Marking and metadata in practice
Marking is the easy part; keeping it is not. A provenance manifest written at generation and stripped by an image CDN two steps later leaves you with the obligation and none of the evidence.
The three mechanisms
| Mechanism | What it is | Survives |
|---|---|---|
| C2PA Content Credentials | Signed manifest of how the file was made | Copying; not re-encoding that drops metadata |
| IPTC Digital Source Type | A metadata field naming the source as algorithmic | The same conditions as any IPTC data |
| Invisible watermark | A signal embedded in the pixels | Re-encoding, resizing, moderate cropping |
C2PA is the most informative and the most fragile. A watermark is the most durable and the least informative. Neither replaces the other, which is why the answer is both plus your own record.
Where marking gets destroyed
- Image CDNs. Most strip all metadata on resize by default, in the name of file size. This is the most common cause.
- Build pipelines. Optimisers such as imagemin or squoosh drop metadata unless told not to.
- Ecommerce platforms. Many re-encode on upload and generate their own derivatives.
- Feed processing. Any step that fetches and rewrites images rather than passing URLs through.
- Social platforms. Most re-encode on upload; assume metadata does not survive and rely on the watermark and on platform-level disclosure.
Test it, do not assume it. Take one generated image through your entire pipeline to the live product page, download it from the page, and inspect the metadata. Most teams find it gone at step two.
Keep your own record
Whatever survives in the file, keep a record you control: asset id, source photograph, tool and model version, prompt or template version, operator, timestamp, and where it was published. This is the artefact that answers a question from a regulator, a platform or a customer, and it does not depend on a CDN’s configuration.
Practical configuration
- Turn metadata preservation on explicitly in your CDN’s image pipeline.
- Add a check to CI that a known generated asset still carries its manifest after processing.
- Store originals with marking intact, separately from web derivatives.
- Where the platform will strip everything anyway, disclose at the platform level using its own AI-content controls.
Visible disclosure is a different thing
Machine-readable marking is invisible and belongs in the file. Visible disclosure, meaning telling a reader that an image is generated, is a separate obligation that applies in narrower circumstances, and it belongs on the page rather than burned into the image, where Merchant Center would reject it.