ChatGPT Ads: The Complete Guide [2026]

I went through every official OpenAI source on ChatGPT advertising — the policy pages, the help centre, the market expansion announcements, the advertiser API docs. Then through every independent test with real spend I could find, where a named agency published what it paid and what it got back. The reason is simple. In six months the internet filled up with numbers that have no source: "51% of answers carry ads", "Newegg 7x ROAS". Go looking for where those came from and there is nothing there.

This guide separates the verified from the invented. Everything stated here as fact comes either from an official OpenAI source or from a named agency that published its budget and its result. Where something is only a vendor claim with no audit, it says so. Where something does not exist despite being widely described as if it does, it says that too. Everything is current as of 1 September 2026, and I update the guide as the product changes — which it does, fast.

What ChatGPT Ads are and where they appear

The ad appears in exactly one place: below the end of the answer. Not mid-answer, not in a sidebar, not as a banner somewhere in the interface. The user asks a question, gets a full answer, and only underneath it sees one or more clearly labelled ad units. That is the entire placement surface.

That single sentence explains almost everything that will surprise you in the numbers further down. In search, the user sees the ad before getting an answer, while the question is still open. In ChatGPT they see it after the answer, when the question is already closed. Psychologically that is a completely different moment. The low CTR you will see below isn't a defect in the format — it is a consequence of where the ad sits.

The formats that actually exist:

• Standard card with advertiser name, logo, headline, body copy, image, and a link to the site
• Product carousel assembled automatically from a merchant product feed based on conversation signals; in beta since early August 2026
• Dynamic CTA — Shop Now, Book Now, Sign Up, Learn More — chosen automatically by OpenAI rather than by you
• "Ask ChatGPT" on the ad itself, letting the user pull the ad back into the conversation and ask about it; the advertiser has no influence over what ChatGPT then says

Who actually sees the ads:

• Only logged-in users on the Free and Go plans. Plus, Pro, Business, Enterprise and Edu are ad-free.
• No ads for users under 18, none in Temporary Chats, none in the Atlas browser.
• A Free user can opt out of ads in settings, in exchange for a lower message limit.

Where they are available: two different lists

This is the single most common reporting error about ChatGPT Ads, and it is worth understanding before anything else. There are two separate country lists, and almost everyone collapses them into one.

A country can sit on one list and not the other. When you read "ChatGPT Ads are now available in X countries", the first question is which of these two things is being claimed. For a real decision you need both: an entity in a List B country so you can buy, and an audience in List A countries so you have someone to sell to.

How the rollout went:

There has been no single global launch, and OpenAI itself says it is "still early". Serving ads to users and opening self-serve buying arrive separately, often weeks apart, so do not trust an article that gives one date per market for both.

If your market is not on either list, you can start creating an account and leave your details to be notified, but that gives you no earlier access. There is no waitlist in the sense of queue priority. The clean paths are a genuine legal entity that actually operates in a supported market, or running through a client's entity that already does.

How to buy them: account, models, auction

There are five routes to the inventory, and for most businesses only the first one matters.

• Self-serve Ads Manager at ads.openai.com, in beta. Open in the US since 5 May 2026, and for India, Europe, the Middle East and South Africa since 31 August 2026.
• OpenAI Ads Solutions, the managed service for larger advertisers.
• Agency partners: Dentsu, Omnicom, Publicis, WPP.
• Over 50 tech and measurement partners, including Adobe, Criteo, Kargo, Pacvue and StackAdapt.
• Advertiser API for teams building their own automation.

Billing models:

• CPM for reach, available since February 2026.
• CPC for clicks, since 5 May 2026.
• oCPC, where you pay per click but the system optimises toward a conversion. A bid cap is mandatory, the objective and conversion event are locked at creation, and tracking has to be live before the campaign starts. OpenAI says CPC with outcome bidding now carries the majority of campaigns.

The auction is relevance-weighted second-price. In plain terms: the highest bid doesn't win, the best combination of bid and relevance to the current conversation does, and you pay based on the runner-up. Max bid sits at ad group level. If that sounds familiar from Google Ads, it is the same principle — except here you get no Quality Score diagnostics telling you where you are losing.

Onboarding runs in order: business details, identity verification, account review, then campaign, ad group, creative and policy review. Assume review is not instant and do not plan a same-day launch.

Targeting without keywords

This is where most people coming from Google Ads make their first conceptual mistake. There are no keywords in ChatGPT. There are context hints: a free-text description of the topics and situations you want to appear in, set at ad group level. Not exact match, not phrase match, and no guarantee of delivery. It is a signal to the system, not a rule.

What you can control:

• Context hints at ad group level
• Geo: country everywhere, plus state, DMA or ZIP in the US
• Platform: iOS, Android, web
• Custom audiences: uploaded emails or phone numbers as SHA-256 hashes, up to 512 MB and 5 million identifiers, with a minimum of 25,000 matched users before a list is usable; available as inclusion, exclusion or bid multiplier

What does not exist:

• Keyword lists and match types
• Demographics and interests
• Site-visitor retargeting
• Query-level reporting. You cannot see which prompt triggered which ad. For me this is the biggest practical gap: there is no equivalent of a search terms report, which means there is no weekly cleanup loop of the kind I run on every Search account.

On privacy, the advertiser never receives conversations, history, memories, name, email, precise location or IP address. You get aggregated performance data. Ads-related data is retained for 30 days after deletion. ChatGPT was designated a Very Large Online Search Engine in the EU on 31 August 2026, so by the end of December 2026 it must meet DSA obligations, including a public ad repository listing ad content, advertiser, serving period and targeting parameters — effectively an ad library for ChatGPT, with a known deadline.

Measurement and the click gap

Ads Manager gives you impressions, clicks, spend, CTR, average CPC and CPM, conversions, CSV export, an API and a change log. For conversions there are three mechanisms:

• The OAIQ measurement pixel: a JavaScript SDK with a first-party cookie lasting around 30 days, standard and custom events.
• Conversions API: server-side events with event-ID deduplication, so the same action isn't counted twice.
• Automatic Advanced Matching: captures form data on your site and hashes it in the browser before sending. On by default for new pixels, and switched on automatically for existing ones on 17 August 2026.

UTM parameters pass through on click, and there are dynamic parameters for campaign, ad group and ad ID. Attribution covers clicks plus, where enabled, a one-day view-through window.

Now the part guides rarely include. The most common practitioner complaint is a click gap: the platform reports materially more clicks than analytics records sessions. The publicly documented cases, per MediaPost's 28 August 2026 report: Cleverly reported 57 clicks against fewer than 20 sessions; Symphonic 53 against 35, a gap of roughly 34 percent; Out of the Box verified 13 percent of roughly 140 clicks in July, rising to 68 percent in August; one case ran 120 against 8.

The cause has not been established. I am not going to write fraud or double counting, because there is no evidence for either, and click-versus-session counting differences exist on every platform. What is established is that the gap narrowed over the summer, which points toward measurement maturity rather than the inventory itself.

All of this assumes site tracking already works before you ever open a ChatGPT account. The fundamentals and the order I set them up in are in my conversion tracking guide.

Who is allowed to advertise

The policy is narrow, and this is probably where most businesses drop out before the budget conversation even starts.

I see this gap missed constantly. Articles list the supported countries, and the whole thing gets read as "so we can run it". Category is checked before market, not after. Fully prohibited categories include adult content and dating, alcohol, tobacco and vaping, gambling and betting, drugs and cannabis, politics, individual real estate and job listings, scams, sensitive events, and wellness claims without evidence.

There is also a placement-level brand safety layer, separate from advertiser eligibility: ads do not serve in conversations about mental health, in emotionally sensitive contexts, or across 19 further defined content categories. Review itself runs at three levels — advertiser, creative plus landing page, and placement — mostly automated with human escalation.

What the real numbers say

Here I separate three things that most coverage blends together: independent tests with published spend, OpenAI's own claims, and ad density studies.

Independent tests with real money (none audited, but all named with published budgets):

Three findings repeat across every source. First, the recommended CPC understates real cost in competitive verticals by two to four times. Second, CTR converges at 0.65-1.3 percent. For orientation: Google Search campaigns average roughly 2-3 percent CTR, and well-built campaigns on branded and highly relevant terms run 5-10 percent. The difference is not a defect — it follows from the ChatGPT ad sitting below an answer that has already resolved the question. Third, most zero-conversion reports come from campaigns run before August, meaning before oCPC, the pixel and the Conversions API matured, so early results probably understate what the platform can do today.

How many answers actually carry ads is the question most often misreported, because the denominator changes everything:

• SE Ranking, August, 50,000+ commercial prompts: ads on 25.94% of commercial prompts; 14.35% of ads semantically unrelated to the topic.
• Adthena, March to May, 850,000 queries of all kinds across US and UK: ads on 4.47% of all queries, averaging 1.06 sponsored items per answer; retail and fashion account for 39% of placements.
• Similarweb: 26% of answers carry ads, CTR 0.50%.

Both headline studies can be true at once, because they measure different things over different windows. So treat every ad density figure as inseparable from its denominator and its date. Without those, it is just a number.

Hype vs reality

On 31 August 2026 OpenAI announced a $1 billion annualised run rate in under 200 days. The figure is official, but a run rate is an extrapolation of current pace, not realised annual revenue. Reputable reporting puts the internal 2026 target somewhere in the $2.3-2.5 billion range.

Against that, eMarketer forecasts that all AI chatbot advertising will stay under $1 billion in 2026 and reach around $5.4 billion by 2030. If that forecast holds, OpenAI misses its own projection by roughly 90 percent. Digiday reported in May and June that advertisers on large commitments had spent less than 10 percent of their budgets due to inventory constraints, and that around half of inventory went unfilled in June.

The most instructive precedent isn't a number, it is an event: Perplexity paused onboarding new advertisers in late 2025, citing weak results and difficult measurement. That does not mean the same happens to OpenAI, which has disproportionately more reach. It means the "ad below the AI answer" model has already failed to deliver once, and that is worth holding in mind when someone plans next year's budget around it.

Should you test it

My filter is short, and I apply it before we ever open the budget conversation.

If you do go in, go in with a question rather than a target ROAS. The question is whether people find you through this channel at all, and how their on-site behaviour compares with traffic from your other sources. That is answerable in four to six weeks. A target ROAS on a channel with no query-level reporting and a documented measurement gap is not something I would promise. The same logic applies to splitting budget between channels you already run, which I covered in my comparison of Google Ads and Meta Ads.

How I run a pilot

The order is always the same, and the first two steps happen before a single ad is written.

The full process, and what working together looks like, is on my ChatGPT Ads service page.

Frequently asked questions