ChatGPT ads arrive in the UK

OpenAI started testing ads inside ChatGPT in the United States on 9 February 2026, as TechCrunch reported at the time. Search Engine Land reported on 8 June that the programme would extend to the UK, Japan, South Korea, Brazil and Mexico, and on 19 June that the self-serve Ads Manager beta had begun reaching UK businesses by invitation. Serving to UK users is the step that landed this month.

Two formats are running: a static unit with a headline, description, image and URL, and a product unit carrying pricing and star ratings. Targeting works nothing like Google Ads. There are no keywords and no demographic selectors. Advertisers write freeform 'context hints' describing the conversation topics they want to appear next to, and can add geography plus first-party customer data above a matching threshold, according to Search Engine Land's 10 August walkthrough of early campaigns. OpenAI is also testing multi-advertiser units that group several ads into one placement.

One advertiser case study covering roughly 1,500 users, published by Search Engine Land on 10 August, compared ChatGPT ad traffic against Google Ads traffic for the same business. Average engagement was 17 seconds from ChatGPT against 41 seconds from Google. Conversion was 0.21 per cent against 3.71 per cent. Cost per click sat in the two to five dollar range, with a suggested minimum bid of three dollars.

That is one advertiser, one category, and a placement a few days old in most markets, so do not read it as a verdict. It does suggest something about intent. A person deep in a ChatGPT conversation is being answered, not shopping a results page, and an ad underneath the answer interrupts that rather than completing it.

The commercially valuable moment in ChatGPT is the sentence that names three products and explains why. That sentence is generated from the model's memory and from whatever pages it retrieves, not from an auction. The ad appears below it, after the recommendation has already been made.

OpenAI is explicit that ads do not influence the answers, and that advertisers get no access to conversations. Nobody outside the company can verify either claim, and the separation is a stated policy rather than an observable mechanism, so it is worth testing rather than trusting. Run your buyer questions through ChatGPT and see whether the brands it names look like the brands buying placements. If the two lists diverge, the separation is doing what OpenAI says it does.

Measure the organic side before you spend anything, because it is the part you cannot buy back later. Our free AI visibility check runs live queries about your brand through Claude with web search and needs no login, and it shows the specific claims an assistant makes about you. Paid runs cover ChatGPT, Perplexity and Gemini as well. For how ChatGPT in particular sources what it says, and what changes when you fix the pages it reads, see our ChatGPT visibility tracker page.

If you do test the ads, write context hints the way your buyers actually phrase the problem rather than as keyword lists, since the system is matching conversation topics. Keep the two metrics apart in your reporting. Ad clicks and cost per acquisition tell you about the unit at the bottom. Whether ChatGPT names you, and what it says when it does, tells you about the answer above it, and that is the number that moves when your product facts and third-party sources improve.

Also make sure the answer has something accurate to work with. Pricing, availability and specifications as plain crawlable text on your own pages, matching what review sites and directories say about you, is still the work that changes what an assistant states. In our July 2026 test of 30 brands, 13 of 30 were described with at least one materially false claim even with live web search switched on. No ad budget corrects a wrong price in the paragraph above the ad.

Ads in ChatGPT are now a real channel in the UK, and worth a controlled test if you already run paid search. They are not a route into the recommendation. The answer still gets built from what the model knows and what it can fetch, which means the unglamorous work of keeping your facts correct and findable remains the only lever on the part of the screen buyers actually read.