Ads in ChatGPT are sponsored blocks below the assistant's answer, which OpenAI began showing on 9 February 2026. They are labelled as advertising, kept separate from the answer, and matched to the topic of the conversation rather than to keywords. As of July 2026 they run in the USA, Canada, Australia and New Zealand, with a pilot announced in five more countries. Not one of them is in the European Union.
This is the only article on the subject that opens with bad news for European businesses. The rest open with "the new channel is here, claim your spot". We checked the primary source — OpenAI's official page with three dated updates — and below we go through what it actually says, what follows from it, and what a company in the EU can usefully do while buying the ads is not an option.
What OpenAI actually launched
Here it is by the official announcement, to separate facts from retellings.
| Parameter | As stated |
|---|---|
| Start date | 9 February 2026, test in the USA |
| Who sees them | Signed-in adult users on Free and Go plans |
| Who doesn't | Plus, Pro, Business, Enterprise, Education — no ads |
| Format | A block below the answer, labelled sponsored, visually separated |
| How they're matched | Topic of the current conversation, past chats, past interactions with ads |
| What the advertiser gets | Aggregate impressions and clicks only. No access to chats, history, memory or personal data |
| Where they don't appear | Health, mental health, politics. And in accounts of minors |
| User controls | Hide an ad, view impression history, delete ad data, turn off personalisation |
Worth noting separately is OpenAI's wording on answer independence: advertising does not affect what the assistant replies. The answer is optimised for usefulness; the ad lives in its own block. That is a fundamental difference from search results, where paid and organic are mixed into a single list.
Geography: why Europe is missing and what that means
The rollout timeline, by date, from the official updates:
| Date | What happened |
|---|---|
| 9 February 2026 | Test launched in the USA |
| 26 March 2026 | Pilot expanded: Canada, Australia, New Zealand. OpenAI reports seeing "no impact on trust metrics and a low ad-hiding rate" |
| 7 May 2026 | Plans announced for the UK, Mexico, Brazil, Japan and South Korea |
Nine countries in five months, and not one from the European Union. The UK is on the list, but it left the EU and operates under a different regulatory regime.
OpenAI does not explain the reason publicly, so what follows is our interpretation rather than a fact. But it suggests itself: in the EU an ad model built on profiling from conversation history runs straight into several regulatory layers at once — the GDPR on profiling, the Digital Services Act on ad transparency and the ban on targeting by sensitive categories, plus AI Act requirements.
The practical conclusion does not change: there is nothing to budget for in ChatGPT advertising for the European market today. Anyone promising to launch it for you either has not checked, or means a different country.
You cannot buy a place in the assistant's answer. You can only earn it — and in Europe that is the only route in.
Ads and organic mentions in an assistant are different things
This is easy to confuse, because the market sells both under the single banner of "AI promotion". The difference is fundamental.
| Ad block | Organic mention | |
|---|---|---|
| Where | A separate block below the answer | Inside the answer itself |
| How to get in | Pay | Only through your site's content and external sources |
| Available in the EU | No | Yes, working right now |
| Labelled as advertising | Yes | No |
| User trust | Lower — the label is visible | Higher — read as a recommendation |
Work on organic mentions is called GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization). We covered in detail which of these practices are backed by research and which are not — including a table of evidence levels.
What the research says, as opposed to the vendors
The topic is new and surrounded by promises. We deliberately looked for academic sources rather than agency blogs. Here is what is established reliably and what is not.
A critical review of the field for 2023–2026, published on arXiv, ranks claims by evidence level. High confidence applies to two things: relevance to the query and the position of the source within the context are the main factors in getting into an answer. A factorial experiment across 252,000 trials in six language models showed that moving a source higher in the context matters more than most text rewrites.
Medium confidence applies to the finding that extractable facts and clear structure make text easier for a model to use. And separately: when everyone adopts the same techniques, the advantage to any individual player erodes.
Low confidence applies to the claim that white-hat optimisation reliably improves visibility across different assistants. End-to-end verification has barely been done.
Very low applies to any link between being cited in AI answers and clicks, conversions or revenue. Causation has not been established.
The claim that "optimisation raises AI visibility by 40%" is explicitly rejected as a general statement in that review: it is a maximum for one metric in one configuration. In the C-SEO benchmark, only three of 54 method-and-domain combinations were significantly positive, and none in question-answering tasks. Some techniques actively made positions worse.
If an agency promises you specific percentage gains in assistant visibility, they are selling something the research does not yet support.
What you can actually do now
The confirmed factors point to concrete actions. None of them requires ads to open up in the EU.
Check what is being said about you today
Fifteen minutes of work. Open ChatGPT while signed out, or in a temporary chat — otherwise its memory of you distorts the answer. Ask ten questions your customer would ask: not "best agency", but the way they would describe their problem. Save screenshots. Repeat monthly with the same questions — the trend is your metric. Do the same in Perplexity and in Google with AI answers enabled.
Put a definition in the opening sentences
A Carnegie Mellon study found that the first 150–200 tokens of a page carry disproportionate weight when a source is selected. Pages whose first sentence contains a clear definition score noticeably higher. In practice: a page should open with the answer to "what is this and who is it for", not with a greeting or company history.
Replace adjectives with verifiable facts
Models prefer specifics to general claims: statistics, definitions, comparisons, prices, dates, source references. Explicit prices and recent dates have a measurable effect; formatting changes alone have a weak one. "We are market leaders" gives nothing to quote. "Cut cost per acquisition by 55% in four months" does.
Answer questions with questions
People query an assistant with whole questions, not keywords. A question-and-answer block marked up with the FAQPage standard gives the model ready-made units it can lift and attribute.
Work on other people's platforms
Assistants do not only cite your site. Reviews, directories, industry associations and trade media that already rank for your queries are the sources a model assembles an answer from. Presence there often does more than edits on your own site.
Tidy up what people find when they check you
A separate and frequently ignored layer: your profiles on review platforms, entries in company databases, and what appears when someone searches "company name + reviews". When an assistant answers a question about you, it draws on those too.