ChatGPT Ads in the UK: worth testing yet?

by Harry Watts · 18 August 2026 · updated 26 September 2026 · 9 min read

Yes. ChatGPT ads are labelled sponsored cards shown below ChatGPT's answers, and UK advertisers can buy them now: Free and Go tier users have seen them since 6 June 2026, and the self-serve Ads Manager beta takes a daily budget from £15. On cost there is still no reliable UK benchmark, and OpenAI says it has none either. Its guidance, checked on 26 September 2026, is a starting bid of US$3–5 per click, roughly £2–£4.

Whether you should is the question every "get in early" newsletter skips. It depends on two things almost nobody leads with: whether your category is allowed to advertise at all, and whether the targeting reaches the people you actually sell to. For a decent share of UK businesses, one of those two ends the conversation today, though since early September the targeting half is no longer the flat no it has been all year. For the rest, a small, properly measured test is defensible, as long as it's sized like a test, not a bet.

What's actually live

• UK users see ads on the Free and Go tiers since 6 June 2026, the first European market. Paid tiers never see them. Under-18 accounts are excluded.
• UK advertisers can buy through the Ads Manager beta, opened later in June. The old US$50,000 pilot threshold is gone, and the minimum is now a daily budget of £15. As of late August, OpenAI is even paying for first tests: invited new advertisers are being offered a £500 spend-match credit. Those terms are from the UK invite wave of late August 2026 and OpenAI's credit terms govern, as does the rule that every one of these offers is earned by spending first.
• One main format: a labelled sponsored card below the answer, carrying your name and logo, a headline, copy, an image and a link. Retailers can build campaigns from a product feed. Sponsored Agents, a chat with a business's own AI representative opened from an ad, are in a limited US test.
• Three ways to buy: per thousand impressions, per click, or optimised towards a conversion, which is billed per click or, in a beta that is widening, per thousand impressions. You never pay per conversion. Ads are chosen in a relevance-weighted second-price auction.
• Measurement exists: the OpenAI Pixel and a server-side Conversions API, deduplicated by event ID, plus UTM parameters for your own analytics. Choosing your own click and view attribution windows is promised "when available"; until then the window is OpenAI's.
• New since 16 September: suggested copy and images when you draft an ad, opt-in text customisation that adapts and translates your copy to each conversation, and HubSpot and Shopify integrations. The Shopify app opened to merchants outside the US on 23 September.

How it differs from the ads you already buy

Forget keywords. You write context hints (plain descriptions of the conversations your ad belongs in), and OpenAI's ranking decides when to show it, reading the conversation and, where a user has ads personalisation switched on, signals from their wider ChatGPT use. There are no bids on search terms and no demographics, and OpenAI documents no lookalikes or website retargeting. Customer lists arrived in September: you can upload your own customers or prospects to target, exclude or bid up, but targeting a list needs at least 25,000 matched people, which rules it out for most small businesses. Excluding existing customers works at any size. You control the location, the device, the creative, the landing page and the bid ceiling. Everything else is the algorithm's call.

Inventory is smaller than the headline numbers suggest, too. ChatGPT has over 16 million monthly UK users, but ads reach only the free tiers, minus under-18s, minus anyone who took the ads-free option, and during the US pilot, fewer than 20% of eligible users were shown ads on a given day. This is a narrow, high-attention placement, not a firehose.

What it costs, as far as anyone knows

There is no reliable UK benchmark yet. Anyone quoting one is guessing. OpenAI says so itself: the platform "does not yet have performance benchmarks" across advertisers, industries or campaign types. What it does publish: a recommended starting bid of US$3–5 per click (call it £2–£4), no recommended price for impressions or for conversion-optimised campaigns, and a minimum daily budget of £15. Where that lands against UK Google Ads CPCs depends entirely on your sector: cheaper than legal clicks, pricier than retail ones.

The early signals are real but second-hand: Reuters reported the US pilot passing US$100 million in annualised revenue inside six weeks with 600+ advertisers, and Criteo observed LLM-referred traffic converting at roughly 1.5× other referral channels across a 500-retailer sample. Evidence the channel works for someone. Not a forecast that it works for you.

First filter: are you even allowed in

OpenAI's ad policy, updated on 10 September 2026, keeps the early phase to a short list of consumer categories: household and consumer goods, local services, travel and entertainment, and digital products and education. Outside the US it excludes healthcare, financial services and legal services, and everywhere it excludes gambling, alcohol, dating and political advertising. Software for health or money, a budgeting app say, can qualify where the service itself can't. Ads are also barred from sensitive conversations (mental health, self-harm and similar), regardless of advertiser.

Read that list against the UK's highest-CPC sectors and the irony is hard to miss: the businesses paying £8+ per Google click, who'd benefit most from a cheaper channel, mostly can't use this one. If you're a solicitor, an IFA, a clinic or a bookmaker, the ChatGPT Ads question answers itself for now. OpenAI says it expects to widen the eligible categories over time; worth rechecking quarterly, not weekly.

Second filter: the targeting gap, which is narrowing

All year this has been a flat no. Targeting was country-level: no radius, no city, no postcode, so a plumber in Leeds paid for clicks from Truro to Inverness. That was still what OpenAI documented a week ago. Its help centre, as updated on 5 September 2026, said the Ads Manager beta "supports campaign-level country targeting and, in the United States, supported state, designated market area (DMA), and ZIP code targeting", and the geotargets file OpenAI publishes for advertisers listed five UK rows: the country, plus England, Northern Ireland, Scotland and Wales.

The product got there first. On 6 September 2026 the campaign location picker in my own UK account accepted LS1, Leeds and E4, Chingford: postcode districts, in Britain. The screenshot sits on the credit article, with what it does and doesn't prove: the picker takes them and the interface states the campaign will only show in selected locations, but nothing of mine has run through them yet, so delivery is unverified. On 12 September OpenAI's documentation caught up, and the same help page now says states or regions, cities, markets and postcodes are targetable "where available", with availability varying by country. One exception is new: a product-feed campaign created now is limited to whole countries.

For a local business that changes the order of operations rather than the verdict. Open the picker and search your own town before concluding anything, because OpenAI's own instruction is to do exactly that: "Search for a location during campaign setup to see the options available to you." If it offers you nothing below United Kingdom, everything above stands and tight-geography businesses get more from the same money in Google and Microsoft, where a mile radius is a checkbox. If it does offer your postcode, this became a materially more interesting channel for local advertisers in early September, with no benchmarks and no delivery evidence yet, which is a reason to test small rather than to move budget.

One thing the picker doesn't fix: with context hints instead of keywords, you still can't buy "emergency plumber leeds" the way you can in a search auction. Geography is not intent.

So who should test now

• E-commerce selling UK-wide. Product-feed campaigns, conversion bidding and now a Shopify app are being built for you, and Criteo's 1.5× conversion signal came from retail.
• Local services, where the picker offers your area. Local services are one of the categories OpenAI's policy names as open, and postcode targeting makes them reachable. Delivery through a postcode is still unproven, so this is the test that most needs its measurement in place first.
• B2B with national reach. Considered purchases, comparison-heavy buying, educational mid-funnel content: the conversations ChatGPT hosts all day.
• Anyone whose customers research before buying and whose offer fits in one card pointing at one relevant page.

Since the targeting changed and OpenAI started offering credits, I've been encouraging every client I work with to at least consider it. Considering it is where the two filters above come in.

And who should wait: excluded categories, local-radius businesses whose location picker still offers nothing below United Kingdom, anyone who couldn't treat the spend as a testing budget, and anyone whose existing accounts still leak. Fixing known waste in a proven channel beats funding an experiment in an unproven one.

If you do test, test properly

• Treat it as a testing budget. At this stage that's what it is. For some businesses ChatGPT Ads has high potential, and there the testing budget can be chunkier. I run it on its own or alongside Google and Microsoft, and adding it to either doesn't double my fee.
• Wire measurement before the first pound. Pixel and Conversions API together, deduplicated. Without it you'll be reading screenshots, not results.
• Check your robots.txt. OpenAI's rules say landing pages must not block OAI-AdsBot or OAI-SearchBot. Block either, or challenge them with aggressive bot rules, and ads fail review.
• Decide on text customisation. Since 16 September it's opt-in. Switched on, it adapts and translates your headlines and descriptions to each conversation, so wording you haven't read can run under your name.
• Write context hints like briefs, not keywords. Describe the situation your buyer is in. Ten specific variants beat one generic one.
• Judge at 60–90 days against blended cost per customer: the same arithmetic as any other channel. Then keep it or kill it on the numbers.

Quick answers

Want a straight answer for your business

A free intro call covers it in thirty minutes: whether your category is allowed, whether the targeting fits how you sell, and whether a test slice makes sense this quarter or next year. If the answer is "wait", you'll hear that. If it's "test", here's how I run ChatGPT Ads for clients.