why does the openai api write worse product descriptions than chatgpt itself?

Look at your temperature and max tokens settings too. A low max tokens cuts descriptions off or makes them oddly clipped. And a lot of people set temperature to 0 thinking it means accurate - for copy it mostly means the same safe sentence structure on every product. We got better results somewhere in the middle, then tightened it if it started inventing features.

Usually because the chat app isn't a bare model. It has its own long system prompt and defaults behind it, and the API gives you none of that, so you're comparing a tuned setup with a raw one. Also double check you're calling the same model, not an older or smaller one. What fixed it for us: write a proper system message (brand voice, length, what to avoid), paste two or three of your best existing descriptions as examples, and send the product attributes as structured fields instead of one blob of text. Output got a lot more consistent after that.