What Content Credentials on a product photo actually say
A C2PA manifest is a signed list of claims attached to the file (the mechanics are in our labeling guide). For a merchant, what matters is not the cryptography but the handful of statements a viewer will read back. They fall into three groups.
| What the record says | What it means for a listing | How a platform tends to treat it |
|---|---|---|
| Captured with a camera | A signing device took the original exposure | Strongest authenticity signal available |
| Edited (crop, colour, exposure) | Ordinary adjustments recorded as actions after capture | Usually no label |
| Edited with generative AI | A model re-drew part of the frame, however small | Often an "AI info" style label |
| Created using generative AI | The whole image came from a model | Labelled, and required to be on some channels |
| No credential at all | Nothing can be said either way | Treated as unknown, not as fake |
Two of those rows map onto the IPTC vocabulary platforms already read: trainedAlgorithmicMedia ("Created using Generative AI") and compositeWithTrainedAlgorithmicMedia ("Edited with Generative AI"). If you have read our guide to product image metadata, these are the same distinctions, now signed so they cannot be quietly rewritten.
The last row is the one most product photos sit in today, and it is worth being clear about: a missing credential proves nothing. The vast majority of genuine photos on the web carry none, because most cameras do not sign and most pipelines strip what is there.
Why a real, retouched photo can get labelled as AI
The most practical risk for a seller is not being caught using AI. It is a genuine photo being presented as AI-made because of one small edit.
Meta learned this in public. In 2024 its "Made with AI" label began appearing on real photographs that had been touched up in Photoshop; PetaPixel's tests found that removing a single speck with Generative Fill was enough to trigger it. After photographers complained, Meta renamed the label "AI info" in July 2024, and in September moved it into the post menu for content that was only edited rather than generated. The detection did not change: Meta still reads C2PA and IPTC metadata, and a generative edit still records itself as one.
A credential records which tool touched the pixels, not how much of the product changed. Removing a dust speck with a generative fill and inventing a whole scene can produce the same "edited with generative AI" assertion.
For a product photo this cuts both ways. If you use a generative tool for the background but leave the product as shot, the honest record still says "edited with generative AI", and a shopper or platform will read it that way. If you want a photo to read as a straight capture, the edits on it have to be ones that record as ordinary adjustments: crop, exposure, colour, and non-generative cleanup.
That is a decision to make per image type, not per file. Hero shots that need to prove the product is real are candidates for a clean chain. Lifestyle scenes built with AI should be labelled as such, and our post on AI image disclosure and customer trust covers how to do that without hurting conversion.
Where the chain breaks in an e-commerce pipeline
A credential survives only if every step between camera and shopper either keeps it or adds a signed entry of its own. Product images pass through more steps than most photos do.
The weakest links are the steps that re-encode. Any editor, retouching service or AI tool that decodes the image and writes a new file will drop the manifest unless it is built to sign its output. That is not a fault of any one product; most image tools today, including AI retouching tools, write a fresh file without a manifest. Treat it as the default and plan around it.
Delivery is the other weak point. Until early 2025, Cloudflare Images discarded credentials on every transform. Since February 2025 it has offered a Preserve Content Credentials setting that keeps the existing record and appends a signed entry for the resize or format change Cloudflare made. With the setting off, credentials are still discarded. Other CDNs and storefront image services vary, and few document it, which is why the only reliable answer comes from testing your own live image.
If a step in your pipeline makes a new file and does not say it signs, assume the credential ends there.
How to check what your live listings carry
You do not need engineering help to find out whether your product images carry credentials. It takes about ten minutes per channel.
- Pick a test photo with a known credential. A JPEG straight from a Pixel 10 camera works, as does an image exported from a tool that signs its output.
- Inspect the original. Drop it into the Content Authenticity Initiative's free verify page at contentcredentials.org/verify, or run the open-source
c2patoolon it from a terminal. Note what it reports: captured, edited, which tools. - Run it through your normal workflow. Retouch it, export it, upload it to the product exactly as you would a real listing.
- Download the image a shopper actually receives. Open the live product page, save the image the browser loads (not the file in your admin), and inspect it again.
- Repeat per channel. Your own store, each marketplace and each social shop can treat the same file differently.
Record the result in one line per channel: "kept", "kept and extended", or "stripped". That table tells you whether signing anything is worth the effort before you buy a camera or change a tool.
Checking the wrong file
- Inspecting the master in your DAM
- Inspecting the file in the store admin
- Assuming the tool's spec sheet is your pipeline
Checking the right file
- Saving the image from the live product page
- Testing each channel separately
- Re-testing after any CDN or theme change
Trust signals for real photos: when signing is worth it
For most catalogs, a full signed chain from camera to shopper is not achievable today, and that is fine. The question is where a credential earns its keep.
It is worth it where authenticity is the product. Resale and pre-owned goods, collectibles, vintage, art and anything where condition drives price all benefit from proof that the photo shows the actual item. A signed capture of the scuff on a pre-owned bag is a stronger answer to "is that really the one I'll receive" than any caption.
It is worth it when you supply images to others. Wholesale partners, marketplaces and press increasingly ask where an image came from. A credential on your masters answers that once, and our guide to image handoffs covers keeping that information intact when files change hands.
It is rarely worth it for routine catalog shots of new, identical stock, where nobody doubts the product exists and the pipeline would strip the record anyway.
Keep the limits in view. Credentials depend on the signer being trustworthy: in September 2025 Nikon suspended its Z6III authenticity service after a user showed the camera could sign an image it had not taken, and the company later revoked the certificates it had issued. A credential is evidence, not proof, and a platform's reading of it can change.
Generated images: where credentials are not optional
For fully AI-generated product images the question flips from "is it worth it" to "do not remove it". Google Merchant Center's policy asks that images created with generative AI carry the IPTC trainedAlgorithmicMedia tag and that sellers do not strip embedded metadata such as the IPTC DigitalSourceType from images made with generative tools. The EU's transparency rules point the same way, which our post on EU AI Act labeling covers in detail.
Practically, that means a generated image needs its origin recorded somewhere your team can rely on, even if a CDN later drops the embedded copy:
- Keep the tool's original output, with whatever metadata it wrote, as the master.
- Track which images were generated or edited with AI in your catalog data, not only in the file.
- Re-apply the IPTC DigitalSourceType at export if your editor dropped it.
The second point is the one teams skip, and it is the one that survives every re-encode. Choosing settings helps here too: in Retouchable's Studio, every setting the model works on is marked ✦ before you press, and Exact keeps the product's pixels as shot, so you know at the moment of making an image which kind it is. Whichever tool you use, write that distinction down where your listing data lives.
A simple policy for a mixed catalog
Most stores now hold three kinds of product image side by side: straight captures, real photos with AI-assisted edits, and fully generated scenes. A one-page policy keeps the credentials question from becoming a per-image debate.
| Image type | Credential goal | Visible label |
|---|---|---|
| Straight capture, ordinary edits | Keep the capture record where your pipeline allows | None needed |
| Real product, AI-built background or cleanup | Let the record say "edited with generative AI" | Light disclosure where a channel asks |
| Fully generated scene or model | Never strip; re-apply IPTC at export | Required on some channels |
| Pre-owned or one-of-a-kind item | Sign at capture, minimal edits, preserve end to end | "Photo of the actual item" |
Revisit it twice a year. Camera support, platform labels and CDN settings are all moving, and the right answer for your catalog in spring may be different by autumn. The test in the section above is the cheapest way to know when it has changed.