ChatGPT Ads: how the channel works and how to run a first test worth learning from

Updated: 1 September 2026. ChatGPT Ads launched across 31 European markets on 24 August 2026, and buying access now reaches the EU — the earlier version of this piece said a European brand could be advertised to but not advertise, which no longer holds.

Most teams open the ChatGPT Ads Manager and recognise it instantly. Campaign, ad group, ad. Bid, budget, pixel. The furniture is familiar, so the Google or Meta playbook comes out of the drawer, gets applied, and produces a flat first month.

The furniture is familiar. The mechanics underneath are not. This is what changes, and how we would structure a first test.

What you are actually buying

A ChatGPT ad is a labelled card that appears below an answer the user has just read and, in most cases, trusts. It carries an advertiser name, a headline, a short description, an image and a link. OpenAI states that ads run on systems separate from the chat model and that advertisers cannot shape, rank or alter what ChatGPT says.

That placement is the whole proposition. On social you interrupt someone who came to be entertained. On search you meet someone mid-query. Here you arrive immediately after a considered answer, in a conversation where the user has often spelled out their situation in detail. Copy built to interrupt reads as noise in that context.

Ads are served to Free and Go users. Plus, Pro, Business, Enterprise and Edu accounts see none, nor do accounts identified as belonging to under-18s. Temporary chats and the ChatGPT Atlas browser are excluded during the test phase.

Where you can buy it today

The Ads Manager is available across North America, Latin America, the Asia-Pacific region, the Middle East and North Africa, and — since late August 2026 — most of Europe. OpenAI keeps the current list in its Ads Manager availability article; check it for your billing country rather than trusting a list in any guide, this one included.

The rollout has been quick — US test from 9 February 2026, self-serve Ads Manager in beta from 5 May, the UK from 6 June, then the rest.

Europe closed that gap fast. ChatGPT Ads went live across 31 European markets on 24 August 2026, and buying access followed: EEA advertisers can now register through ads.openai.com, with the Ads Manager localised per market. One rule still separates Europe from the rest: personalised ads remain unavailable in the EEA and Switzerland, so matching there runs on the current conversation, approximate location and device type alone. If you are used to layering behavioural signals, that constraint is the thing to plan around.

Context hints are not keywords

This is the single biggest source of weak first results.

At the ad group level you write context hints: descriptive signals about the conversations you want to appear in. OpenAI's documentation calls them broad thematic signals rather than exact keyword matching, and describes them as covering topics, conversation themes, user needs or situations. They guide matching; they are not exact-match rules.

A hint written out of Google Ads habit:

“CRM sales software best tools comparison”

A hint written for this channel:

“A founder evaluating CRM options for a small remote sales team struggling with pipeline visibility and inconsistent follow-up”

The second one names a person, a decision and a moment of uncertainty. The system aligns it with conversations carrying the same intent. The first one hands a semantic matcher a list of tokens and wastes what it is good at.

A useful formula: persona + intent + scope. Who is this, what are they deciding, and what narrows the conversation toward your category. Read the hint aloud — if it does not sound like how you would brief a colleague on a buyer scenario, rewrite it.

The structural consequence is that ad groups stop mapping to product categories and start mapping to stages of awareness:

• Early. „A small business owner trying to work out whether they need a dedicated CRM or can keep using spreadsheets.” Educational creative, low-friction call to action.
• Middle. „A SaaS founder comparing CRM options for a team of 10–15 with both inbound and outbound motions.” Differentiators and proof.
• Late. „A VP of Sales weighing a migration off Salesforce to something lighter without losing pipeline history.” Risk reduction: trials, migration support, references.

One caveat on the folklore: a 280-character limit on context hints circulates widely in industry newsletters and appears in none of OpenAI's documentation, help centre articles or bulk upload schema. Treat it as unverified until the interface tells you otherwise. There are also no audience size estimates or reach forecasts, so you cannot check how broad your matching is before you spend. Run three or four hint variations in parallel and read the delivery pattern instead.

Creative in a hundred characters

Specifications are tight. Headline up to 50 characters, with 16–24 recommended. Description up to 100 characters, with 32–48 recommended. Square image, PNG or JPG, from 256×256 px. Longer text gets truncated. Product feed ads generate creative from your catalogue, and a multi-product carousel is in testing.

There is no room for a pitch. There is exactly enough room to demonstrate recognition:

• Search-style: „Best CRM for Startups. Try Free.”
• Conversation-style: „Still updating your pipeline by hand?”

The second works because it names the thing the reader was already describing to the assistant a moment earlier.

Landing pages carry more weight here than most advertisers expect. Someone arriving from a ChatGPT ad has been primed by a conversation in which they articulated their own situation, so a page saying „built for 10–50 person outbound teams” outperforms „the #1 CRM platform” by a wide margin. Specificity beats polish.

Bidding and budgets, precisely

Three models: CPM for impressions, CPC for clicks, and oCPC — conversion-optimised CPC — in beta since June 2026 and extended to product feed campaigns in August.

Two details are routinely misreported.

oCPC still bills per click. Your bid cap is expressed in CPA terms and the system converts it into an auction bid using predicted conversion probability, but you pay for valid clicks. It requires working measurement and a standard conversion event; custom events are not supported as an optimisation target.

Daily budget is a seven-day average. A single day can spend up to twice the amount you set, with a weekly ceiling of seven times the daily figure. Judge pacing across the week, not day by day.

On numbers: OpenAI's help centre recommends starting with a maximum bid of $3–$5 per click, and the interface warns below $3. Observed averages in the market sit in a slightly wider $2–$5 band. Reported daily minimums are around $25 with no total spend requirement. All of this moves as the auction fills, so the figure that matters is your own break-even CPC, not the platform's suggestion.

Measurement: wire it before the first dollar

Reporting here is thinner than you are used to, so the gaps have to be closed on your side.

Pixel and Conversions API. OpenAI provides a measurement pixel — the JavaScript function is `oaiq`, hence the nickname — and a server-side Conversions API with event deduplication by ID. Both shipped on 5 May 2026 alongside the Ads Manager. There is an official Google Tag Manager template in OpenAI's GitHub repository, not yet in the Community Template Gallery, so it imports manually. One server-side constraint to plan around: the API accepts events timestamped within the last seven days.

Attribution windows. View-through is fixed at one day. The click window is configurable and OpenAI does not publish a default. The 30-day figure in circulation is the pixel cookie lifetime, which is a different thing.

Effects outside the click. Assistant conversations run long and decisions land late. A meaningful share of the channel's contribution shows up as branded search and direct traffic days after the impression, so track those alongside campaign reporting or you will systematically undervalue the channel.

Custom audiences, with one regional caveat. First-party lists are supported from 25,000 matched users, but the Ads Manager states they are unsupported for campaigns targeting the EEA and Switzerland, where personalised ads are not yet available. Suppression is the fastest win this channel offers — everywhere it is available.

A 30-day test that produces answers

Four things in place before launch: pixel and conversion events confirmed firing, three or four context hint variations running in parallel, at least two creative variants per ad group, and bids at the upper end of the recommended range until you have delivery data.

Then watch three signals over the month:

• Delivery across hint variations. This is the most valuable output of the first test. It tells you which problem states your category actually appears in, which is information you cannot buy anywhere else.
• Click-through rate by creative. Whether a card earns a click mid-conversation is a direct read on whether it lands as relevant or as an interruption.
• Branded search movement. Measured in Search Console in the two weeks following launch.

What you should have after thirty days: a map of the conversations your category lives in, a baseline CTR, a read on branded lift, and two or three concrete decisions for the next cycle. What you should not expect is a clean ROAS figure. That is not what a first test on a channel this young is for.

Where the channel fits

Strong fit: anything bought after research. B2B software, professional services, education and certification, and e-commerce categories where buyers ask about ingredients, specifications or suitability. If your product has a comparison page, buyers are already having these conversations.

Weaker fit: impulse and visually-driven categories — fashion, accessories, décor — where the purchase happens through images rather than description.

Check the policy first. Gambling, alcohol, tobacco, adult content, politics, drugs, legal services, and housing and job listings are disallowed. Financial and health services are restricted: permitted case-by-case for licensed advertisers in the US, generally not outside it. Ads are also withheld from conversations about health and other sensitive contexts.

One demographic note. Pew Research Center data published in June 2026 (fielded in February) puts US ChatGPT usage at 61% among adults aged 18–29 and 19% among those 65 and over. Ads reach Free and Go tiers only, so if your buyer skews older, the addressable slice today is narrower than the headline user numbers suggest.

FAQ

The advantage available right now is not clever bidding. It is that most advertisers are still running search copy against keyword-shaped hints and concluding the channel does not work. Writing hints as buyer scenarios, creative as recognition rather than pitch, and measurement wired before launch is, for the moment, itself the edge.

If you want a second opinion on whether this channel fits your funnel, let's talk — or hand ChatGPT Ads management to us end to end.