Before the budget and before the creative, this channel asks a question the others do not: which country are you selling into. Financial, health and legal services are open in the United States and prohibited everywhere else.
That is not a reading between the lines. The advertising policy says that ads for financial services outside the US are generally prohibited, the same for health services, and that ads for legal services outside the US are currently prohibited. For a broker, an accountancy practice, a wealth manager, a law firm or a clinic, the answer to whether this channel is worth testing depends entirely on where the ad will be shown.
And the geography keeps deciding things after that. Personalised ads and custom audiences are both switched off for campaigns targeting the European Economic Area and Switzerland. Which leaves four levers, and the main one is the sentence the person is in the middle of typing.
This page sets out who is allowed in, what can be targeted where, what you are billed on, what the measurement does by default, and the European deadline that will eventually let you see what your competitors are running. It describes a product that is still in beta and still moving: the bidding controls change three days after this was written.
Check you are allowed in, because the answer depends on your country
This is the first thing to do, before the budget and before the creative, and it settles the matter entirely for a large share of service businesses.
Four categories are supported: household and consumer goods, local services, travel and entertainment, and digital products and education. The policy adds that all other categories are disallowed at launch.
Three sectors turn on geography, and the wording matters. Ads for financial services outside the US are generally prohibited. The same for health services. For legal services the policy says they are currently prohibited outside the US.
Read that from each side and you get two different products.
From the United States, those three sectors are open, and open with conditions rather than freely. Financial advertisers may be allowed on a case-by-case basis after approval, across a documented list: auto loans and leasing, credit cards, credit monitoring, deposit accounts, financial planning, insurance, brokerage and investment, mortgages, personal loans and payment services. Health has its own approved list: consumer medical devices, wearables and apps, dental, supplements, disease awareness campaigns without political advocacy, health insurance, hospitals and urgent care, testing and diagnostics, minimally invasive cosmetic procedures, and vision. Legal services require the advertiser to be licensed to practise in the jurisdiction where the ad is shown.
From anywhere else, including the United Kingdom, a broker, an accountancy practice, a wealth manager, a law firm or a clinic has no access to this channel. That is not a budget question or a creative question.
The general prohibitions apply everywhere: alcohol and tobacco, gambling, adult content including dating and sexual health, politics, individual job and housing ads, recreational drugs, counterfeits, debt settlement and credit repair, scams, sensitive topics and events, unsubstantiated wellness claims, and unapproved or experimental treatments.
Plus one rule specific to this surface, and creative teams should know it: imitating the ChatGPT interface is prohibited.
Two process notes. Review is mostly automated with human oversight. And the publisher reserves the right to reject an advertiser for any reason.
What you can target, and the question the documentation will not answer
What is removed in the European Economic Area and Switzerland is written twice. Personalised ads are not initially available in the EEA or Switzerland. And custom audiences are not supported for campaigns targeting the EEA or Switzerland, where personalised ads are not yet available. No arrival date is published.
What exists where personalisation is available. Custom audiences take email addresses, phone numbers, SHA-256 hashed versions of either, and advertising identifiers. They can include, exclude, or apply a bid multiplier from 0.1x to 10x. The matching floor is 25,000 matched users for inclusion and multipliers, with 100,000 recommended. That floor is worth knowing before you promise anyone a retargeting plan: a list of eight thousand customers does nothing here.
And now the question a British advertiser cannot answer from the documentation. The restrictions name the European Economic Area and Switzerland. They do not name the United Kingdom, which appears as its own entry in the availability table, so the publisher is not using a regional grouping that would sweep it in. The reasonable reading is that a UK advertiser has personalisation and custom audiences. But the publisher never says so. We are inferring it from an absence, and we would rather say that than present a deduction as a fact. If this decides your plan, put the question to the ad team in writing.
What everyone gets, wherever they are. Four levers.
Geography, at country, region, city or postcode level where available. Worth noting: without an explicit choice, a campaign can target every available location.
Platform: iOS app, Android app, desktop web, iOS web, Android web.
Context hints, written per ad group. The documentation warns against mistaking them for keywords: they are not exact-match and they guarantee no delivery in any particular conversation.
And above all the conversation in progress, which is not a setting but the engine of the system. The documentation lists the signals: the context and intent of the current conversation, the ad’s landing page, headline and text, the advertiser’s context hints, and, where ad personalisation is on, certain signals from the user’s wider experience. If a user turns personalisation off, they continue to see ads based on their current thread. The live conversation always feeds the targeting. What switches off is history and memory.
In the other direction, and this deserves credit. Advertisers receive none of it. The documentation states that advertisers do not have access to your chats, chat history, memories or personal details. Only aggregated reports come back. Placement exclusions exist around mental and personal health, emotional reliance and politics.
What is not documented in anything we could open: retargeting from site visitors, lookalike audiences, and demographic or interest targeting. We cannot say those do not exist. We can say they are not in the documentation, which is not the same thing and is the honest version.
What you pay, and the one number the channel refuses to give
The billing model. The channel supports impression and valid-click billing. Three objectives exist: views, billed per thousand; clicks, billed per click; and conversions, billed either on an optimised cost per click or an optimised cost per thousand. On that last one the documentation is unusually direct: you do not pay per conversion. As everywhere else, the objective on the label is not the event on the invoice.
The mechanism. A relevance-weighted second-price auction.
And the bidding controls change on 9 November 2026, which is three days after this page was written, so check it against your own account. Target CPA replaces maximum bid for new conversion campaigns, Target CPC becomes available for new click campaigns, and maximum bid and maximum CPC are withdrawn from existing campaigns on 1 January 2027. The corresponding programmatic fields arrive on 18 November. If you are reading a guide to this channel written before November, its bidding section is already out of date.
The published floors, and they do exist. The minimum daily budget is 25 US dollars, 15 pounds or 15 euros, set out in a currency table covering 38 currencies. That is worth stating plainly because the channel’s reputation for publishing nothing is only half deserved: the floor is public, it is low, and it is one of the few numbers here that is.
Payment. A card is required, billing is in arrears, and the first invoices trigger at 25 dollars, with the threshold rising as payment history builds.
Spend pacing. The daily budget is an average over seven days. A single day can reach twice the budget, and a week will not exceed seven times it.
Now the number you will not get. The channel publishes no floor bid, no rate card, no average cost per click and no average cost per thousand. The only figure resembling a price is a starting suggestion: begin with a maximum bid of three to five dollars per click. A recommended opening bid is not a market price.
And the channel states, itself, that it has no benchmarks: it does not yet have performance benchmarks across advertisers, industries or campaign types. That is a rare and honest sentence, and it is the only defensible answer to what this costs. Every figure you read elsewhere is either a third-party reading of a different market, or that opening bid recycled as a tariff.
The formats, and the one you cannot have
There is no official catalogue of formats carrying commercial names. The accurate labels are the ones in the technical documentation, and they describe four things.
The standard ad appears beneath the ChatGPT response. It consists of the advertiser name, a favicon, a headline, body text, an image and a destination page. The published constraints are a headline of 3 to 50 characters, body text of at most 100 characters, and an image of at least 640 pixels on a side. The channel states that ads are clearly identified as sponsored and visually separated from the ChatGPT response.
The product ad is generated from a catalogue feed and shows images, titles, star ratings, prices, sale prices and brand. A useful point: products in that feed serve the ads only, never the organic answers.
Hotel ads are in limited beta.
Sponsored agents are a format that lets people chat with a business’s AI representative. They are in alpha, restricted to selected advertisers, and no access requests are being accepted. It is the format that will get talked about, and it is precisely the one you cannot have.
Two absences worth noting. No video format is documented. And ads appear neither in temporary conversations nor in the publisher’s own browser.
Measurement, and the consent setting that is on by default
The measurement stack is complete for a channel this young, and it carries one default setting that an advertiser in a consent jurisdiction has to change.
What is measurable: impressions, clicks, spend, click-through rate, average cost per click, average cost per thousand, conversions. At campaign, ad group and ad level, with export. Refresh delays are published and they are uneven: about a quarter of an hour for impressions and clicks, seven to eight hours for spend, twenty-four to forty-eight hours for conversions.
The pixel. It captures a click identifier passed as a URL parameter and writes it into a first-party cookie lasting thirty days, with a second cookie lasting a year. It offers advanced matching on an email address or phone number hashed in the browser, and it detects site forms automatically.
Here is the trap, and it is explicit in the documentation: the pixel is active by default. There is a call that declares a consent refusal and deletes both cookies, but you have to make it yourself, before initialisation. An advertiser in the United Kingdom or the European Union who installs this pixel without wiring their consent platform is therefore dropping trackers with no prior collection. That is a compliance matter, not an implementation detail, and we have set out the general case in pixels, browsers and server-side conversions.
The conversions API exists, server to server, with its own key and deduplication by event identifier. Its event window is seven days.
Attribution. Thirty days after a click, one day after a view, by default, with windows configurable since late September 2026. Worth knowing: the channel models part of its measurement and says so, noting that reported conversion totals may include modelled conversions where available. Not all of your numbers are observed.
Two limits set out expressly. Dynamic macros in URLs are not supported, only static parameters work, which complicates per-campaign tracking. And the publisher acknowledges in advance that figures will diverge from analytics tools, noting that a difference does not necessarily indicate an error.
One sentence to re-read in six months. The documentation says data collected by the pixel is not currently used to personalise ads at user level, and adds that if that changes in future, the opt-out setting will be respected. That is written in the future tense, it is public, and it announces a change.
And a technical requirement that gates everything else: the ad validation crawler must be able to reach your landing page. The documentation is imperative on allowing OAI-AdsBot. A web application firewall that answers it with an error blocks the ad from being approved. That crawler has nothing to do with model training: the documentation states the data it collects is not used to train generative AI foundation models.
The ad library that does not exist yet, and will not cover you
This is the most useful date to have on this channel, and almost nobody is carrying it.
The fact, and it is counter-intuitive. On 31 August 2026 the European Commission designated ChatGPT a very large online search engine under the Digital Services Act, not a very large online platform as is often written. The Commission describes it as a hybrid service that qualifies as a search engine. Reddit and Roblox were designated the same day, as platforms.
What the designation is not. It is not a finding of infringement and not a penalty. It is a classification, and it triggers enhanced obligations. The compliance period is four months from notification, which the Commission states as by January 2027. The notification date itself is not published, so no exact day can be given.
What follows for an advertiser. Article 39 imposes a public ad repository on very large platforms and on very large search engines that present advertising on their interface. The minimum content is specific: the content of the ad, who paid for it, the period it ran, the targeting and exclusion parameters, and the total number of recipients reached, broken down by member state.
No such repository exists on this channel today. We looked, including on the publisher’s own regulatory and transparency pages. So you cannot currently see what your competitors are running here, and European law provides that you will be able to.
And now the part that matters most to an American or British reader. That obligation comes from European law and reaches European delivery. There is no equivalent repository for ads delivered in the United States or the United Kingdom, and nothing on the horizon creates one. This is the same pattern we found on TikTok’s library, which covers thirty-two countries and not the United States, and on LinkedIn’s, which shows impression ranges and targeting parameters only for ads aimed at the European Union.
The pattern is consistent enough now to state as a rule: advertising transparency is a European product. If you buy outside Europe, you buy in the dark, and so does everyone watching you.
How many people, and why no figure answers the question
The global figure: 1.2 billion people a week. That is what the publisher states, and it attaches no definition of use to it. It covers every plan, paid tiers included.
The European figure, published twice and differing by 26 million. The publisher declares approximately 185.6 million average monthly active recipients for ChatGPT search in the European Union, for the six months ending 30 September 2026, adding that the figure is computed solely for regulatory compliance and should not be relied on for other purposes. The European Commission’s own supervision page, updated 28 October 2026, shows 159.1 million. Both numbers originate with the publisher. As on every platform under this regime, these are self-declared figures the Commission reproduces as such, not audited measurements.
Note also what the European figure covers: ChatGPT search, the search functions, not the whole product.
The United States: 44% of adults. The Pew Research Center finds 44% of American adults saying they ever use ChatGPT, in a survey of 5,119 adults conducted between 17 and 23 February 2026, up from 18% in 2023, 23% in 2024 and 34% in 2025. One caveat on that series, and Pew is transparent about it: until 2025 the question was only asked of people who had heard of ChatGPT, and the 2026 wave also covers other chatbots. The level is solid, the trend is fragile.
The United Kingdom: no figure. The publisher publishes no country-level count, and the national regulator’s measurement was not accessible to us.
And here is why none of this sizes your audience. Ads are shown only to users on the Free and Go tiers. Plus and Pro carry none. Business, Enterprise and Edu carry none. Minors receive none. There is also an ads-free setting on the Free tier, which removes advertising in exchange for lower usage limits and reduced features.
So the number of people actually exposable to your ad is published nowhere, in any market. Not in the United States, not in the United Kingdom, not in Europe. We are not going to estimate it, and neither should anyone selling you a plan for this channel.
What does not hold up
This channel is new, so the folklore around it is dense and recent. Here is what we were able to trace.
“The cost per thousand is sixty dollars.” The figure existed, and it does not describe what you can buy. It comes from the American pilot phase of February 2026, in managed buying through agencies, reported by the trade press. It does not describe the current self-serve product. A revealing detail: the only occurrence of a sixty dollar cost per thousand in the channel’s own documentation is a worked example of unit conversion.
“Entry starts at two hundred thousand dollars.” Same origin, same pilot phase, same press. The current self-serve product publishes no minimum at all and is explicitly open to small and medium businesses.
“A click costs three to five dollars.” That is a recommendation turned into a tariff. The documentation suggests starting with a maximum bid of three to five dollars per click. A recommended maximum bid is not a price, it is a starting point.
“The click-through rate is 0.68%.” Those are third-party panel measurements whose denominator is not what you think, and the channel writes that it has no benchmarks of any kind. Set them aside.
“The ad product is sold by Microsoft.” It is not. The confusion probably dates from 2023, when Microsoft announced it was exploring ads in its own assistant. Microsoft appears in none of the partner lists OpenAI has published, and it is OpenAI’s advertising system that controls all delivery decisions. We have gone through what does and does not connect the two in the Microsoft Advertising file.
“Conversions here are one and a half to four times better than on Google.” Unverified, and contradicted by the publisher itself, which states it holds no performance benchmarks.
“Advertisers can see the conversations.” No. The documentation states that advertisers never receive your conversations, your chat history, your memories, your name, your email address, your precise location, your IP address or sensitive information. Only aggregated reports come back.
How we work on this
We arrive at this subject with almost always the same question: should we go now.
We start with eligibility. It is a ten minute check that settles the matter entirely for a large share of service businesses. Better done before the creative than during the review.
We look at whether the targeting available to you is enough for your offer. Where personalisation and customer files are unavailable, this channel suits what gets triggered inside a conversation and suits less what requires recognising a person. That is not a flaw, it is a scope.
We wire the consent refusal before the pixel. Since it is on by default, placing the tag and dealing with consent later is a breach, not a sequencing choice.
We set the decision threshold before the first impression. On a channel with no benchmarks, the only discipline that protects a budget is writing down in advance what would make you stop, and by when.
And we treat this channel as what it is. A young product, with no ad repository in most of the world, no published performance benchmark, and targeting rules its own publisher describes as liable to change. That can justify a framed test. It does not justify moving a budget.
Where to start
Open the advertising policy and find your sector. If you sell financial advice, legal services or healthcare, your answer depends entirely on which country you are selling into.
Check your firewall lets the ad validation crawler through. An ad whose landing page answers that crawler with an error will not be approved.
Wire your consent refusal before installing the pixel if you deliver into a consent jurisdiction. It is on by default and switching it off is your job.
Write your context hints as descriptions, not keywords. The documentation warns they are not exact match and guarantee no delivery.
And write your stopping threshold before the first impression. On a channel with no public benchmark, that is the only protection there is.