ChatGPT Ads creative: format, copywriting and the rejections to avoid

What an ad is made of

The ad is the third level of the ChatGPT Ads structure, below the campaign and the ad group. It carries four elements: a headline, a description, an image and a landing URL.

That is not much, and that is the point: there are no extensions, no assemblable headline variants, no documented sitelink block. Every word counts because there are no words in reserve.

Targeting is not set at this level: it comes from the ad group (context hints, bid) and the campaign (country, dates, audiences). The full map is on our ChatGPT Ads Manager page.

Where the ad appears, and to whom

Two documented facts that govern everything about the copy:

• Ads appear below ChatGPT's responses, and they stay separate from the model's answers. An ad is not woven into generated text and does not influence it.
• They are only served to users on the Free and Go plans. Plus, Pro, Business, Enterprise and Edu accounts see no ads, and OpenAI does not serve ads to accounts identified as belonging to users under 18.

The practical consequence for a B2B advertiser: your most solvent audience may be exactly the one that will never see your ad, because it is on Plus or managed under a Business plan. That is not a reason to skip the channel, it is a reason to calibrate expectations and to keep running ChatGPT SEO work in parallel, which reaches every plan.

The copy guidance OpenAI publishes, and the limits it does not

The ad creation documentation sets out explicit guidance, which we repeat as written:

• Keep titles concise, with clear, focused language, staying within the recommended character limits.
• Prioritise clear, useful details over broad marketing language.
• Add new information in the description rather than repeating the title's message.
• Keep the message easy to understand, using straightforward language that helps users quickly evaluate the offering.

The through-line is consistent: information beats promise. On a channel where the user has just received a structured, qualified answer, a slogan stands out, and not in a good way.

A simple test before you approve a headline plus description pair: strip out your brand name. Does the ad still say something concrete? If not, you wrote a slogan.

Let us be blunt, because this is where third-party content contradicts itself most. OpenAI asks you to stay within the recommended character limits, but we find no officially published numeric values for the headline and description. The sheets announcing 24, 30, 48 or 60 characters do not agree with each other and cite no primary source.

So we flag those figures as unconfirmed, and we suggest you do the one thing no article can do for you: read the limit shown in the Ads Manager composer as you write. That is the only authoritative value, and it can change without announcement.

What can be said safely: the format is short. Write the densest version you can, then check the counter. Doing it the other way round produces truncated ads.

Imagery: what the documentation asks for

On images, the official guidance is short and specific:

• Keep images simple and relevant, with clear visuals that support the ad message.
• Avoid overly abstract or cluttered visuals.
• Align the image with the ad message where possible.

These are editorial instructions, not technical specifications. Accepted dimensions, file formats and weights are stated by Ads Manager at upload time: that is the only current source, and the only one that will reject a non-conforming file on the spot.

Our own recommendation, consistent with the legibility guidance: an image rendered small carries neither burned-in text nor dense composition. One image, one subject.

The landing URL: the quietest failure point

The documentation asks you to link to the most relevant destination: a product page, a collection or a content page, rather than defaulting to the homepage. It also requires landing pages to be valid and reachable.

And above all it sets a technical requirement most advertisers discover after the fact: the page must not block OpenAI's user agents, in particular OAI-AdsBot and OAI-SearchBot.

This is the channel's nastiest failure, because it is invisible from inside Ads Manager and usually comes from a decision taken elsewhere in the company: a web application firewall rule, a robots.txt block inherited from an anti-AI policy, a CDN bot protection. Our ChatGPT Ads landing page checker tests those crawlers' access directly, alongside the content criteria.

The triangle: hint, ad, page

Three elements have to tell the same story, or you lose the user or get the ad rejected:

The official rule meets commercial logic here: an ad may be rejected if it is misleading or inconsistent. A promise nowhere to be found on the landing page ticks exactly that box.

Our method for writing hints is on the ChatGPT Ads context hints page.

Why an ad gets rejected

All ads go through review before they can serve. The documented grounds for rejection:

• the ad is misleading;
• it is inconsistent, between its own elements or with its destination;
• it contains disallowed content.

Compliance is assessed across the whole unit: copy, image and landing page must all comply with OpenAI's ad policies, updated on August 10, 2026 and built as an allowlist, where only explicitly permitted categories may run. The detail is on our ChatGPT Ads policies page.

If rejected, there are two routes: edit and resubmit, or request another review if you believe the rejection was in error. To open a useful support case, the documentation suggests gathering the ad title and copy, the disapproval code or message, the creative, the landing URL and the time you received the notice.

Write for a conversation, not for a query

The context shift changes how you write, and that is the real subject of this page.

On a search engine, the user compressed a need into a few words and is comparing a list. On ChatGPT, they have just received a structured, often qualified answer, sometimes with recommendations attached. Your ad arrives after that answer. It does not need to re-explain the topic: it needs to supply what the answer did not give, usually a specific offer, an availability, a price, a condition, a lead time.

Three concrete habits:

1. Start from the question asked, not the keyword. The reader wrote a sentence, not a phrase.
2. Give one verifiable fact in the description: a number, a lead time, a condition. That is exactly what the official guidance asks for.
3. Send traffic to the page that continues the answer, not to the brand showcase.

Several ads in one group: how to test

An ad group is a set of ads around one theme. Nothing forces you to write only one, and everything argues for writing several to compare angles.

A sober protocol, consistent with the documented markers:

1. Write two or three ads per group, differing on a single axis: the promise, the proof or the image.
2. Leave them alone, respecting the advice to wait at least 24 hours before drawing any conclusion about delivery.
3. Wait for enough click volume before cutting: a few dozen clicks separate nothing.
4. Read CTR and average CPC per ad in the reporting, then replace the weakest rather than adding a fourth variant.

The documentation does not describe a dedicated A/B testing feature with controlled traffic splitting: we flag that as unconfirmed rather than asserting it will never exist.

Ads generated from a product feed

For a retailer, product feed campaigns build ads from catalogue attributes instead of by hand. The consequence is direct: your feed titles and descriptions become your ad creative.

So the same guidance applies to the feed: concise, informative titles, descriptions that add information rather than repeat it, simple images consistent with the copy. A catalogue written for a price comparison site does not transfer as-is.

The full feed walkthrough is on our ChatGPT Ads for ecommerce page.

Checklist before you submit

Eight checks, ordered from cheapest to most expensive to fix later:

1. The headline fits the limit shown by Ads Manager, with no truncation.
2. The description adds information to the headline.
3. One concrete fact appears in the ad: a number, a lead time, a condition.
4. The image is legible at small size and consistent with the copy.
5. The URL points to the most relevant page, not the homepage.
6. The page is reachable and needs neither a login nor a prior acceptance to display the offer.
7. OAI-AdsBot and OAI-SearchBot are not blocked on that page.
8. The offer's category is permitted under the current ad policies.

What is not publicly documented

Points we flag as unconfirmed for lack of an official source:

• Numeric character limits for the headline and description are not published: rely on the Ads Manager composer.
• Technical image specifications (dimensions, ratio, formats, weight) are not reproduced here: Ads Manager states them at upload.
• The number of ads shown per conversation is not publicly documented.
• The existence of a dedicated A/B testing tool is not documented.
• Review turnaround times carry no public commitment.
• No guarantee of delivery, placement or outcome is offered by OpenAI, and nobody can offer one on its behalf.