Advertising in ChatGPT: what to prepare before opening an account

Since 31 August, a business in France can set up its own advertising campaigns in ChatGPT. What can delay a launch, or create a privacy compliance problem, sits in OpenAI's ad policies, on your website and in your consent banner.

OpenAI has been serving ads in 31 European markets since 24 August, to adult users on the Free and Go plans, and its Ads Manager has been open for self-serve (in beta) since 31 August. Access, the budget for a first test and how to measure the channel's effect are covered in our ChatGPT Ads guide. What remains are the checks to run before you create the account.

Is your sector even allowed?

This is the fastest check, and it can end the project. OpenAI's ad policies prohibit several categories (gambling, alcohol, dating, adult content, political advertising, legal services other than general legal information) and restrict others, such as health, where advertisers are admitted case by case after manual review. Financial services are a separate case: outside the United States, they are generally prohibited. An online bank or a broker in Europe has nothing to prepare for now.

There is also a placement rule, distinct from advertiser eligibility. OpenAI announced in January that ads do not appear near conversations about health, mental health or politics; the rule now targets the user's personal health. A food supplement brand may therefore have its ads approved, then excluded as soon as the user talks about their own health. These policies have been rewritten several times since the spring: read the version in force on the day you open the account.

The OpenAI crawler you must not block

Before serving an ad, OpenAI sends a crawler, OAI-AdsBot, to the landing page. Its documentation gives it two jobs: checking that the page complies with the ad policies, and using its content to decide when to show the ad. OpenAI states that what it collects is not used to train its models.

Many sites have blocked OpenAI's crawlers to opt out of training. A robots.txt rule aimed at GPTBot has no effect on OAI-AdsBot, which has its own user agent; a blanket disallow (User-agent: *) does block it. Bot protection at the CDN or firewall level (Cloudflare, Akamai or similar) can also stop it before it reaches the page, even though OpenAI says Cloudflare recognises it as a verified bot. OpenAI publishes the crawler's IP addresses, so it can be allowed explicitly. According to OpenAI's advertiser help centre, an ad whose page could not be checked is unlikely to be approved. Have these settings checked before opening the account. If you want to keep refusing training, nothing prevents you from letting OAI-AdsBot through on its own, on landing pages only.

Because the page content also decides when the ad is shown, a vague page (a slogan, a video, a form) gives the system little to work with. It should state plainly what you sell and to whom.

The pixel and your consent banner

For measurement, OpenAI provides a JavaScript pixel and a Conversions API. One line of the pixel documentation matters directly for European sites: it "initializes consent to true by default unless you set it to false". Installed as is, it sets its cookies for visitors who refused advertising, whereas the ePrivacy Directive, enforced in France by the CNIL, requires prior consent for advertising trackers.

Before the first impression, declare the pixel as a vendor in your consent management platform, call oaiq("consent", false) before it initialises, then update the value according to the visitor's choice. If the site uses a Content Security Policy, add the OpenAI domains listed in the documentation, or the script will not load. And if you also send conversions through the API, reuse the same event ID on both sides so they are not counted twice.

Targeting without keywords or audiences

Ads Manager has no keywords. Each ad group takes context hints: phrases describing the conversations, needs or situations where the ad is relevant. According to OpenAI, these hints guide selection without working as exact match.

In Europe, audiences cannot take over. OpenAI says ad personalisation is not active in the European Economic Area or Switzerland: selection relies on the current conversation, approximate location, language and time, with no chat history and no audiences. Retargeting and customer lists have no equivalent.

Your context hints and your landing page therefore carry all the targeting. They should be written from the questions your customers actually ask, and a Google Ads keyword list is a poor starting point. Take a payroll software vendor selling to small businesses. The keyword "payroll software smb" describes no conversation. A useful hint looks more like this: a small business owner who wants to bring payroll in-house and is comparing software before talking to their accountant. The raw material sits with your sales team and in customer support tickets. Also prepare the list of contexts where you do not want to appear: in early October, OpenAI opened negative phrases to some advertisers.

On the creative side, the ad comes down to an advertiser name, a favicon, a title, a short description, an image and a URL. OpenAI recommends very short copy and advises against using the logo as the main visual: plan an image of the product or of it in use.

The answer shown above your ad

In January, OpenAI set out a principle of answer independence: ads do not influence ChatGPT's answers, and they appear below the answer, separated and labelled. Your ad therefore sits under text you do not control, which may recommend a competitor, quote an outdated price or describe your offer wrongly.

Before launching, ask ChatGPT the questions your context hints describe, in your customers' language and several times (answers vary between sessions), and note what it says about you and your competitors. If the answer recommends a competitor by name, your ad will appear right below it, and its message has to account for that. This work belongs to generative engine optimisation (GEO) and is worth doing even if you never run a campaign.

In what order to proceed

We would run these checks in order of the cost of getting them wrong, starting with the one that can make the project impossible:

• Eligibility: the ad policies in force, for your category and each of your products.
• Website: OAI-AdsBot allowed in robots.txt and the firewall, and pages that say what you sell and to whom.
• Consent: the pixel declared in the consent platform and set to false before initialisation, the CSP updated.
• Context: hints written from real customer questions, and the list of contexts to exclude.
• Answers: what ChatGPT says about you and your competitors on those same questions.

Once these are settled, the test framework can be set using our guide. If you would rather hand over the account setup and management, that is what our ChatGPT Ads service covers.