1. A new picture versus the item on the shelf
ChatGPT, Gemini and Midjourney are examples of general-purpose image generators. They are built to create a new picture from a request, and a good result is one that looks convincing and fits the request.
A product photo is judged by a different test: does it show the item that ships? The buyer compares the picture with the parcel. A print that has moved, a logo with a changed letter or a button the real garment does not have turns into a return or a complaint, however good the picture looks.
2. What must stay the same
For a fashion item, five things carry its identity: the print (motif, scale, colours and placement), the logo and any text, the buttons, zips and pockets, the colour, and the cut — length, neckline, sleeves and fit. A buyer who orders from the picture expects each of them.
A prompt describes a product in words; a photo of the item shows it. Whichever tool produces the image, the result is a new image, and only a side-by-side comparison with the real item shows whether those five details came through.
3. Google Merchant Center: match the product you sell
Google’s image requirements ask sellers to “match the image you submit to the product that you're selling, and ensure your image shows the correct color, pattern, and material”, and to show the correct variant.
For AI-generated images, Google requires metadata that marks them as AI-generated, using the IPTC DigitalSourceType tag TrainedAlgorithmicMedia, and asks sellers not to remove it. Such images are accepted in image_link, additional_image_link and lifestyle_image_link.
4. Amazon and Zalando: people and distortions
Amazon asks sellers to add the metadata keyword contains-synthetic-performer, with an IPTC-compatible image editor, to images that show photorealistic people created entirely by AI. Images of real people are excluded, even when they were edited with AI.
Zalando’s AI image content policy states that synthetic content must meet its editorial and technical criteria, and that assets which fail to represent the product accurately can be rejected. Its examples of unacceptable distortions include distorted text, mispositioned buttons, artifacts in logos and details that show inconsistently across an image set. Visible on-image labels are not allowed; identifying data belongs in the metadata or an invisible watermark.
5. Where a general-purpose generator fits
Images in which no stocked item has to be reproduced are a different case: a mood board, a campaign concept, a background texture or a social post around a theme. There, a convincing new picture is the goal.
Once an image goes into a listing, the tool that made it matters less than the check it passes. The list below applies to every AI product image, from any source.
Checklist
- Lay the real item next to the screen and compare in daylight, colour first.
- Print: the same motif, scale and placement; a back print stays on the back.
- Logo and text: every letter as on the item, in the same position.
- Buttons, zips and pockets: the same number, in the same place.
- Cut: length, neckline, sleeves and fit as on the item, with nothing added that the item does not have.
- Metadata: the AI marker still in the file after resizing; for Amazon, contains-synthetic-performer when the image shows an AI-generated person.
Sources
- Google Merchant Center Help — Image link [image_link]
- Google Merchant Center Help — AI-generated images
- eWeek — Amazon requires sellers to disclose AI-generated people
- Zalando Partner University — Image guidelines
Checked 26 September 2026. Rules change; the official pages are the final reference.