ChatGPT ads arrive in the EU: what the launch means

A client called last week with a question that did not exist a month ago. He had asked ChatGPT for a supplier in his own category, got three recommendations, and underneath the answer sat an ad from a fourth company that the text had never mentioned. His question was simple: should we buy that slot, or is it a trap?

It is a good question, and the answer turns on something most coverage has skipped. The ads and the recommendations come from two separate systems.

Here is what actually happened, what can be bought, and what still has to be earned.

What actually happened

Ads inside ChatGPT reached Europe in the week of 24 August 2026. They rolled out across 31 countries: all 27 EU member states plus Iceland, Liechtenstein, Norway and Switzerland. Denmark was in from the start. The United Kingdom got them back in June as a separate wave.

They appear only to users on the free plan and the Go plan. Plus, Pro, Business, Enterprise and education accounts see no ads at all, and users under 18 are excluded entirely, both from account data and from an age prediction.

The format is tight. The ad sits below the answer, clearly labelled as sponsored and visually separated from the text. There are 24 characters for the headline and 48 for the body copy, in every language. That is shorter than a text message subject line, and much shorter than most marketing departments picture when they hear the word advertisement.

Self serve advertiser accounts opened at the same time for businesses registered in the eligible markets. The account is created against the registration country, the registered name is checked against national records, and both country and currency are permanent once set. Targeting works at country level. There is no city or postcode targeting outside the United States yet, so a carpenter in Vejle can buy Denmark, not Vejle.

What the ads do not do

This is the part that matters most, and it is missing from almost all the coverage.

OpenAI states that ads do not influence ChatGPT's answers, and that they run on separate systems. An advertiser cannot buy a mention inside the text itself. Advertisers receive aggregate reporting only, meaning views and clicks, and no chats, chat history, memories, name, email, precise location or IP address are shared.

There are places the ads do not go either. Not in Temporary Chats, not in the Atlas browser, and not next to sensitive or regulated subjects such as physical health, mental health and politics.

Personalisation is opt in with consent, and here is the detail people misread: opting out does not remove the ads. It only changes which ones you see. Without consent the system uses what is in the situation itself, meaning the topic of the current conversation, approximate city level location, device type, time of day and language. With consent it may also use previous ad interactions and inferred interests. Removing ads altogether means moving to a paid plan.

An ad can buy you a place under the answer. It cannot buy you a place in the answer.

Why this changes the arithmetic and not the direction

That separation between the two systems is the whole point for a company deciding what to do.

When somebody asks for a supplier in your category, two things happen on the same screen. The model writes an answer naming the companies it has come to associate with that category. Below it sits an ad chosen from the topic of the conversation. The reader sees both, but does not read them the same way. What sits in the answer reads as advice. What sits below reads as advertising, and it has read that way online for thirty years.

That does not make the slot worthless. It means the slot solves a different problem than the one most buyers think they are solving. An ad can build awareness with somebody considering your category right now. It cannot make the model recommend you in the next conversation.

It is the same distinction we describe in what AI visibility is. Being recommended is not the same as being shown.

And there is an asymmetry worth taking seriously. If you are neither in the answer nor under it, while a competitor is in both, you lose twice on the same screen.

What can actually be bought right now

Four things are worth knowing before anybody writes a budget.

• Targeting is coarse. Country level, with language following the conversation. A Danish ad reaches Denmark, not a neighbourhood.
• The slot is small. 24 characters for the headline. Write the name and the one thing you want remembered. There is no room for a value proposition.
• The audience is the free users. Your B2B buyer inside a larger company often sits on a paid plan and will never see your ad.
• The context is bought, not earned. You pay to stand beside a subject. You are not buying a recommendation, and you cannot get one.

That last point is where price becomes a question of expectations. An ad beside an answer that recommends your competitor rarely converts well. An ad beside an answer that names you converts far better, because the two reinforce each other.

What it says about the direction

There are two roads to take when an answer has to earn money, and the two largest players have taken one each.

OpenAI keeps the ad below the answer, labelled and visually separated, and states plainly that the two systems are separate. Google has gone the other way and places ads inside the AI generated answer in the markets where they have rolled out. Those are two different answers to the same question: how much a payment is allowed to move what the reader experiences as advice.

The boundary between them is the thing worth watching over the next year. As long as the ad sits below the answer, a reader can tell the difference, and an earned mention is worth more than a bought one. If the boundary moves, the arithmetic changes for everybody who invested in being recommended organically.

There is a regulatory layer here too, and it has not been tested in practice yet. The Digital Services Act requires that ads are clearly identifiable and that there is no profiling on sensitive categories or on minors. OpenAI meets that on paper through labelling, age exclusion and a ban on ads near health and politics. Whether a recommendation phrased inside a flowing answer can ever be labelled as clearly as a box beneath it is a question no regulator has answered.

What to do about it

The order of work is the same as it always was, and the arrival of a paid slot has not changed it.

Find out first whether the models mention you in your category at all, and what they say. Without that number you cannot tell whether an ad has to build awareness from nothing or amplify something already working. That is what Signal measures: live queries against ChatGPT and Gemini in your real category, a score across five pillars and a benchmark against three competitors. The output is a 10 to 15 page PDF and a 30 minute walkthrough call within one business day. €690.

After that the decision is simple enough. If you are already named, an ad can amplify that, and it is worth trying at small scale. If you are not named, the money goes further fixing the cause first, because the ad disappears the second you stop paying while a place in the answer stays.

Either way there is one thing worth doing this week: ask it yourself. Open ChatGPT on a free account, put the three questions your customers would ask, and look at who is in the answer and who is underneath it. It takes ten minutes, and most people are surprised.