Key Takeaways
- OpenAI started serving ads in 31 European markets in the week of August 24, 2026, six months after the US pilot opened in February. Ads reach logged-in adults on Free and Go only, and every paid tier stays ad-free
- The targeting unit is a context hint, a written description of the conversation you want to appear in. It sits at ad group level and is matched semantically, so there's no keyword, no match type and no cookie underneath it
- The input got much longer. SimilarWeb puts the average ChatGPT prompt at 60 characters against 3.4 for a Google search, and Nectiv found 77% of the resulting search queries run five words or more
- That makes phrasing a media buying asset. You can't pull search volume for a description of a buying moment, so the only honest place to get one is from people describing that moment in their own words
- Nobody has real benchmarks yet. OpenAI's own position is that there are no cross-advertiser numbers by industry or objective, so treat published CTR figures as marketing until a methodology shows up
Edu here. A friend running a supplements brand messaged me last week asking whether he should be buying ads in ChatGPT now that they've reached Spain. Fair question. My first instinct was the boring answer, which is that it's a week old and nobody knows anything yet.
Then I went and looked at how the buying actually works, and something in it changed my mind about why this matters. It isn't the placement. It's the field you type into.
In Google, you buy a keyword. In Meta, you buy an audience. In ChatGPT, you buy a description of a conversation. That's a different kind of input, and it comes from a different place than the one most brands have been shopping in.
ChatGPT Ads replaces keyword targeting with context hints, short written descriptions of the conversations where a product is relevant, matched semantically against the live chat rather than against a typed string. OpenAI expanded the format to 31 European markets in the week of August 24, 2026, six months after beginning its US pilot in February 2026.
What actually launched, and where?
Ads reached 31 European markets in the week of August 24, 2026, including Germany, France, Spain, Italy, Poland, Sweden and the Netherlands (Digiday, OpenAI's ads business hits Europe at the six month mark, August 2026). That's the six month mark since the US pilot opened in February 2026.
The route there went US, then Canada, Australia, New Zealand and the UK in June, then Mexico, Brazil, Japan and South Korea, then Europe. Roughly one new region a month, which for an ad product that didn't exist in January is quick.
Who sees them matters more than most coverage suggests. Ads show to logged-in adults on the Free and Go plans. Plus, Pro, Business, Enterprise and ChatGPT Edu, the university plan, all stay ad-free. So the surface excludes, by design, the people paying twenty dollars a month or more for the thing, which in plenty of B2B and premium DTC categories is a meaningful slice of the buyer.
Density is low too. MediaPost reported ads appearing on fewer than 5% of searches at launch, with a maximum of one ad per search (MediaPost, OpenAI Expands Ad Pilot Across European Markets, August 25, 2026). One unit, below the answer, in a labelled box that's visually separated from the response.
Buying is agency-first for now, through Publicis, Omnicom, WPP, Havas, Dentsu and MediaPlus, with OpenAI saying it'll open Ads Manager to local advertisers in these markets within two to four months. So if you're a small brand in Europe reading this, you can't actually buy yet. Which is a decent window to think rather than spend.
Why isn't this keyword targeting?
Because there's no keyword field. The primary targeting signal in OpenAI's Ads Manager is the context hint, described as a short description of the conversations, topics or keywords where your product or service is relevant. Hints sit at ad group level, in between the campaign and the ad itself.
The recommended way to write one is audience plus intent plus topic. A published example reads: "RevOps leaders at Series A to C US healthtech companies evaluating CRM alternatives to Salesforce because they need HIPAA eligibility and EHR integration." Guidance suggests five to fifteen of these per ad group, written as variants on the same combination (Context Hints, The Definitive Guide to Targeting in ChatGPT Ads, 2026).
Read that example again and notice what it is. It's a sentence about a person in a situation. It isn't a query, it isn't a demographic, and there's no volume number attached to it anywhere.
| Keyword targeting | Context hints | |
|---|---|---|
| What you submit | A literal string someone types | A description of a conversation |
| Match logic | Token based, exact, phrase or broad | Semantic, across the whole prompt |
| Intent signal | Inferred from query modifiers | Stated directly in the conversation |
| Research input | Search volume tools | No volume data exists |
| Identity layer | Cookies, device graph | Conversation context, plus opted-in personalisation |
The bottom row gets the least attention and probably matters most. Each conversation is evaluated on its own content, which means advertisers can't retarget an individual across sessions the way cookies allowed. OpenAI states that advertisers never receive chats, chat history, memories, names, emails or precise location.
Of course that's OpenAI describing its own product, and I'd hold it loosely. But if it holds, the practical effect is that behavioural history stops being the advantage and description quality becomes it.
Context hints replace three things at once: the cookie, the device graph and the exact-match keyword. Because each conversation is scored on its own content, cross-session retargeting is not available to advertisers, and OpenAI states that advertisers never receive chats, chat history, memories, names, emails or precise location from ChatGPT ad placements.
The targeting input got seventeen times longer
SimilarWeb data covering October 2023 to September 2025 puts the average ChatGPT prompt at 60 characters, against 3.4 characters for a Google search and 10.4 for Google's AI Mode. That's roughly seventeen times the length of a search query, and six times the length of an AI Mode one.
Nectiv's study backs up the shape of it from a different direction. Analysing more than 8,500 ChatGPT prompts across nine industries, they found 31% triggered at least one search, producing 2,600 queries at an average of 5.48 words each, with 77% running five words or longer (Search Engine Land, ChatGPT performs a search in 31% of prompts, new data reveals, October 2025).
Here's the part I think gets missed. A three character query carries almost no information about the person. A sixty character one carries constraints, a budget hint, a previous product they hated, the reason they're looking at all. So the targeting unit moved from a token to a phrasing, and phrasing is not something a keyword tool has ever been able to hand you.
The bit I haven't seen said out loud: for twenty years, targeting research meant compressing human situations down into the shortest string that had volume attached to it. Context hints reverse that. You're now rewarded for writing the situation out at full length, in the words the person would use. That's the same raw material good ad copy has always needed, except now it goes into the targeting field too.
So where do you get the sentences?
Not from a keyword tool, because there's no volume for a description. Not from your own team either, and this is the failure mode I'd bet on being most common. Someone writes fifteen context hints in an afternoon using category vocabulary, then wonders why the campaign matched into the wrong conversations.
The gap is a familiar one. Brands describe their product in the language of the category. Buyers describe their problem in the language of their life, usually before they know the category name at all. Someone in a Reddit thread doesn't write "evaluating magnesium glycinate for sleep quality." They write that they wake up at 3am every night, that they already tried the gummies, and that the gummies did nothing.
What I keep seeing in corpora: when I code a few thousand comments in a category, the phrasing people use to open a problem is remarkably consistent, and it almost never matches the brand's own product page. Same problem, two vocabularies. Until now that mismatch cost you conversion on the landing page. In a context hint system it costs you the match itself, which sits upstream of everything else.
So the method is the one that was already worth doing, with a new payoff attached. Go to the places where people in your category describe the problem to each other, collect how they open it, and count what repeats. Then write your hints in that vocabulary instead of yours. The step by step version is in how to find the exact words your customers use, and the copy side of the same idea is in why your ad copy speaks your language, not theirs.
Also worth saying: this cuts both ways. The same research that gives you the hints tells you whether your buyers are even on a free tier. That's a much cheaper question to answer than a campaign is.
Your ad gets read right after an answer you didn't write
The sponsored box sits below the response, and OpenAI's stated position is that ads don't influence the answer. Take that at face value and it still leaves an awkward structural fact. Whatever the model just said about your category is the context your ad is read in, and you have no control over that paragraph.
Search-trigger rates vary a lot by vertical, which tells you how often that preceding paragraph is freshly retrieved rather than recalled from training. Nectiv found local intent triggering a search 59% of the time, against 19% for fashion and 18% for credit cards.
In fashion, four times out of five, the paragraph above your ad is the model's stored impression of your category rather than anything current. That's a reason to care about what the model already believes about you, which is the whole subject of how customer language drives AI visibility, and it's now sitting directly upstream of a paid placement.
OpenAI's ads VP Dave Dugan framed the creative shift like this: "In the attention economy, brands had to be interesting; in the intelligence economy, they have to be useful." It's obviously a pitch for his own product. It also happens to describe the format accurately. An ad arriving after a paragraph that has already answered the question has to add something the answer didn't.
What I wouldn't do yet
I wouldn't trust any performance number you read about this. OpenAI's own position is that ChatGPT Ads has no cross-advertiser benchmarks across industries, objectives or campaign types, and most advertisers running these placements have weeks of data rather than years. Every CTR and conversion figure circulating right now traces back to an agency blog with no published methodology, so I've deliberately left them out of this piece.
I also wouldn't move budget. One ad per answer, on under 5% of searches, on free tiers only, in a market where you can't self-serve buy for another couple of months. That isn't a channel yet. It's a thing that might become one.
And I'd be careful with the DTC assumption specifically. If your product sells to people who pay for their AI subscription, the ad-free tiers are exactly where those people are. That's not a small carve-out in premium categories.
What I would do is the cheap part. Find out how your buyers open the problem in their own words, because that's now useful in three places at once instead of one.
How does your buyer describe the problem before they know your category exists?
Insightios collects the threads where your market talks to itself, codes what repeats, and hands you the phrasings and objections ranked by how common they are, with the real quotes attached. Flat fee, fixed turnaround.
Frequently asked questions
Are ads live in ChatGPT in Europe?
Yes. OpenAI began serving ads across 31 European markets in the week of August 24, 2026, six months after the US pilot started in February 2026. Ads show to logged-in adults on the Free and Go plans. Plus, Pro, Business, Enterprise and ChatGPT Edu, the university plan, all stay ad-free.
How does ChatGPT ad targeting work?
Through context hints, which sit at the ad group level and are short descriptions of the conversations where your product is relevant. The system matches them semantically against the current conversation rather than against a typed keyword string. There's no cookie, no device graph and no exact match type.
What is a context hint in OpenAI Ads Manager?
A written description of a buying moment, structured as audience plus intent plus topic. A published example is RevOps leaders at Series A to C healthtech companies evaluating CRM alternatives because they need HIPAA eligibility. Guidance suggests five to fifteen hints per ad group, written as variants on one audience-intent-topic combination.
How long are ChatGPT prompts compared to Google searches?
About seventeen times longer. SimilarWeb data covering October 2023 to September 2025 puts the average ChatGPT prompt at 60 characters against 3.4 characters for a Google search and 10.4 for Google's AI Mode. Nectiv found 77% of ChatGPT's resulting search queries ran five words or longer.
Do ads change the answer ChatGPT gives?
OpenAI says no. The sponsored unit is a labelled, visually separated box placed below the response, and the company's position is that paid placement is kept separate from the organic answer. In practice this means your ad is read immediately after a paragraph about your category that you didn't write.
Should a DTC brand advertise in ChatGPT yet?
Only as a test with money you can afford to lose. There are no cross-advertiser benchmarks by industry or objective, most advertisers have weeks of data rather than years, and every paid ChatGPT tier is ad-free, so a share of your buyers will never see the unit at all.
What to do next
Don't buy anything yet, mostly because in Europe you can't. Use the two to four months before Ads Manager opens locally to do the input work, since the input is the part that won't be sitting on a shelf when it does.
That means going and reading how people in your category state the problem, at the moment before they've picked a solution. Collect the openings, not the conclusions. Count which phrasings repeat. Those become your context hints, and they double as ad copy and landing page headlines, which is a decent return on one piece of research.
Besides that, I'd check the obvious thing first. Ask whether your buyer is plausibly on a free AI tier at all, because if they aren't, none of this is your channel and you've just saved yourself a quarter.
All in all I think this is a real shift and a slow one. The targeting input moving from a keyword to a sentence is the kind of change that takes a while to show up in anybody's numbers and then looks obvious afterwards. I don't know how the format performs. I do know the sentence has to come from somewhere, and there's only one place that's ever produced good ones.
Sources
- Digiday. (2026). OpenAI's ads business hits Europe at the six month mark, August 2026. Link Retrieved September 1, 2026.
- MediaPost. (2026). OpenAI Expands Ad Pilot Across European Markets, August 25, 2026. Link Retrieved September 1, 2026.
- Context Hints. (2026). The Definitive Guide to Targeting in ChatGPT Ads. Link Retrieved September 1, 2026.
- Search Engine Land. (2025). ChatGPT performs a search in 31% of prompts, new data reveals (Nectiv study, 8,500+ prompts across nine industries), October 2025. Link Retrieved September 1, 2026.
- OfficeChai. (2025). ChatGPT Queries 17x Longer Than Google Searches, 6x Longer Than Google's AI Mode: SimilarWeb Data (SimilarWeb, October 2023 to September 2025). Link Retrieved September 1, 2026.
- ALM Corp. (2026). ChatGPT Ads Explained: Where Ads Appear, Who Sees Them, How Targeting Works. Link Retrieved September 1, 2026.