AI-generated image metadata: Google wants it kept, and your editor probably removes it

The short version

• Google requires AI-generated images in a product feed to carry a provenance tag, and tells you not to remove it.
• The generator does add it. We generated an image and found the exact tag Google names, plus a credit reading “Made with Google AI”.
• Opening the file and saving it again — with no edit at all — destroyed both. Not resizing. Not cropping. Opening and saving.
• Different tools behave differently, and one kept the wrong half: a PNG resize in ImageMagick preserved the “Made with Google AI” credit while losing the tag Google actually names. The file still looks provenanced. It is not.
• There is a fix, and it is opt-in rather than automatic.

What Google asks for

The rule sits in Google’s Merchant Center image requirements, which govern the pictures in your product feed rather than the ones on your own website. In Google’s words:

“All images created using generative AI must contain metadata indicating that the image was AI generated (for example, the IPTC Digital Source Type trainedAlgorithmicMedia metadata tag).”

And immediately after: “Don’t remove embedded metadata tags such as the IPTC Digital Source Type property from images created using generative AI tools.”

Google names three codes and says they “should be preserved”, defining each one itself: “TrainedAlgorithmicMedia: The image was created using a model derived from sampled content.” “CompositeSynthetic: The image is a composite that includes synthetic elements.” “AlgorithmicMedia: The image was created purely by an algorithm not based on any sampled training data (for example, an image created by software using a mathematical formula).”

The first is the one an ordinary image generator sets. Note the verb in that instruction, because the whole of this article is about it: preserved. Google is not asking you to add anything. It is asking you not to lose something.

A footnote on that sentence, which moved while we were watching

We quoted this same passage in a post on 7 September, having checked it word for word. Re-fetching it on 10 September to write this piece, the checker said our quotation no longer matched — so we went looking, rather than assuming either side was wrong.

An archived copy of the page from 8 June 2026 reads “meta data”, “AI-generated” and “DigitalSourceType TrainedAlgorithmicMedia”. The live page today reads “metadata”, “AI generated” and “Digital Source Type trainedAlgorithmicMedia”. Google copy-edited the sentence some time between 7 and 10 September. The requirement did not change; the words did. We have updated our earlier post to match the live page and said so there.

It is a small thing, but it is the reason we date every check. A quotation is a claim about what a page said on a particular day.

The test

We generate an illustration for every article on this blog, which gives us a steady supply of genuinely AI-generated images. The picture at the top of this page is the one we tested.

First we looked inside the file as downloaded. It carried two separate things: an XMP packet containing Iptc4xmpExt:DigitalSourceType set to trainedAlgorithmicMedia — exactly what Google names — and a separate IPTC profile crediting “Made with Google AI”. So the generator does its part properly.

Then we did eight ordinary things to it and looked again.

Tested on 10 September 2026 with Pillow 12.2.0 and ImageMagick 6.9.13-29, on one image. The rows marked “the other tool” are ImageMagick; the rest are Pillow.

The result that should worry you

Line two. We opened the file and saved it again. We did not resize it, crop it, compress it or convert it. Both provenance markers were gone.

That matters because “open it and save it” is not an edit anyone would think to mention. It is what happens when a plugin optimises your images, when a content system generates thumbnails, when someone opens a picture to look at it and clicks save on the way out. Nobody involved would describe what they did as removing metadata.

And the library is not losing the information through carelessness. It reads it perfectly well — the tag is sitting there in the image object, available to anyone who asks for it. It simply does not write it back out unless you tell it to. The provenance is not destroyed by the edit; it is dropped at the save.

The half-kept file, which is worse than a stripped one

The most awkward row is the sixth. Resizing the PNG with ImageMagick kept the “Made with Google AI” credit and lost the Digital Source Type tag.

So the file still announces that it was made with AI. Anyone opening it in a metadata viewer sees provenance and moves on. But the specific machine-readable code Google names — the one a feed check would look for — is no longer there.

A file that has plainly lost everything at least tells you it has lost everything. A file that has kept the human-readable half is the one that gets waved through.

What to do

1. Check one image rather than assuming. Take a picture that came out of an AI tool, put it through whatever your normal process is, and look at what came out the other end. You are looking for “DigitalSourceType” or “trainedAlgorithmicMedia” in the file.
2. Suspect the automatic steps most. Image optimisation plugins, thumbnail generators and upload pipelines are where files get opened and re-saved without anyone deciding to edit them.
3. Keep the original. If the tag is gone from the working copy, the untouched download is where it still exists.
4. Do not rely on metadata for disclosure that matters. We composite a visible “AI GENERATED” badge onto every illustration we publish, because a mark in the pixels survives what a mark in the header does not — and because the EU AI Act asks for disclosure that is clear at first exposure, which a file header is not. We wrote about that requirement in what the AI Act actually asks you to disclose.
5. Remember the scope. This is Google’s product feed rule. It is not a law, and it does not govern the photographs on your own website.

We ran it again on a different image

The test above was run on 10 September using this page’s own illustration. Before publishing, we generated a completely different image from the same tool on 12 September and put it through all eight operations again, because one image proves very little.

Every row came back the same. The fresh file arrived carrying both markers. Opening and re-saving it with no edit removed both. Resizing, cropping and saving as JPEG through the same library removed both. Converting to JPEG with the other tool kept both. And the awkward one repeated exactly: resizing the PNG with the other tool kept the “Made with Google AI” credit and lost the Digital Source Type tag.

So this is not a quirk of one file. It is how these tools behave.

One more check, and it is slightly at our own expense. We looked inside the finished illustration at the top of this article — the one you are actually looking at, after we cropped it, drew a headline on it and added our logo. Both provenance markers are gone. Our own publishing pipeline strips exactly what this article is about, which is precisely why the visible badge in the corner is there and the file header is not doing the work.

What argues the other way

This is one image, two tools, one afternoon. We cannot tell you what Photoshop, Canva, Lightroom, a phone gallery app or your shop platform does on upload, and it would be dishonest to imply otherwise — the honest instruction is to check your own chain rather than trust our table.

The behaviour is also not a bug. Stripping metadata on save is often deliberate, because metadata carries things people do not want published — location, camera serial numbers, editing history. A tool that quietly discards it is protecting somebody. The awkwardness is that the same broom now sweeps away a tag a platform wants kept.

And provenance metadata was never robust. It is a note attached to a file, and it survives exactly as long as everything handling that file agrees to carry it. That is a social arrangement more than a technical one. Anyone treating an IPTC tag as proof of how an image was made is trusting every piece of software that touched it in between, which is a longer list than it looks.

We would rather say that plainly than sell the tag as a guarantee. It is worth keeping. It is not evidence. That distinction is the same one we ran into when looking at what AI watermarks actually prove, and the answer there was similarly modest.