A ChatGPT ad is small on purpose: a headline capped at 50 characters, a description capped at 100, plus an image, all packaged into what OpenAI calls a "chat card." There's no room for the persuasive padding a lot of ad copy leans on, and the writing style that wins a Google auction often reads oddly inside a conversation.

Here's the detail most guides skip entirely: 50 and 100 characters are the input limits, not what actually displays. The headline visually truncates around 24 characters and the description around 48, roughly half the space you're allowed to type. Write past that point and the words doing the actual persuading may never get seen by most readers. Everything below accounts for that.

  1. Front-load your first ~24 characters of headline and ~48 of description. The card visually truncates well before the input limit, so the specific claim has to land in that first window. Treat anything past it as a bonus for readers who happen to see the expanded card, not the actual pitch.
  2. Write like a specific, honest recommendation, not an ad. Copy that reads like something a knowledgeable person would actually say performs better than copy that reads like marketing. The context is a conversation, and the ad has to fit it.
  3. Lead with a specific, verifiable claim. A number or fact a reader can check beats a vague benefit every time. "Ships in two days" outperforms "fast, reliable shipping," and it fits inside the truncation window besides.
  4. Don't reuse Google or Meta copy as-is. Copy optimized for a search auction or a social feed is written for a different context and a different reading pattern, and neither was written against a 24-character visible headline. Write fresh for the conversation format.
  5. Match the message to where the conversation is. A first-time question deserves copy that hooks a real problem. A comparison-stage conversation deserves copy that differentiates clearly. A decision-stage conversation deserves copy that removes risk.
  6. Meet the actual image spec, not just "add a picture." The format requires a square, 1:1 image at a minimum of 256×256px, ideally 512×512px or higher for retina displays, in JPG, PNG or WEBP. A stretched, low-resolution or generic stock image undermines relevance the same way weak copy does, since the image is one of the four inputs the auction actually scores.
  7. Write context hints around the buyer's situation, not your product. Context hints are the plain-language descriptions that tell the auction when your ad is relevant, and they're scored as part of the same relevance calculation as your copy. "Someone comparing project management tools for a small team" works better as a hint than a list of your product's features.
  8. Test more variety than you would on a mature platform. Because the format and audience behavior are still new, assumptions about what performs carry more risk here than on a channel with years of benchmark data.
  9. Keep claims something the model would confidently repeat. Vague or exaggerated copy doesn't just underperform with people. It's also less likely to be the kind of content a model treats as trustworthy elsewhere.
The ads that work here don't sound like ads. They sound like someone who actually knows the answer chiming in, and they say the important part in the first 24 characters.

Once your creative is dialed in, see what that's likely to cost in our ChatGPT ads pricing guide.