SourceCited / Provenance
Guide · Trust signals
Provenance labels are arriving in search results. C2PA credentials will not rank you — but on high-trust pages, credentialed original media is becoming worth the ten extra minutes.
C2PA Content Credentials (now ISO/IEC 22144) and Google SynthID are reaching Google Search and Chrome, and OpenAI now attaches both to generated images. Provenance is not a ranking factor and not a penalty on AI-written text — it is an emerging authenticity layer. Use credentialed original imagery on high-trust pages, and know the limits: re-encoding strips metadata, and absence proves nothing.
C2PA Content Credentials are cryptographically signed manifests recording a file's origin and edit history — now an ISO standard (ISO/IEC 22144). SynthID is Google's watermark family for AI-generated media, with 100B+ items watermarked and roughly 50M verifications to date. The 2026 shift: at I/O, Google announced provenance verification surfacing in Search (via Lens, AI Mode, and Circle to Search) and in Chrome, and OpenAI joined the C2PA steering committee (May 19, 2026), now pairing C2PA metadata with SynthID on generated images.
Plainly: provenance is a labeling and trust layer, not a ranking factor. There is no confirmed citation or ranking lift for credentialed media, and no penalty for AI-written text. Where it earns its keep: original, credentialed photography on high-trust pages — product originals, expert headshots, first-hand documentation — as authenticity labels start appearing beside media in search surfaces. If you publish AI-generated media to EU users, note the EU AI Act's Article 50 marking obligations apply from Aug 2, 2026 — a legal question to confirm with counsel.
Screenshots and re-encoding strip C2PA metadata, so a missing credential proves nothing. Detection tools and provenance are different things. Watermark robustness varies. Treat provenance as a signal you attach to what you originate — not a verdict on anything else.
The most common failure is not philosophical, it is plumbing: image CDNs, compression plugins and resize services strip metadata by default, quietly deleting credentials between upload and delivery.
Watermark-style marks (SynthID-class) survive some transforms that strip embedded manifests, and manifests carry detail watermarks cannot. That is exactly why Google pairs them — and why you should not rely on either alone.
Not because it is AI. Google targets scaled, low-value content regardless of who wrote it, and says so in its spam policies. Provenance labels flag media origin; they are not a text penalty.
No evidence yet — provenance is a trust layer, not a citation lever. The case for it is authenticity labeling on your original media as those labels reach search interfaces.
Use capture or editing tools that sign credentials (recent cameras, Adobe tooling) or the open c2patool, then verify the file at contentcredentials.org before publishing.