Last tested: Oct 27, 2026. This page is re-tested quarterly and updated with a visible date.
Two weeks ago we explained why provenance metadata dies in distribution: platforms re-encode every upload, and the rebuild keeps the pixels, not the baggage. That piece made a general argument. This one replaces the argument with data.
We took certified AI images, each carrying a full, verified provenance set (a signed C2PA manifest with a trusted timestamp, and an invisible watermark), published them on four platforms the way any brand would, downloaded the public copies back, and verified what survived. Here is the matrix, cell by cell, with the method underneath so anyone can reproduce it.
The matrix
| Platform | C2PA manifest survives? | Provenance badge shown? | Watermark survives? | What a verifier finds afterwards |
|---|---|---|---|---|
| ❌ | ❌ | ✅ | No Content Credentials in any verifier; the watermark still decodes on 5/5 images. The desktop uploader crops images to 1:1 square by default | |
| ❌ | ✅ | ✅ | The post shows a "CR" Content Credentials badge ("issued by Modelia, Inc."), but the downloaded file has no credentials; the watermark decodes on 5/5 | |
| X | ❌ | ✅ | ✅ | X automatically labels the post "Made with AI" (no disclosure option was selected), but the served file has no credentials; the watermark decodes on 5/5 |
| ❌ | ❌ | ✅ | No Content Credentials, and no label in the chat; the watermark decodes on 5/5 (standard and HD) |
On all four platforms, the C2PA manifest was gone from the file the public receives; the watermark still decoded on 5 of 5 images on every platform. Two of the four, LinkedIn and X, still showed a provenance signal on the post itself, even though the downloaded file carried none.
How we tested (reproduce it yourself)
- 1. Source images: 5 certified images exported from Modelia Studio, each independently verified before upload (manifest valid, timestamp trusted, watermark detected).
- 2. Publication: uploaded through each platform's standard web or app flow, as a normal public post, no API shortcuts, because we are testing what happens to a brand publishing normally.
- 3. Retrieval: the public copy of each image was downloaded back, logged out, as the file the platform actually serves: on X, a browser fetch of the public media file at full size and default display size; on LinkedIn and Instagram, a browser fetch of the images served on the public post page; on WhatsApp, the receiving account’s Download button (WhatsApp Web).
- 4. Verification: every retrieved file was checked in our public verifier and cross-checked in an independent C2PA inspector. A layer counts as surviving only if it verifies, not merely if fragments are present.
- 5. Evidence: verification screenshots for every cell are archived with the test date. Behaviour may change after our test date, which is why this page carries one and is re-tested quarterly.
Reading the matrix as a brand
Three practical conclusions follow from the pattern, whatever the individual cells say.
- Your evidence lives at export. The moment your image is published on a re-encoding platform, the file the public sees is not the file you exported. Your proof of proper marking is your export record, not the wild copy. Keep it where you can produce it.
- The watermark is the working layer in public. The layer built into the pixels is the one still answering on stripped copies, which is exactly why the EU requires two layers instead of trusting metadata alone. When you evaluate any provenance solution, the survival column that matters most is the watermark's.
- The visible label is strip-proof. A visible label on deepfake-class imagery survives every platform in this table by definition: it is part of the pixels. Whatever else changes in this matrix over time, that row never will.
What this matrix is not
It is not a ranking of platforms by virtue. Stripping metadata on re-encode is largely an infrastructure behaviour with legitimate bandwidth and privacy roots, older than AI. And it is not permanent: platforms are moving, some toward preserving Content Credentials, and we expect cells in this table to improve. That is what the quarterly re-test is for. What the matrix is: the current, tested answer to the question every compliance and brand team eventually asks, so that nobody has to plan around guesses.
Verify any of this yourself with the public verifier, and find the full picture of duties and dates in the Article 50 guide. Read how Modelia images carry their marks, and what is deliberately not inside them.
Every image exported from Modelia carries signed provenance and an invisible watermark by default, and anyone can check any image, free, no account, in our public C2PA checker. The full picture, rules and industry data, is in The State of AI-Generated Content in Fashion 2026.




