AI product images rejected: the three real causes

Start by sorting the rejection into one of three buckets

Rejection messages are vague on purpose, so the useful first move is triage rather than troubleshooting. In our experience almost every rejected product image falls into one of three buckets, and each one needs a different fix.

• •The image does not match the product the customer will receive.
• •The provenance metadata is missing, or your pipeline stripped it out.
• •The image breaks an ordinary image rule that has nothing to do with AI.

Work out which bucket you are in before you change anything. Fixing the wrong one is how teams end up regenerating a whole catalogue and getting rejected all over again.

Bucket 1: the image does not match the product

This is the big one. A seam that is not there, a handle that runs differently in real life, a logo in the wrong spot, a bath that sits too small in the bathroom. Marketplaces do not call this an AI problem. They call it accuracy, and the rules for it predate generative models by years. What changed is volume: you can now produce a hundred images in an afternoon instead of ten, so every weakness in your review process shows up a hundred times faster.

In our own quality checks, this is what the conversation is almost always about. Not "does this look AI" but "does that hinge actually sit like that". We went into the detail in a separate piece on AI product photo accuracy and invented details.

Look at the image above. The logo, the bolts, the hinge, the vent wheel: those are exactly the places a model will improvise if you give it room. Hence the rule of thumb that works best for us. Let AI build the environment, not the product. The product comes from your own packshot or your own reference material. Anything the model is allowed to fill in, it is also allowed to invent.

Bucket 2: the provenance metadata is missing

Modern generators write their origin into the file. Usually that is the IPTC field digitalSourceType with a value such as trainedAlgorithmicMedia, often wrapped in Content Credentials under the C2PA standard. Platforms read it.

Here is the part that surprises a lot of brand teams: that metadata is on your side. Google Merchant Center asks for it explicitly. Images created with generative AI must contain metadata indicating that, and Google states plainly that you should not remove embedded IPTC tags. Missing is worse than present.

What usually goes wrong is not intent, it is plumbing. A WebP conversion, a resize script, an upload through an older integration, and the metadata is gone. Now a detector flags the file as generated while the file itself claims no origin at all. So check what comes out of the end of your pipeline, not just what comes out of your generator.

Bucket 3: an ordinary image rule, with no AI involved

Sometimes it genuinely has nothing to do with AI. Promotional text burned into the image, a watermark, a discount flag, a logo added in post, resolution below the minimum, or a lifestyle shot sitting where the main image should be. These rejections existed long before the first generator. They simply surface more often once volume goes up. We mapped the hard pixel requirements per channel in product image resolution per channel.

What each platform actually says

Amazon

Amazon does not ban AI imagery. Its image rules are about what you show: the main image shows the real product on a pure white background, the product fills at least 85 percent of the frame, the longest side is at least 500 pixels, and from 1600 pixels the zoom function kicks in.

One thing was added on 22 July 2026. If a listing image, A+ content module or product video contains a photorealistic person generated entirely by AI, you have to add the keyword contains-synthetic-performer to the dc:subject (XMP) field before you upload. It applies across Amazon's worldwide stores, and you add it yourself with an IPTC-compatible metadata editor; there is no checkbox in Seller Central. It does not apply to real people, even if you edited them with AI tools, to people who are clearly not photorealistic, or to images with no people in them.

Google Shopping and Merchant Center

Google accepts AI imagery in your feed on one condition: the file carries metadata saying it was made with generative AI. Google lists three valid IPTC values, TrainedAlgorithmicMedia, CompositeSynthetic and AlgorithmicMedia, and tells you not to remove embedded IPTC tags.

The ordinary feed rules still apply on top of that. No overlays, watermarks, prices or text such as "Buy", "Shop now" or "Coming soon" in the image, and the image has to match the exact variant you are selling. From 31 January 2027 the minimum size is 500 by 500 pixels, with 1500 by 1500 or larger recommended.

Meta: Facebook and Instagram

Meta mostly labels rather than blocks. When an ad is created or significantly edited with Meta's own generative ad features, an "AI info" note appears under "About this ad". Minor changes such as cropping or colour correction do not trigger it. Only advertisers running ads about social issues, elections or politics are required to disclose themselves that an image, video or audio was digitally created or altered.

Practically, that means if your product ad gets rejected, look at the message rather than the method. A misleading representation of the product, a before and after that promises too much, a claim the image cannot support. And a label is not a rejection, nor a penalty.

Regional marketplaces

Smaller marketplaces tend to say nothing about AI at all, which means you fall back on their normal image rules, and those are often stricter than Amazon's. In the Benelux, bol.com asks for a white neutral background with no visible shadow, the item out of its packaging, the front view, no lifestyle shot, and no models, text, logos, icons or watermarks on the main image, at between 500 by 500 and 6000 by 6000 pixels. Since 1 July 2025 bol no longer actively enforces those image guidelines apart from self-promotion, but not enforcing is not the same as dropping them: bol can still swap or adjust your main image, and visibility in search still depends on it. Worth checking the equivalent page for every marketplace you sell on. We wrote about what the strongest listings do differently in standing out on marketplaces.

If you sell into the EU: what the AI Act does and does not ask of you

The transparency obligations in Article 50 of the EU AI Act have applied since 2 August 2026. Plenty of brands read that as an obligation to label every AI image they publish. That is not how it is built.

The marking obligation sits with the provider of the AI system, which has to mark output in a machine-readable way so it can be detected as AI-generated. As a brand you are a deployer, and the deployer's own disclosure duty covers deepfakes and text on matters of public interest. A lifestyle shot of your own bath in a bathroom scene is not a deepfake.

That does not make it irrelevant to you. You buy from exactly those providers, so the marking arrives inside your files through your tooling. Leave it there. And ordinary consumer law still applies: an image may not mislead the buyer, AI or not. We covered rights and policy in more depth in using AI-generated images commercially.

What not to do

There is advice circulating on how to get past detection: paste EXIF data from a real camera into the file, add digital noise, strip the Content Credentials. We would advise against it, and not only because it goes against the rules of the channel you are selling on.

It solves a different problem from the one you have. If you were rejected because the image does not match the product, noise changes nothing; the next audit, the first return or the first review will surface it anyway. And if you were rejected because provenance data was missing, stripping makes it worse. Google writes it out in plain language: do not remove the embedded IPTC tags.

Our position is simple. An image that is accurate does not need camouflage. An image that is not accurate does not get better with it.

Checklist for images that clear every channel

• •Put every image next to a real photo or the real product and check shape, proportion, colour, logo, text and finish.
• •Let AI build the environment, not the product itself.
• •Use your own packshot for the marketplace main image and keep lifestyle work for the supporting slots.
• •Check at the end of your export chain that the IPTC and C2PA metadata is still in the file.
• •If a photorealistic, fully AI-generated person appears in the image, add contains-synthetic-performer to dc:subject before uploading to Amazon.
• •Keep overlays, prices, discount flags, logos and watermarks out of the frame.
• •Check the hard requirements per channel: background, how much of the frame the product fills, and pixel dimensions.
• •Decide who does the final check, and do it before publishing instead of after the rejection.

How we handle this

At LensLab a creator from our own team makes the visuals in Visual Studio Pro, and a person reviews everything before it reaches you. That review is not a repair step, it is a quality choice: precisely because AI makes it possible to fill a whole catalogue in days, someone has to know the product that is in the frame. We deliver more options than you need and you only pay for the visuals you accept.

If you sell through dealers and retailers, all of this counts twice. An image that gets rejected on your own listing is, within a week, live on dozens of sites you do not control. Correcting it there costs far more than checking it once up front. You can see examples of what we made for brands on our use cases, how the process runs on see how it works, and what it costs on our pricing page.

Sources

Amazon Seller Central, announcement on tagging AI-generated people in listings and A+ content (22 July 2026): sellercentral.amazon.com

Google Merchant Center Help, Image link [image_link]: support.google.com

Meta, How AI-generated images in ads are identified and labeled: meta.com

bol.com partner platform, rules for your images: partnerplatform.bol.com

European Commission, Guidelines on transparency obligations for providers and deployers of certain AI systems (Article 50 AI Act, applicable from 2 August 2026): digital-strategy.ec.europa.eu