In short
- Ads opened inside ChatGPT on 24 August 2026 across thirty-one European markets, France included. They appear only for Free and Go accounts, never for paid subscriptions.
- In Europe the targeting is contextual: the topic of the live conversation, approximate location, device. No chat history, no memory. Which means no demographic targeting, no lookalikes, no site retargeting.
- Sign-up is open, serving is not. The account is created self-serve, then goes through a business verification and a manual review, with no announced delay and no way to speed it up.
- No sector-level profitability benchmark exists. The platform publishes none, and the conversion-rate-by-industry tables now circulating declare no sample, no period and no method.
- What you can prepare now, without spending: the crawlers allowed in your robots file, the measurement script allowed by your security policy, consent wired before the pixel initialises, your conversion event chosen, and URLs that survive rewriting.
Run the test before reading on. Open ChatGPT on a free account and ask it to compare two suppliers, two models, two offers, the way one of your customers would. Since Monday 24 August, below the answer, you will see a box marked as sponsored, visually detached from the generated text. Advertising has just opened in thirty-one European markets, France among them.
The question everyone is asking is what it costs and whether it works. That is the wrong question, and not because it is premature. Because nobody can answer it for your sector today, and because the real window for action is somewhere else.
Three things to know before anything else
1. The audience is not “all of ChatGPT”. Ads show only on Free and Go accounts. Plus, Pro, Business, Enterprise and Edu display none. So you are not reaching the heaviest users, who are often the most professional ones. Temporary chats are excluded too.
2. In Europe the targeting is constrained, and that shapes everything. At launch, selection rests on the topic of the live conversation, an approximate location, the device type, the hour and the language. Chat history and memory are not used. Personalisation will come later, on explicit and revocable consent. For a media buyer that means no demographic targeting, no lookalike audiences, no retargeting on site visits. You do not pick keywords, you describe conversation themes at ad group level.
3. Sign-up is open, serving is not. This is the part last week’s coverage missed. The advertiser account is created self-serve, but the path includes a business verification by a third party, then a manual review. The platform states plainly that request volume is high, that the queue runs continuously, and that no request can be expedited. Three settings are final at account creation: country, currency and time zone. Getting one wrong means starting a new account, with a new email address.
The trap is believing the subject is buying
The natural reflex with a new channel is to open the account, put a small budget behind it and watch. That is exactly what failed in the United States in February.
The American launch was a measurement problem before it was a performance problem. The first advertisers bought with no usable reporting: no conversion tracking, no automated buying, deals done over spreadsheets and email.
The second layer of the trap is the profitability benchmarks already circulating. This week you will come across conversion-rate-by-industry tables, very tidy, very detailed, telling you that one sector converts at five per cent and another at less than one. Look at their methodology: there is none. No sample size, no period, no account count. One of the most reproduced admits mixing internal client data, an organic conversion report and old paid-search research. The reality is simpler to state: the platform has published no sector benchmark at all, no cost per click, no cost per thousand, no click-through rate, no cost per acquisition. The only traceable data available comes from a single account over fifteen days, and its author points out these are attributed conversions rather than a demonstrated incremental gain.
It is the same mechanism we described in our piece on citability by generative engines: a figure everyone quotes and nobody produces.
What worked, what did not, and who says so
Six months of hindsight allows some sorting. The table below gathers what is public and attributable, separating what the platform claims from what advertisers observed.
| Case | What happened | Where it comes from |
|---|---|---|
| Newegg, Best Buy, Lowe’s, VistaPrint | The product’s official showcase. The testimonials speak of presence at the moment of decision, never of return on investment. No figure accompanies them | The platform’s advertiser page |
| One advertiser in the American pilot | Spent roughly one hundredth of the budget it had committed, over four weeks, for two hundred impressions. The pilot then covered less than ten per cent of the eligible audience | Digiday, May 2026 |
| Two agency executives | Unable to prove any commercial result to their clients. Several large brands declined to return without a trusted intermediary | The Information, via Search Engine Land, March 2026 |
| An advertiser tracking conversions through URL redirects | Efficiency close to its own non-brand search. The only one in the set the press notes had set up independent measurement before spending | Digiday, May 2026 |
| One account tracked over fifteen days | Return on ad spend of 1.49 and a conversion rate of 2.35 per cent. The author states these are attributed conversions, with no incrementality measurement | Opascope, data from its own account |
| The platform itself | Cost per acquisition down sixty per cent in six weeks, cost per thousand impressions down twenty per cent, fill rate up thirty to fifty per cent in the spring | OpenAI communications, unaudited |
Read the first and fourth rows together, that is where the lesson sits. The official showcase lines up four large American retailers and not one claims a measured gain: the performance marketing lead at one of them explains he wants people researching a purchase to think of his brand. That is an awareness objective, stated by a performance lead. Meanwhile, the only advertiser reporting efficiency comparable to its reference channel is also the only one known to have built its own tracking.
So the dividing line is neither the sector nor the budget. It runs between those who could observe what was happening and those staring at an empty dashboard. The channel has genuinely improved since: prices came down, cost-per-click then cost-per-action billing arrived, fill rate recovered. What has not changed is that an advertiser without measurement will not be able to benefit from it.
What about apparel?
It is one of the most present categories, and that is verifiable. Between 10 February and 24 May 2026, retail brands accounted for close to forty per cent of the advertising impressions recorded inside ChatGPT, with clothing among the leading subcategories. For the fortnight of 12 to 26 February alone, retail and grocery represented forty-four per cent of impressions in ChatGPT against thirty-seven per cent on Google search: the category is over-represented relative to its usual weight. These readings come from a panel of volunteer users on the American Android app, which identified 1,419 distinct brands over the period.
| Signal | What it indicates | What it does not |
|---|---|---|
| Two clothing retailers, L.L. Bean and Quince, among the ten largest advertisers in July | Five months after opening, they are still buying. Sustained presence is the only satisfaction signal available today | Neither publishes a return on ad spend, a cost per acquisition or a sales volume |
| Gap recorded in the second half of February | Fashion retailers tested very early, alongside Best Buy, Target, Sephora, Ulta and Wayfair | Presence over a fortnight does not say whether the campaign was renewed |
| Advertiser count rising from about 300 in April to more than 820 in July | The channel is filling up, and nearly one July advertiser in five was new | 1,419 brands had already tested between February and May: many tried without staying |
One qualification is needed, and it comes from the source tracking the market continuously: the share of retail brands in ChatGPT advertising spend fell from about thirty-six per cent in April to twenty-five per cent in July, while financial services rose from two to twelve per cent. Retail remains the leading category, but its lead is narrowing as other sectors arrive. A high share of voice today does not guarantee it holds.
Failure, in this category, is not documented brand by brand. It is documented through price. At launch, cost per thousand impressions sat at roughly three times the average observed on Meta. One analyst sums the problem up in a sentence: high rates can be justified, provided you help brands measure them. That is precisely what was missing. Add the reception context, which apparel cannot ignore: sixty-three per cent of American adults say advertising inside an AI-assisted search reduces their trust in the results.
So run the numbers with yours rather than with a table found online. Publicly reported costs per click range from roughly 1.70 to 4.90 dollars depending on the account and the bidding model. Take your real conversion rate on a collection page, your average order value and your return rate, which weighs heavily in fashion. Compare the resulting cost per acquisition with your non-brand search. If the gap does not hold, the answer is no, whatever the surrounding enthusiasm.
One point of honesty to close: as of today, no apparel case study publishing a return on ad spend is publicly available. What circulates under that name belongs to the method-free tables described above.
First come, first served: what it is actually worth
On Monday, France went overnight from zero advertising inventory to the full inventory, while the number of French advertisers starts at zero. That configuration will not happen twice.
The argument rests on auction mechanics, and the American figures make it legible. At the February launch, cost per thousand impressions was around sixty dollars, with a high minimum commitment. When inventory opened more widely, that cost fell to about twenty-five dollars and the minimum disappeared. Then demand caught up with supply: advertiser count nearly tripled between April and July, from about three hundred to more than eight hundred and twenty. In Europe the order is reversed. Inventory across thirty-one markets arrives at once, local demand will take months to build, and an imbalance between supply and demand always ends up showing in the price paid.
Three qualifications, otherwise the promise would be dishonest.
The scarce thing right now is not inventory, it is your place in the queue. The bottleneck is the business verification and account review, which the platform itself says takes time and cannot be expedited. Signing up early costs nothing and commits you to no spend: it gets you a position. Those signing up in November will meet the same queue, longer.
What you learn early cannot be bought back later. Contextual signals resemble neither keywords, nor audiences, nor demographic targeting. You describe kinds of conversations. Nobody has more than six months of practice with this mechanism, and nobody has any in French. The real lead is not on cost per click, which will normalise, but on understanding what causes an ad to surface inside a conversation.
The advantage does not go to the first to arrive, it goes to the first who can measure. That is the whole lesson of February: those who arrived first without a measurement setup did not gain a lead, they funded everyone else’s learning. Arriving early without being able to read what happens means paying the pioneer’s price without collecting the benefit.
Hence the only sequence worth following today: take your place now, since it is free, and use the waiting period to build measurement. What follows is exactly that list.
Five checks to run this week, with no tools and no spend
While your account sits in the queue, here is the self-assessment. If you cannot answer all five, you are not ready to spend a euro on this channel.
1. Do your robots let the platform’s crawlers through? Two agents must reach your pages: the one validating ads and assessing their relevance, and the one exploring content. If they are blocked by your robots file or your application firewall, your ads will be invalidated without an obvious reason. Open your robots.txt and look for blanket rules blocking anything resembling an AI crawler: many sites added them in 2024 and never revisited them.
2. Does your content security policy allow the measurement script? Three directives are involved, not one, and this is the most common mistake: script-src to load the script, connect-src to fetch the pixel configuration and transmit conversions, img-src for the image-request fallback. One directive missed and your dashboard will report zero conversions on campaigns that produce them. If your counter reads zero rather than merely low, look here before anything else.
3. Is consent actually wired? This is the most sensitive point, and the quietest. The pixel starts in active collection by default. It must be explicitly held before initialisation, then released when the person accepts. Pasting the script without that guard means collecting before consent, in the region that has just been given reduced targeting for privacy reasons.
4. Do you know which event you will optimise for? A conversion-optimised campaign takes a single standard event, and that choice is locked at creation. Purchase, appointment, account creation, quote request: it has to be settled beforehand, not after. And the event must already exist, cleanly, on your site.
5. Do your URLs survive redirection? Tracking relies on a parameter appended to the destination address, then stored in a first-party cookie. Any aggressive URL rewriting, any parameter stripping by a caching or security module, breaks the chain. It is the same blind spot we described about server-side tracking.
Four possible stances, and who each suits
| Stance | What it gives you | Who for |
|---|---|---|
| Wait six months | You let the market take the hits and you get real cost benchmarks | Regulated sectors, low-margin catalogues, teams with no technical resource available |
| Sign up now, do not spend | You take your place in the verification queue and use the delay to wire up measurement | Most online retailers and service businesses that own their site |
| Test small as soon as access lands | You learn the mechanics of contextual signals, which resemble no other channel | Those who already have reliable measurement and a short purchase cycle |
| Go through an agency or a partner | You get support and third-party tooling, at a higher entry ticket | Significant media budgets, multi-country brands |
These stances combine. Signing up without spending, then testing small, is in fact the most sensible sequence for a smaller company.
One thing to check before all the rest: your sector may simply be excluded. Dating, alcohol, tobacco, gambling, politics and weapons are prohibited. Health, health insurance, consumer credit, financial advice and investment services are reviewed case by case, currently in the United States only. For those sectors the profitability question does not arise yet: the inventory is closed to them.
The tooling
To make the self-assessment above concrete, we publish a free measurement kit in our resources: a generator that produces the measurement script with the consent guard already in place for the main consent platforms, the naming convention for campaign links, a checker where you paste your robots file and your security header, and the list to run through before the first campaign. Like our other resources, it is dated and maintained: the product is in beta and its settings will move.
On the server side, sending conversions without depending on the browser is a matter of server-side tracking infrastructure. For our clients we use the platform published by Datafirefly Limited, our agency’s sister company, which opened a ChatGPT Ads destination on the day of the European launch. It works no miracles and does not measure incrementality: it only guarantees that the conversions you actually earned reach the tool deciding your bids. It is a foundation, not an answer. Method beats tooling, and the requirements above hold whichever solution you choose.
What this says more broadly
Search is fragmenting. Ten years ago a smaller company could reasonably concentrate its acquisition on a single engine. Today part of the comparison phase happens inside conversations, part in marketplaces, part in generated answers that link nowhere. Each new space arrives with its own identifiers, its own attribution windows, its own blind spots.
In that landscape, the only thing that transfers from one channel to the next is your own ability to measure what happens on your side. Platforms change, targeting rules change, prices change. A clean conversion event, correctly wired consent and a stable naming convention remain valid everywhere. That is the only asset you build for yourself rather than for whichever ad network is current.
Sources
The figures in this article link back to the document that produced them. We do not cite sources we have not read.
- OpenAI, advertiser page for ChatGPT Ads, consulted on 24 August 2026. Sign-up path, testimonials from Newegg, Best Buy, Lowe’s and VistaPrint. Open
- OpenAI, Ads Manager Beta account setup, help centre, updated August 2026. Business verification, review queue, country and currency locked. Open
- OpenAI, Quick start: launch your first campaign, help centre. Campaign, ad group and ad structure, native metrics, CSV export. Open
- OpenAI, Measurement Pixel, developer documentation. Script, standard events, consent handling, first-party cookie. Open
- OpenAI, Conversions API, developer documentation. Server-side sending and deduplication by event identifier. Open
- Search Engine Land, March 2026, reporting an investigation by The Information. Absence of reporting and of proof of results for early advertisers. Open
- Digiday, May 2026, seven advertising executives interviewed. Pilot under-delivery, fill rate recovery, efficiency close to non-brand search for one equipped advertiser. Open
- Digiday, August 2026. European opening, thirty-one markets, partner agency groups, evolution of bidding models. Open
- Opascope, benchmarks from fifteen days of real spend on a single account, first-party tracking. Return on ad spend, cost per click and conversion rate. Open
- Sensor Tower, April to July 2026, United States. Share of advertising spend by category and leading advertisers. Open
- eMarketer, July 2026. Top ten advertisers including L.L. Bean and Quince, advertiser count rising from about three hundred in April to more than eight hundred and twenty in July. Open
- eMarketer, retailer economics. Cost per thousand impressions compared with Meta, retail share in ChatGPT and on Google search from 12 to 26 February 2026, and an Ipsos survey from January 2026 on trust. Open
- AI Ad Economy, analysis of Sensor Tower readings from 10 February to 24 May 2026. 1,419 distinct brands and the breakdown of impressions by category. Open
FAQ
Is there a minimum budget to start?
The minimum commitment that existed at the American launch was removed in the spring. There is no announced floor today. Watch one operational detail though: a campaign budget is a total spending cap, not a pacing control. A poorly bounded campaign can burn through it quickly.
How long between sign-up and the first ad serving?
No delay is announced. The path includes a business verification, an account review, then a review of each campaign. The platform states the queue runs continuously and that no request can be expedited. So plan to build measurement without knowing when access will land.
Can competitors appear in an answer that mentions my brand?
Selection rests on the conversation theme, not on keywords bought in the classic sense. An advertiser describes kinds of conversations relevant to its ad group. There is no public mechanism today for bidding on a third-party brand, nor brand protection comparable to search engines.
Can Dotsland help?
Yes, on the part that actually matters today: measurement. Start with the five-point self-assessment in this article, it needs no tools and will tell you where you stand. If the answers are missing, or if conversion tracking is already fragile on your current channels, let’s talk before you add another one.