Ads Reached ChatGPT in Europe Before the Advertisers Did

A shopper in Lyon opens ChatGPT, asks which running shoe suits a wide foot, and reads a paragraph of advice. Under it, from late August, sits a small labelled card with somebody's product in it. That card is new in Europe. Whether you can buy one is a separate question, and the answer for most sellers reading this is not yet.

The story broke in two halves that most coverage merged. One half is display: ads reaching European eyeballs. The other half is buying: a self-serve tool that decides whether a one person shop can participate at all. They arrived months apart, in different countries, and the gap between them is the part worth understanding before you plan any budget around it.

What actually changed in August?

OpenAI notified European users on 15 August 2026 that ads would begin appearing on the Free and Go tiers across the European Economic Area and Switzerland later that month, according to the notification email reported by PPC Land, sent by OpenAI Ireland Limited as the data controller for European users. Digiday puts the start date at 24 August across 31 European markets and reports that a new EU privacy policy went up on 14 August to cover the data handling underneath it.

Paid tiers are untouched. Plus, Pro, Enterprise, Business and Education stay free of ads. Free users who dislike the change can move to a reduced configuration with lower message limits instead. Ads do not appear near personal health, mental health or political topics, accounts identified as belonging to under 18s are excluded, and Temporary Chats and the Atlas browser carry none.

The two outlets publish different market lists, which looks like a contradiction and is not. Digiday counts nine markets live before Europe, including Canada, Australia, the United Kingdom, Japan, South Korea, Mexico and Brazil. The buying documentation still describes ads as visible in four. Both are accurate for the day they were written, which tells you something useful on its own: this is a rollout measured in weeks, so any list you read is a snapshot with an expiry date.

Six months separate the first American impression from the first European one. If that cadence holds for the buying side, European sellers are looking at a wait rather than a launch.

What does an ad in ChatGPT actually look like?

Smaller than anything you have written before. The unit is an advertiser name, a favicon, a title of 16 characters, 32 characters of body copy, an image asset and a landing page link, per the format documented by WebFX. Sixteen characters is roughly "Wide fit runners". Thirty two is roughly "Sizes 38 to 50, returns in 30 days", and that is already over.

The card sits below the response text, labelled as sponsored and visually separated from it. It is not woven into the paragraph the model wrote. Anyone who has watched sponsored results creep into search results will recognise the shape, with one difference that matters: there is no headline auction for the answer itself, because the answer is not for sale.

Targeting is conversational rather than demographic. An advertiser supplies context hints describing the conversations, topics or keywords they want to sit near, and can apply bid multipliers at ad group level plus geographic exclusions. At launch in Europe the selection uses the current conversation topic, general location and device type. Past chats and stored memories are excluded until a user opts in to personalisation.

What comes back is thin by design. Advertisers receive aggregated performance metrics, total views and clicks, with no access to individual chats or personal details. The reporting surface covers impressions, clicks, spend, click through rate, average cost per click, average CPM, conversions, attributed sales value and return on ad spend, with a Conversions API, a pixel and UTM tracking to feed it.

Can a small seller buy one yet?

In the United States, yes, in beta. Elsewhere, not on any published date. OpenAI launched a beta self-serve Ads Manager with cost per click bidding and conversion measurement, and Search Engine Journal reports that United States advertisers can register for access, upload ads, manage budgets, control pacing and monitor performance in the platform. Before that, buying ran through agency groups such as Dentsu, Omnicom Group, Publicis Groupe and WPP, or through technology partners including Adobe, Criteo, Kargo, Pacvue and StackAdapt.

The eligibility line is the one to read twice. The self-serve platform describes United States based businesses of all sizes. The European announcement, by contrast, does not confirm when European advertisers gain direct buying access, noting that availability may continue to evolve as testing expands. So a French shoe shop can have its customers shown ads this month while having no documented way to buy one.

The gating that did disappear is the money one. The pilot carried a 50,000 dollar minimum spend. That minimum has been removed, and any approved advertiser can now set a budget that fits a test, with approval and ongoing review working much the way automated ad review does on the established platforms. For the first time this is a channel a business without a media agency could plausibly try, which is why the removal matters more to a small seller than the European display date does.

What does a first test cost?

Start with the published bids rather than a guess. The recommended starting maximum bid for cost per click campaigns is 3 to 5 dollars per click. The default maximum for CPM campaigns is 60 dollars, the same figure the United States pilot ran at in February.

Put that next to a small budget and the arithmetic gets uncomfortable quickly. A 500 dollar test at a 4 dollar click buys about 125 clicks. If your store converts visitors at 2 percent, that is between two and three orders, far below the traffic a creative test needs before a winner is real, which tells you almost nothing statistically and will not pay for itself unless your average order is large or your customers come back. This channel currently rewards a high value product, a strong repeat rate, or patience.

One targeting tool is simply out of range for most readers here. Custom audiences built from email or phone lists require a minimum of 25,000 matched users. A shop with 4,000 subscribers cannot use it, and there is no partial version. The context hint targeting is the accessible half of the toolkit.

Does paying change the answer above the card?

No, and that is the load bearing claim of the whole product. OpenAI frames the model as grounded in answer independence, privacy and user control, in the words of David Dugan, its head of global solutions, quoted by Search Engine Journal. The sponsored card is a separate object from the paragraph above it.

Which leaves a merchant with two distinct jobs rather than one. The card is bought. The mention inside the answer is earned, by the same unglamorous work that decides whether an assistant can read your catalogue at all. We looked at what actually moves that in our breakdown of what changes your citation rate in AI answers, and the honest summary is that structured, specific, machine readable product data does more than any amount of copy polish.

The same split showed up when supermarket assistants started carrying sponsored placements. We wrote about sponsored listings inside a grocery AI assistant when Kroger shipped them, and the awkward part was identical: sellers who could not buy the placement were competing against sellers who could, inside an interface the shopper trusted as advice. Europe has now imported that awkwardness at the scale of a general assistant.

What does the consent model mean for a European seller?

It means the audience you can eventually reach with personalised ads is an opted in subset, not everyone on the free tier. Personalised advertising under GDPR needs a lawful basis, and rather than lean on legitimate interest, a route regulators have rejected for other platforms, OpenAI is asking for explicit consent. Users choose between personalised ads, which draw on conversation data, purchase history and ad engagement, and contextual ads that use only the current conversation and rough location. Consent can be withdrawn in settings at any time.

There is a fairness question underneath that nobody has answered yet. Clara Westbrook, head of data privacy at Arbor Law, told Digiday that because ads apply only to the free plans, users may feel pressured onto a paid plan to avoid having their data used for advertising. Whether a regulator eventually reads that as freely given consent is an open matter, and it is the kind of open matter that can change a channel's rules a year after you build a habit around it.

For your own store the practical read is narrower. If you run retargeting off a pixel, ChatGPT traffic arriving from a sponsored card is subject to your own consent banner, not OpenAI's. The AI platform's consent covers the ad it served, not the tracking you run afterwards.

How big is the ad eligible audience?

Almost all of it. Digiday cites reporting from The Information putting paying users at roughly 35 million against a global base of about 900 million. Run that division and paying subscribers are under 4 percent of the population, which means the ad free tiers are a rounding error and the ad carrying tiers are effectively the product.

That inverts the usual worry about advertising inside a subscription service. The concern with most freemium products is that the free tier is a small, low intent slice. Here the free tier is where nearly everyone lives, including plenty of people who would happily spend 200 euros on a pair of shoes. The audience quality argument against this channel is weak; the measurement argument against it is strong.

Set that against the excluded contexts and you get a clearer picture of where your card can appear. Health, mental health and political conversations are off limits. Under 18 accounts are excluded through account age data combined with behavioural signals. Temporary Chats carry nothing. If your product sits close to any of those categories, a meaningful share of the conversations you would want to reach are structurally unavailable, and no bid changes that.

Which businesses should try this first?

The ones whose economics survive a 4 dollar click. Work through it honestly with your own numbers rather than the ones a case study offers you.

A jeweller with a 600 euro average order and a 55 percent margin can absorb an expensive click and still learn something from 125 of them. A shop selling 18 euro candles cannot, and no amount of creative work fixes an arithmetic problem of that size. Between those poles sit the businesses where the decision turns on repeat purchase: if a first order reliably becomes four, the cost that matters is per acquired customer, not per first sale, and this channel becomes arguable.

Service businesses have a quieter advantage here. A booked consultation or a quoted job carries a value per lead that most product sellers cannot match, and the conversations people have with an assistant before hiring somebody are long, specific and full of the context hints the targeting runs on. If you sell time rather than objects, you are closer to the front of this queue than a small retailer is.

What is worth doing this month?

Three things, none of which require the ad platform to open in your country.

First, find out whether assistant traffic already reaches you. Most analytics setups lump it into direct or referral. Give it its own segment now, before the ads land, so that when the volume changes you can tell paid from organic rather than guessing. A month of clean baseline data is worth more than an early campaign.

Second, make the catalogue legible. Every improvement that earns an unpaid mention also improves what a paid card can point at, and unlike the ad account it is entirely under your control. Product pages that answer size, material, delivery time and return window in plain text do measurably better with machine readers than pages that hide the same facts in an image. We covered the specific failures in how AI shoppers read product pages, and most of them take an afternoon to fix.

Third, write down what you would test if buying opened tomorrow. One product, one audience, a budget you can lose, a number that decides pass or fail. Teams that treat a new channel as an experiment with a stop rule spend a fraction of what teams who treat it as a land grab spend, and they learn the same thing. If you are still assembling the storefront those tests would point at, our AI ecommerce store builder is where that side of the work starts.

The part nobody can price yet

Nobody knows what a click from inside an answer is worth, because the behaviour on the other side of it is new. A person who clicked a search ad typed a query and chose a result. A person who clicks a card under an AI answer has already read a recommendation, possibly one that named a competitor. They arrive warmer or colder than search traffic, and the published data does not yet say which.

Search Engine Journal makes the same point in plainer language: benchmarks are limited, measurement standards are still developing, and behaviour inside these platforms keeps moving. Anyone quoting you a confident return figure for ChatGPT ads in August 2026 is extrapolating from a few months of one market.

What is not speculative is the direction. A surface that started as an assistant now has an ad load, a bidding model, a conversions API and an agency ecosystem, in under a year. The interesting question for a small seller is not whether to buy the card. It is whether the answer above the card still mentions you, because that is the placement no competitor can outbid you for.