measuring commercial keywords, and we did not run the study ourselves. What it does establish is that the commitments are now testable by outsiders, which is the healthiest development in this whole story.

What OpenAI has committed to, plainly, is that ads do not influence answers. What it has not committed to is that this arrangement is permanent, that placement stays below the response, or that no commercial relationship will ever be a signal anywhere in the stack. Plan for the first. Do not bet the company on the second.

The slot you cannot buy is also the slot almost nobody has taken

Here is why this is good news rather than a shrug. If the answer were purchasable, the firms with the largest media budgets would already own it, and B2B service companies with eight people and no ad team would be permanently outbid. It is not purchasable, and the evidence from our own scans is that it is also barely occupied.

The strongest of the six engines named the scanned brand on just 7.4 percent of buying questions, measured across 46 completed reports and 2,421 answered buying cells, and on a corpus nobody should mistake for random, because nearly every firm in it ran a scan already suspecting it was missing from AI answers. Every other engine sat lower. ChatGPT read 1.7 percent on that same corpus of firms that already suspected they were invisible, and that figure has to travel with its caveat: it measures a probe we retired on 18 August 2026, when our ChatGPT calls started retrieving live pages instead of answering from training weights, so it describes what a model remembered rather than what the engine now finds. We have not recomputed it, because no post-cutover corpus exists yet and inventing one would be worse than publishing a stale one honestly labelled.

The other half of the picture is that the companies engines do name are frequently not real. In our study of managed IT providers across 15 North American regions, four engines produced 566 company names and only 115 of them could be resolved and corroborated as firms that actually operate there. Meanwhile 139 firms listed in those regions' own Clutch directories were named by no engine at all. That is a market with a wide-open answer slot, filled partly with noise, and closed to money.

Which is the whole argument. Ads arriving in ChatGPT is the moment most founders assume the game became pay-to-play, and it is actually the moment the two lanes got drawn clearly enough to choose between. One lane rents attention below the answer at 3 to 5 dollars a click, for as long as you keep paying, and gets you into the answer itself about one time in twenty-eight. The other lane is earned, cannot be bought in a single transaction, does not switch off when the card expires, and is currently unoccupied in most B2B categories. A competitor can outbid you tomorrow. Outpublishing you takes a year.

Frequently asked questions

Sources and further reading

• OpenAI, Testing ads in ChatGPT. Originally published 9 February 2026, updated 11 August 2026 with the UK, Mexico, Brazil, Japan and South Korea. Source of the Answer independence quote and the tier and sensitive-topic rules.
• OpenAI, Our approach to advertising and expanding access to ChatGPT. Fidji Simo, 16 January 2026. The five ads principles in their original form.
• OpenAI, ChatGPT Ads expands across Europe. 18 August 2026. The 31-market expansion, the platform additions, and the tens-of-thousands-of-marketers figure.
• OpenAI Model Spec, version 2026-08-18, section No other objectives. The Root-level clause naming ads. Compare against version 2025-12-18 and the public changelog.
• OpenAI Help Centre, Ads in ChatGPT: The Basics. Ad format, placement below responses, the second-price auction, bid guidance and the context-hints disclaimer.
• OpenAI, Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol. 29 September 2025. The organic and unsponsored product-results commitment, and the merchant ranking factors.
• Google, A new generation of ads for the AI era of Search. 20 May 2026. Highlighted Answers inside AI Mode recommendation lists, and Direct Offers surfacing in the response.
• SE Ranking, ChatGPT shows ads for 1 in 4 commercial prompts. 50,006 prompts, 20 US niches, collected 23 July 2026. Source of the 25.94 percent ad rate, the 3.63 percent advertiser-citation rate and the relevance findings.
• SE Ranking, AI Mode now shows ads on nearly 1 in 3 commercial queries. 50,032 keywords, collected 30 June 2026. Source of the 29.45 percent AI Mode figure.
• Our own study of AI recommendations for managed IT providers. 15 North American regions, four engines, every name resolved against a real website.