What ChatGPT ads can't buy you.

You can now buy space inside ChatGPT. The part worth understanding before you do is where that space sits: OpenAI puts ads below the answer and says they stay separate from it. So the interesting question is not how to run a campaign. It is what the answer above your ad looks like, and whether the ad account changes it. We have measured the first and OpenAI has documented the second.

What happened

OpenAI opened an advertising product inside ChatGPT, run through a self-serve tool it calls Ads Manager Beta at ads.openai.com. It names Newegg, Best Buy, Lowe's and VistaPrint among early advertisers. The documentation is still moving: the help articles we read carried update ages ranging from "last month" to fourteen hours, so treat any date you see, including ours, as a reading date rather than a launch date.

The facts, as documented:

• Ads "appear below ChatGPT responses" and carry an advertiser name, favicon, title, copy, landing page and image.
• Buying is CPM or CPC, with an oCPC objective documented in the campaign guide. OpenAI recommends starting at a maximum bid of $3 to $5 per click.
• Selection runs through "a relevance-weighted, second-price auction".
• Reporting covers impressions, clicks, spend, CTR, average CPC, average CPM and conversions, and UTM parameters "persist on ad clicks".

"Ads remain distinct from ChatGPT's responses." OpenAI, ads.openai.com, read 2026-08-27

Where the ad actually sits

A ChatGPT ad is a labelled unit placed underneath an assistant answer, carrying an advertiser name, a favicon, a headline, a line of copy, an image and a landing page. It is not a placement inside the answer, and OpenAI is unusually direct about the difference.

Its advertiser documentation says ads "appear below ChatGPT responses". The product page lists three commitments alongside that: ads are "clearly identified", they "remain distinct from ChatGPT's responses", and people "control how their data is used for ads" ( Ads in ChatGPT: The Basics and ads.openai.com, both read 2026-08-27).

Read that as a buyer rather than as an advertiser. When somebody asks ChatGPT which agency, clinic or tool to use, the recommendation they act on is generated above the ad, from pages the assistant retrieved and read. The ad is a labelled unit underneath. Both are real placements and both can earn a click. They are not substitutes, and no budget converts one into the other.

What the answer above the ad looks like

We put 50 buying questions to ChatGPT, Claude, Gemini and Perplexity in July 2026, ten questions in each of five industries, and counted what came back. Nothing here is modelled. Every figure is counted from recorded answers, and the prompts, dates and a run-to-run variance control are published with the dataset.

50 buyer questions across five industries, July 2026. Counted from recorded answers.

Two things in that table matter for anyone weighing a campaign. ChatGPT names roughly nine businesses per answer, so the answer is not a single winner-takes-all slot; it is a shortlist with room on it. And ChatGPT leans on directories more than any other assistant we tested, at 26% of its citations, so a listing profile is often doing work your own site is not.

The reach of that reading is wider than it looks. Ten questions per industry produced citations to between 143 and 220 different websites, and across all 50 questions all four assistants converged on the same source only three times. Whatever is happening above the ad, it is not a small fixed list you can enumerate.

The buyers you most want are excluded

This is the single most important line in OpenAI's documentation and the one least likely to make it into a launch roundup:

"We do not show ads to users on Plus, Pro, or any Business plan, or to any accounts identified as belonging to users under 18."

If you sell to businesses, work out what that removes. The buyer researching vendors on a company ChatGPT seat is not in the addressable audience. Neither is anyone who pays for their own Plus subscription, which skews toward exactly the heavier, more committed users a considered purchase depends on. Ads in ChatGPT reach free-tier adults, and that is a real audience, but it is not the whole room and for some businesses it is not the part of the room that buys.

Those excluded users still get answers. They are simply getting the answer without an ad under it, assembled the same way it always was, from pages the assistant could reach and read.

Your landing page is already in the auction

Paid and organic are usually run as separate workstreams. OpenAI's description of ad selection makes that harder to justify here.

The system, it says, "selects and delivers ads based on expected relevance and outcomes". Among the signals it weighs are "the context and intent of the current conversation, the ad's landing page, title, copy, advertiser-provided context hints". It resolves those through what OpenAI calls "a relevance-weighted, second-price auction" ( Ads in ChatGPT: The Basics, read 2026-08-27).

A second-price auction is one where the winner pays just above the next-highest bid rather than their own, so your maximum bid is a ceiling and not a price. Relevance-weighted means that bid is not the only input: how well the ad and its landing page fit the conversation affects both whether you win and what you pay.

The landing page is a selection signal, and the auction weights relevance. So the same work that makes a page legible to an assistant, clear structure, answers to the questions buyers actually ask, and content that matches the conversation it appears in, is feeding both placements. Neglecting the page does not just cost you the citation. In a relevance-weighted auction it plausibly costs you money on the ad too. We have not measured that second effect and will not claim a number for it.

Context hints are not keywords

Anyone arriving from Google Ads will reach for the keyword field and not find one.

Context hintsare descriptions you attach to an ad group of "the conversations, topics, or keywords where their products or services may be relevant", in OpenAI's words. Then comes the line that matters: "these hints help guide ad matching, but they are not exact-match keywords and do not guarantee delivery in specific conversations."

That is a genuinely different model, and it is worth sitting with. You are not buying a query. You are describing a context and letting a relevance system decide whether you belong in it. The lever you have is how well your offer, copy and page match a conversation, which is uncomfortably close to the lever you have for being named in the answer above.

Five things in the docs that will cost you

None of these are hidden. All of them are easy to skip on the way to a launch button, and each is quoted from OpenAI's own pages, read 2026-08-27.

• A campaign-total budget does not pace.OpenAI: it is "a total spending limit, not a pacing control. The budget isn't distributed evenly across campaign dates, and spend may accumulate quickly." A daily budget is a target, not a cap: daily spend "will not exceed twice your selected daily budget".
• Agencies need one account per client."Each advertiser requires its own account. If you are managing multiple advertisers, you will need to create an account for each advertiser." Plan the admin before you promise a client a start date.
• Verification is a queue, not a form.Account checks run through Persona, and OpenAI warns that "because there is a high volume of applications that are reviewed in a rolling queue, account verification may take some time."
• Ads will not serve on an incomplete account.Account name and logo have to be set: "Ads will not serve unless this step is complete." Billing must be fully configured before anything delivers.
• The product is moving faster than its own documentation.On the day we read them, "Ads in ChatGPT: The Basics" described CPM and CPC buying, while the campaign guide, updated fourteen hours earlier, also documented an oCPC objective. Check the article dates before you rely on any guide, including this one.

What we would actually do

1. This week: check you can be read at all. A robots.txt line blocking a crawler is the cheapest problem in this article to find, and it caps both placements at once. OpenAI runs three separate agents, and blocking one does not block the others.
2. This week: ask the assistants your buyers' questions. Not your brand name. The questions somebody asks before they know you exist. Repeat each one, because answers vary between runs, and a single check tells you close to nothing.
3. This month: decide whether your buyers are even reachable. If you sell to businesses, work out what share of your buyers sit on Plus, Pro or a company plan. That share is not addressable by ads at any budget.
4. This month, if you do test: tag everything and cap the spend. UTM parameters "persist on ad clicks", so the traffic lands in the analytics you already have. Prefer a daily budget over a campaign total until you have seen how fast it moves.
5. This quarter: fix the landing pages either way. They feed the ad auction and they feed the answer. It is the one piece of work that is not wasted whichever placement you end up caring about.

Do not:

• Treat the ad as a substitute for being named. Only one of the two survives you turning the budget off.
• Rebuild your site around ad copy. The answer above the ad is assembled from pages the assistant can reach and read, which the ad account does not touch.
• Assume Google Ads habits transfer. There is no keyword to buy and no guaranteed delivery in a given conversation.

Common questions

Sources

Every OpenAI fact above was read from these pages on 2026-08-27. They are a beta product's documentation and they change; where the wording carries the meaning we have quoted it rather than paraphrased.

• Advertise in ChatGPT, OpenAI
• Ads in ChatGPT: The Basics, OpenAI Help Center
• Ads Manager Beta Overview, OpenAI Help Center
• Quickstart: Launch your first campaign, OpenAI Help Center
• Ads Manager Beta Account Setup, OpenAI Help Center

The citation figures are our own. The study sets out the questions, the assistants, the dates and the variance control, and the raw dataset is published under CC BY so you can recount them. If something here does not match what you see, tell us and we will correct it.