Report 01 · Baseline · Content provenance

Content provenance rarely survives the trip to your browser

We scanned — images as actually served from a stratified sample of the Tranco top‑1M. — carried a C2PA manifest. They came from just — sites, and five of those account for — of them.

By Cătălin · Net Census · 9 August 2026 · Methodology · Raw data · Prior work · Limits · CC BY 4.0

—%

of images served carried a content-provenance manifest.
95% CI —% – —% (cluster bootstrap over — sites).

What this does and does not measure. This is a measurement of provenance as delivered — what arrives in a browser. It is not a measurement of adoption. A publisher can embed credentials correctly and still serve images without them, because image pipelines re-encode. Cloudflare Images, for one, discards Content Credentials by default unless explicitly told to keep them. The gap between what is created and what arrives is the subject of this study, not an error in it.

The numbers

—

images fetched and scanned

—

carried a manifest

—

sites measured

—

% coverage of the sampled frame

What was already known

Provenance prevalence has been asserted more often than measured. The most-cited figure in the field comes from the Reuters Institute's Journalism, Media and Technology Trends and Predictions 2026, which states that fewer than 1% of news images or videos published globally include C2PA metadata. It carries no method, no sample and no source. It may well be right; there is no way to check it.

Three pieces of real prior work bound this study, and each measures something different:

Table 1 — Prior work, and what each actually measures.
Work	Measures	Finding
Rijsbosch, van Dijck & Kollnig, Missing the Mark (Policy & Internet, 2026)	Supply side — do generators emit marks at creation	38% machine-readable marking across 50 generative systems
Krawetz / FotoForensics	Uploads to a forensics service — self-selected population	~0.067%, roughly 12 C2PA images per week
Imatag, State of image metadata (2018)	Metadata survival on editorial sites, pre-C2PA	~3% of published photographs retained their metadata
This study	Distribution side — what arrives in a browser	—% of images served

Reconciling the apparent contradiction. Some industry estimates put provenance coverage far higher — one aggregator puts it at 75–85%, derived from coalition member disclosures and platform announcements rather than from measurement. Both figures can be true at once. If generators emit marks at creation and the delivery chain strips them, a high creation rate and a near-zero delivery rate are the same phenomenon seen from two ends. That gap is the finding.

Provenance is clustered, not thinly spread

The — images carrying manifests came from — sites out of — measured. The top five of those sites account for — of the total. Where provenance appears, it tends to appear across a whole site at once — consistent with a pipeline decision rather than per-image intent.

This clustering is also why the confidence interval here is wider than a naive calculation would give. Images are not independent observations; a site serving twenty images through one pipeline is closer to a single observation than to twenty. Treating them as independent would understate the uncertainty by a factor of roughly —× in variance.

It is not concentrated at the top

The expected pattern would be the largest publishers leading, with a long tail of nothing. That is not what the bands show.

Table 2 — Results by Tranco rank band. 700 domains drawn per band by fixed stride.
Rank band	Attempted	Measured	Coverage	Images	With provenance

The top 1,000 sites produced the fewest. The mid bands produced the most. We do not have causal evidence for why, and will not pretend otherwise. The pattern is consistent with provenance arriving through tooling defaults — an application or camera that writes manifests without being asked — while the largest sites run images through optimisation pipelines that re-encode and discard them.

The lightweight tag holds its own

Alongside C2PA we counted the IPTC digitalSourceType property, a single XMP field stating that content was machine-generated. It is far simpler than a signed manifest, and it appeared on — images against — carrying full C2PA — marginally more, in the same corpus, surviving the same pipelines.

Where the standard's own authors sit

We separately sampled the public websites of the content-provenance industry. Across 53 images, one carried a manifest — on opensource.contentauthenticity.org. Leica's own page describing its Content Credentials support returned none across 16 images; so did Truepic, Digimarc, contentcredentials.org and c2pa.org.

This is not hypocrisy and should not be reported as such. It is the same pipeline effect measured everywhere else in this study. These organisations are subject to the same CDNs and content management systems as everyone else. If anything it is the cleanest demonstration available that the problem is structural: the people with the strongest possible motive to serve provenance are not managing to either.

Among large platforms, youtube.com and youtu.be both returned images carrying manifests — evidence that surviving delivery at scale is achievable when a pipeline is built to preserve it.

Why this is being measured now

The EU AI Act's Article 50(2) requires providers of generative AI systems to mark synthetic output in a machine-readable format. It became applicable on 2 August 2026; systems already on the market have until 2 December 2026.

That duty does not fall on the websites measured here. Article 50(2) binds the providers of generative systems, not publishers who host images. Nothing in this study asserts or implies that any measured organisation has any obligation, still less that it has breached one. The Act is the reason the timing matters — it is not a yardstick we are holding anyone to.

The relevance is mechanical rather than legal. If marking is applied at generation and removed in delivery, an obligation can be met at one end of the chain and undone at the other, with nobody having decided to do anything wrong. Measuring the delivery end before the December date establishes a baseline that cannot be reconstructed afterwards.

What we could not see

— of — sampled domains could not be measured. Reporting that matters more than the headline.

Table 3 — Non-response, all causes.
Cause	Domains	Share

The largest cause is not blocking. A substantial share of the sampled domains do not resolve at all — Tranco is built from DNS query logs and so contains hosts that are not websites. Those were never measurable and their absence does not bias the result.

Deliberate blocking matters for a different reason. If it were concentrated among the largest sites the headline would be biased downward. It is not: coverage runs between 55% and 63% across the four bands, with the lowest coverage in the top band and the highest in the mid bands where we found the most provenance.

As an explicit upper bound, if every unmeasured domain carried provenance on every image the rate would be about 41%. That is deliberately absurd. It is stated because it is the first question a reviewer should ask.

Method, in brief

Draw 700 domains from each of four Tranco rank bands by fixed stride — deterministic, reproducible, no randomness.
Render each homepage in headless Chromium; collect image URLs from the rendered DOM including srcset and og:image.
Fetch the first 128 KB of each image with a ranged request; manifests sit near the head of the file.
Scan bytes for embedded manifests (JPEG APP11/JUMBF, PNG caBX, WebP) and remote manifests (XMP dcterms:provenance).
Confirm every candidate with c2patool 0.27.7, the C2PA reference implementation.

The detector was gated before it was trusted. It was run against the reference corpus until it found every known positive — 11 of 11, across JPEG, PNG and WebP, and across both embedded and remote manifests. An earlier version missed remote manifests entirely. Full detail, including the errors we made and corrected, in the methodology annex.

Reproduce it

The complete per-domain results are published with checksums, the exact Tranco snapshot identified by hash, and every field documented. The headline recomputes from the CSV in one line. If your numbers differ from ours we want to know — corrections@netcensus.org.

Next measurement: 2 December 2026, the date the AI Act transition period ends, using the same instrument and the same frame.