From Monday, 24 August, the layer that answers “where do I buy this” starts selling the answer.

OpenAI extends ChatGPT Ads to 31 further European markets (opens in new tab) that day, Germany, Austria and Switzerland among them, following a US pilot that began in February. The company says the ads will be clearly marked and kept separate from the answers themselves, and positions the format for the moment when someone is researching, comparing options or preparing a decision.

Read the booking terms rather than the promise, and the sentence that matters sits further down. At launch, placements are bookable through OpenAI’s own Ads Solutions team and through partner agencies and technology partners; the named agencies are Publicis, WPP, Omnicom, MediaPlus, Havas and dentsu. The self-serve Ads Manager (opens in new tab) is promised for “later this quarter”.

How far that reaches is set out in OpenAI’s own availability table (opens in new tab): the Ads Manager currently works in nine countries — Australia, Brazil, Canada, Japan, South Korea, Mexico, New Zealand, the United Kingdom, the United States. Not one of them is in Europe. Anyone outside those countries can, for now, only register interest.

Then there is the question of who the ads reach at all. They appear only on the Free and Go tiers (opens in new tab); Plus, Pro, Business, Enterprise and Edu stay ad-free, as do accounts presumed to belong to minors. And in the EEA and Switzerland, OpenAI is not serving personalised ads initially — only the current conversation feeds the selection. The paid layer is therefore narrower in this region than the launch date suggests: it reaches the free tier and the cheapest paid one, and it reaches them without audience data.

The shelf goes on sale from Monday, and a mid-sized operator cannot book it themselves on the day. They wait for a tool due in the same quarter.

What the layer cites is not the business

Days before the launch date, a Viennese agency published a first measurement of what this layer actually does with a local trade. SlopeLift ran 14 bakery brands through 80 prompts across six systems (opens in new tab) — ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, Copilot, Gemini — with fieldwork in July.

Two names lead, and neither is the market leader: Joseph Brot at 45.3 percent brand coverage, the organic bakery Öfferl just behind at 44.4. On citations the order reverses, with Öfferl’s site pulled as a source 2,098 times against Joseph Brot’s 1,877.

Then the figure that reframes the exercise. The most-cited sources are not bakeries at all. Google, wien.info, Reddit and a lifestyle magazine are drawn on more often than any individual bakery’s website, and a single Reddit thread about the best bakeries in Vienna is cited more than three hundred times.

That is worth sitting with before Monday. The layer about to sell placement is already assembling its answers largely out of surfaces that belong to other people. Anja Höbarth-Furniss, who leads SEO at the agency, puts it plainly: these systems are no longer neutral search engines, they recommend and they link.

One caveat belongs in this paragraph rather than in a footnote. This is an agency that sells visibility consulting, measuring the value of visibility consulting. The methodology is disclosed and the citation counts are checkable, which is more than most such studies offer — but the interest is real and it points one way.

The layer does not only choose, it narrows

Published the same week, and from a very different corner: researchers at MIT Sloan Management Review (opens in new tab) ran a lab study with 104 participants and a global field experiment with 245, comparing ordinary search against a tool built to pull from semantically distant clusters rather than the most popular hits.

Their uncomfortable result is not about creativity scores. It is that with ordinary search, domain experts showed no statistically significant advantage over novices. The authors call the effect ideation bubbles: because everyone queries the same infrastructure, several people independently arrive at the same idea.

Set that beside the bakery numbers and the shape of the problem changes. It is not simply that the recommendation layer decides who gets named. It is that the layer compresses the range of what gets considered at all — and from Monday, part of that compressed range is for sale.

The lever that is left is the one nobody is selling

If placement cannot yet be bought directly and the answer is built from third-party surfaces, then the work is not on the website and not in a media budget that has no route in.

It is in being present, accurately and in the business’s own words, on the surfaces these systems actually draw from — the city portal, the forum thread, the regional listing, the review that a model will read as evidence. That is unglamorous, it is slow, and it is the only part of this that is currently available to a company without an agency holding on retainer.

What to check on Monday

This note is deliberately published before the event rather than after it, which means it can be wrong in public. Here is the checkable list:

  • Do ads appear in German-language ChatGPT sessions at all on 24 August, or does the rollout slip?
  • Are they marked as clearly as promised, and genuinely separated from the answer?
  • Does an advertised business start appearing in the organic answer as well?
  • Is there a route in without an agency on the day, and when exactly does the Ads Manager arrive?

The last one decides whether this is a wall or merely a head start. Everything else is detail.

Apparatus: The rollout date, market count, sales routes and the Ads Manager commitment come from OpenAI’s own announcement (opens in new tab), 18 August 2026; the list of partner agencies from internetworld.at (opens in new tab), 20 August 2026. The US pilot began on 9 February 2026 and has since been extended to eight further markets. Ads Manager availability, the tier boundaries and the note on initially non-personalised ads in the EEA and Switzerland come from the OpenAI help centre, as of 22 August 2026. The AI-visibility figures are from the SlopeLift “AI Visibility” study on Viennese bakeries, published 13 August 2026 with fieldwork in July — 14 brands, 80 prompts, eight categories, six systems — reported by INTERNET WORLD Austria (opens in new tab) and medianet.at (opens in new tab); the quotation is from Anja Höbarth-Furniss, CSO and Head of SEO at SlopeLift. SlopeLift sells AI-visibility consulting, and that interest is disclosed here rather than footnoted. The search-and-creativity findings are from Moran Lazar, Hila Lifshitz, Charles Ayoubi and Hen Emuna, MIT Sloan Management Review (opens in new tab), 20 August 2026. Nothing here reports what happened on 24 August, because at the time of writing it had not yet happened. Source intake: Radar 2026-W34, cards 01 to 03.