August 17, 2026

ChatGPT Ads Reach Europe, and the Advertisers Are Almost Never the Sources

Period of August 11 to August 17, 2026. Last period was about who gets through your door. This one is about what the door costs and whether you can see who walked in. OpenAI notified European users that advertising arrives in ChatGPT this month, ending the assumption that the EU had a long runway. Independent measurement of that ad surface shows the sponsored slot and the cited-source slot are almost entirely different inventories, which means a budget cannot buy you into the recommendation. A first merchant-side instrument for agent traffic shipped, built on server logs rather than JavaScript. And a consumer study identified the condition that flips people into letting an agent buy for them, which turns out to be a returns policy.

Last week we covered the Ninth Circuit holding that an agent acting for a user is the user, Cloudflare issuing agents an identity and a capped wallet, and Shopify reporting AI orders tripling. The through-line was access: who may reach your storefront, and on whose authority.

This period moved one step downstream. If agents and AI answers are a real channel, two questions follow immediately. What does it cost to be visible there, and can you measure any of it? Both got answers this week, and the European answer arrived earlier than almost anyone had planned for.

Here’s what happened.

Europe gets the paid slot, and it starts unpersonalised

On 15 August, OpenAI Ireland Limited emailed ChatGPT users across the European Economic Area and Switzerland to say advertising will begin appearing on the Free and Go plans later this month, with the privacy policy updated to describe how ads are selected, measured and controlled. The mechanics are deliberately conservative. Ads are not personalised at launch. Selection runs on the topic of the current conversation plus limited contextual signals such as general location and device type, with past chats and memories explicitly excluded. Personalisation requires an opt-in prompt rather than arriving by default, which is the opposite of the US configuration. Plus, Pro, Business, Enterprise and Education accounts stay ad-free.

That lands on top of a rollout that has been moving all year, and it is worth getting the sequence right because a lot of coverage this month has compressed it. The pilot began in the US on 9 February, reached Canada, Australia and New Zealand on 26 March, then went live in the United Kingdom on 6 June, making Britain the first European market, followed by Japan and South Korea on 22 June, the same day the self-serve Ads Manager beta opened to UK businesses. Brazil and Mexico followed in early August, bringing the pilot to nine markets before the EEA notice landed. Several outlets reported an 11 August announcement as a five-country expansion including the UK, Japan and South Korea. Those three were already live, so treat that framing as an error worth not repeating.

Why it matters for merchants: We have been writing since catalogues started becoming ad inventory in May and again after Cannes turned the ad into the store that the AI answer would end up sold. For European merchants the honest position a week ago was that this was coming but not yet here. That position is now gone, and the runway was shorter than the whole industry assumed.

Update: it landed on schedule. Ads went live across 31 European markets on 24 August, the 27 EU member states plus Iceland, Liechtenstein, Norway and Switzerland, unpersonalised as described here.

The unpersonalised launch is the part with real operational consequences, and it cuts in a direction that favours anyone who has done catalogue work. When an ad system cannot use profile history, the topic of the live conversation carries nearly the whole targeting burden. Selection has to lean on what the person just asked and how well an advertiser’s offer matches that semantically. In the US an advertiser can partly compensate for a vague product catalogue with accumulated behavioural signal. In the EEA, at least at launch, there is much less of that signal to compensate with. The same discipline that gets you into the organic recommendation, which is precise, well-attributed, category-legible product data, is also what makes you matchable in the European ad auction.

One thing not to assume: being able to see ads and being able to buy them are separate rollouts. Self-serve buying reached US businesses on 5 May and UK businesses on 22 June, and OpenAI’s own documentation says availability varies by country and continues to evolve. If you operate in the EU, confirm whether you can actually purchase a placement rather than budgeting for one you cannot yet buy.

The measurement: 96 percent of advertisers never appear as a source in their own answer

On 10 August, SE Ranking published a study of more than 50,000 US commercial prompts across 20 niches, finding sponsored placements on 25.94 percent of them. Two further findings matter more than the headline. Around 14.35 percent of ads shown were unrelated to the prompt that triggered them, ranging from 2.6 percent in Pets to more than half in Relationships and in News and Politics. And the overlap between advertising and being cited is close to nothing: only 3.63 percent of advertisers appeared as a source in the response sitting above their own ad, with the exact advertised URL turning up in citations 0.09 percent of the time. Put the other way round, 96 percent of paid placements sat under an answer that did not reference the advertiser at all.

Why it matters for merchants: That figure settles an argument a lot of teams are having internally right now.

Paying for the placement does not get you into the answer. In a conventional search result, the ad and the organic listing at least compete for the same attention on the same page, and a strong ad position carries some halo. Here the answer is generated first and the ad is attached underneath it. Nothing in the SE Ranking data suggests the advertiser list feeds into which products the model recommends, and the near-zero citation overlap is what you would expect if the two processes never touch.

So treat them as two channels with different mechanics. One is bought media. The other is a function of whether your catalogue, specs, stock and prices are machine-readable, which is the work most catalogues still have not done. Budget accordingly and stop managing AI visibility as one line item. If you want to know which of the two you occupy today, the Shopify Feed Previewer shows what an AI actually reads of your products, which is the input to the unpaid slot. Note the study covers US prompts only, so read the ad-density number as indicative rather than as your market’s.

Agent traffic becomes visible, and it was never going to show up in Google Analytics

On 13 August, OtterlyAI launched Agent Analytics, which reads server logs to show which AI agents and crawlers reach a site, which pages they hit, and how that maps to whether the site gets cited in AI answers. The reason this needs to be a product is a technical detail with large consequences: conventional web analytics run on JavaScript, and agents request pages directly from the server without executing client-side scripts, so their visits never appear in an analytics dashboard. The tool separates three classes of visitor that most merchants currently lump together: on-demand fetchers that are answering a live question for a real person, search index crawlers, and training scrapers. It ingests via Cloudflare Workers, WordPress, Netlify, webhooks or direct log upload, processes no personally identifiable data, and had more than 50 sites tracking before launch.

Why it matters for merchants: Ignore the vendor and keep the category. You have been flying blind on the most important traffic segment of the last eighteen months, and the fix does not require buying anything, because your server logs already contain this. Every conversation about agent traffic that begins “we don’t really see much of it” is a conversation being had by someone looking at a JavaScript dashboard that structurally cannot see it.

Internalise the three-way split, and get the names right, because the naming is where merchants go wrong. Anthropic documents three separate user agents: ClaudeBot collects content for model training, Claude-User fetches a page because a person just asked Claude something, and Claude-SearchBot works on search result quality. OpenAI splits the same way between ChatGPT-User and OAI-SearchBot. A training scraper takes your content and gives you nothing this quarter. An index crawler builds the catalogue you might get retrieved from later. An on-demand fetcher is a customer, mid-question, right now. Those three deserve different treatment in your robots.txt, rate limits and bot rules, and blocking the wrong one costs you live customers.

One limit worth stating plainly. This only catches agents that declare themselves. As we noted last week, the Ninth Circuit ruling turned on the fact that Perplexity’s Comet relays through the user’s own machine, arriving from their IP in their browser session, which makes it indistinguishable from the customer at the network layer. Logs give you the declared traffic, not all of it. That is still far more than zero, and it is the same reason signed agent identity keeps returning as the only real sorting mechanism.

What actually unlocks a delegated purchase: your returns window

RTB House published “Who’s Buying? Consumer Trust in the Age of Agentic AI” on 10 August, based on fieldwork across June and July 2026 with 1,840 respondents in the United States, United Kingdom, France and Japan. The discovery numbers are strong: 59 percent of US consumers credit AI with surfacing brands they had not heard of, and 63 percent have used AI to build an initial product list. The friction number is more interesting than it looks: 42 percent of US consumers say AI lengthens their final purchase decision, rising to 48 percent among US Gen Z against 32 percent outside the US, because the assistant hands them more options to weigh rather than fewer.

The finding to keep is the conditional one. 42 percent of US millennials said they would delegate a $250 purchase to an AI agent when a seven-day return window applies. Without that return protection, the figure drops to 34 percent. Retail Dive adds that 35 percent across generations would prefer a human to review the transaction before an agent buys, rising to 44 percent among baby boomers, and that 44 percent trust AI tools for shopping decisions against 59 percent for friends and family, with AI now ranking above influencers, media outlets, TikTok and Instagram.

Why it matters for merchants: An eight-point swing in willingness to delegate, tied to a returns clause, makes your returns policy a conversion lever in the agentic channel rather than a cost centre. That is an unusually cheap thing to act on. If your return window is under seven days, or is stated in a way a machine cannot parse, you are losing delegated purchases at the moment of authorisation and the loss will never appear in any report you run.

Say it in structured data, not only in a policy page written for humans, because the agent constructing the purchase is reading fields. In practice that means schema.org MerchantReturnPolicy on your product markup, with merchantReturnDays, returnFees and applicableCountry populated, plus the equivalent return policy attributes in Google Merchant Center. European merchants have an advantage here that almost nobody publishes: the statutory 14-day right of withdrawal on most distance sales already clears the seven-day bar that moved US delegation rates by eight points. It is worth putting in a field rather than assuming an agent infers it from your jurisdiction.

The “AI lengthens my decision” finding cuts the other way and deserves a moment. It contradicts the tidy story where agents compress the funnel. What is happening is that the assistant widens the consideration set, which is good for you if you were previously invisible and bad for you if you were previously the default. That is the same asymmetry Shopify’s long-tail split showed last week, seen from the shopper’s side.

Bookings move inside the chat, and the merchant loses sight of the channel

On 10 August, Yelp brought Reservations and Waitlist into ChatGPT, letting users in the US and Canada book a table or join a queue without leaving the conversation, with changes routed back through Yelp. Resy, owned by American Express, launched reservations in ChatGPT for US restaurants the same day, and OpenTable is in the same reservation surface, so this is a category move rather than one deal. Morning Consult research commissioned by Yelp, covering 2,202 US adults in late February 2026, found 65 percent consider the ability to take action on trusted platforms important when using AI tools for local discovery, and Yelp’s announcement did not say whether ChatGPT-sourced bookings will be distinguishable in Guest Manager reporting.

Why it matters for merchants: OpenAI stepped back from taking payment in the chat when it retired the first Instant Checkout in March, but it never stepped back from taking the action. Booking a table is a commitment. It just is not a card charge, and it is moving in-chat regardless.

The attribution problem is immediate and it applies well beyond restaurants. A conversion now originates in an interface you cannot instrument, arrives through a partner’s system, and lands in your book with no channel attached. If you run any booking, quoting or appointment flow through a third party, ask that platform one question this month: will inbound from an AI assistant carry a distinguishable source in the data you receive. If the answer is no, your AI channel will look like it produces nothing, and you will underinvest in it for exactly as long as that stays true.

There is a second-order problem too. The same surface that routes the booking also sells advertising on the answer above it. Nobody has published how availability, ranking and sponsorship interact in a local recommendation, and that question is better asked early than after a competitor buys the slot.

The European angle: the runway closed this week

A week ago the defensible read for an EU merchant was that paid placement inside AI answers was a US and UK phenomenon with time to spare. The 15 August notice ended that. Ads land on EEA and Swiss Free and Go accounts this month, administered by OpenAI Ireland, and the compliance shape is visible in the design: contextual by default, personalisation only on explicit opt-in, past chats and memories out of scope at launch. That is a consent-based architecture, which is the predictable consequence of doing ad personalisation under GDPR rather than anything to do with the AI Act transparency duties that took effect on 2 August. Those cover chatbot disclosure and synthetic content marking, not advertising, and conflating the two will send your legal review down the wrong path.

The practical European consequence is the inversion of the advice we would have given seven days ago. There is no longer a window in which product data is the only route into the answer, because the paid route is arriving now. What survives is the stronger half of the argument: with personalisation switched off at launch, European ad selection leans harder on the topic of the live conversation than on accumulated user profiles, so semantic match between the question and your offer does more work here than in the US. That is an inference from the announced design rather than something OpenAI has quantified, but it points the same way as everything else on this page. Clean product data is the input to both slots in Europe, not just the unpaid one.

Two further notes. Buying and being shown are separate rollouts, so an EU merchant should verify self-serve availability in its own market before assuming a budget can be deployed at all. And the RTB House sample includes the UK, France and Japan as a combined non-US group reporting AI lengthening the purchase decision at 32 percent against 42 percent in the US, so the widened consideration set is a smaller effect in Europe so far. No France-specific breakdown was published, so do not read a national conclusion into that.

What moved this period

Development	What happened	Why a merchant cares
ChatGPT ads reach the EEA	OpenAI Ireland notified EEA and Swiss users on 15 August	Paid placement in AI answers arrives in Europe this month
Unpersonalised at launch	Contextual signals only, personalisation on opt-in	Semantic match matters more than profile history
SE Ranking ad study	Ads on 26% of US commercial prompts, 96% of them uncited	Buying the slot does not put you in the answer above it
OtterlyAI Agent Analytics	Server logs split fetchers, crawlers and scrapers	Invisible to JavaScript analytics, visible in logs
RTB House trust study	$250 delegation: 42% with 7-day returns, 34% without	Your returns policy is an agentic conversion lever
Yelp, Resy and OpenTable	Reservations booked inside the conversation	Actions move in-chat with no reliable attribution

What merchants should do this period

1. Split your AI spend into two lines this week and write down the ratio. One line is sponsored placement, the other is product-data work. Only 3.63 percent of ChatGPT advertisers appeared as a source in the answer above their own ad, so these buy different things and a single budget hides which one you are actually funding. If you are in an ad-enabled market, the placement is now purchasable. If you are in the EEA, confirm whether you can buy at all before you plan around it.

2. Pull your server logs and count three things separately. On-demand fetchers such as ChatGPT-User and Claude-User, index crawlers such as OAI-SearchBot and Claude-SearchBot, and training scrapers such as ClaudeBot. You do not need a vendor to start: grep the user agents. Then check what your robots.txt and bot rules currently do to each of the three, because a single blanket policy almost certainly blocks live customers or feeds scrapers, and quite possibly both.

3. Put your returns terms in a field, not a page. Willingness to delegate a $250 purchase moved from 34 to 42 percent among US millennials on the presence of a seven-day return window. Populate schema.org MerchantReturnPolicy with merchantReturnDays, returnFees and applicableCountry, and mirror it in your Merchant Center return policy settings. European merchants: your statutory 14-day withdrawal right already clears the bar, so publish it rather than assuming it is understood.

4. Ask every booking or quoting platform you use whether AI-sourced traffic is distinguishable. Yelp, Resy and OpenTable all now carry reservations inside ChatGPT, and Yelp’s announcement did not say whether those bookings are separable in the operator’s own reporting. If your third-party systems collapse AI-originated demand into an undifferentiated bucket, your AI channel will read as zero and you will fund it accordingly. Get the answer before the holiday season, not after.

5. European merchants: re-run your AI channel plan against a live ad surface, not a future one. Anything written on the assumption that EU paid placement was a 2027 problem is now out of date. The immediate work is unchanged and more urgent: prices, stock, variants and returns terms machine-readable, because with personalisation off at launch the conversation topic is doing the matching, and that is decided by how legible your catalogue is.