A model deciding whether to trust a claim has no editor looking over its shoulder — only whatever signals happen to be attached to the content it retrieves. ClaimReview is one of the oldest and most widely adopted of those signals: a piece of schema.org markup, built for fact-checkers, that turns "we checked this and found it false" into something a machine can parse. Understanding what it does — and where it stops — is a prerequisite for any credibility strategy built for an AI-mediated web.
What ClaimReview actually is
ClaimReview is a schema.org markup type, developed with Google and adopted across the fact-checking community, that lets a publisher attach a structured verdict to a specific claim: who made the claim, what the claim states, who reviewed it, and what rating that review produced, on a scale the publisher defines. It was built for exactly one purpose — making the output of fact-checking organisations machine-readable, so that search engines and, increasingly, AI systems could surface "this has been checked" alongside a claim, rather than leaving that judgment invisible in prose a crawler has to interpret.
It is not a badge you apply to your own content to declare it accurate. ClaimReview describes a review of a claim, produced by whoever performed that review — typically an independent fact-checking organisation, not the entity being written about. That distinction is the whole point: the markup carries weight because it is structurally separated from self-interest.
How fact-checkers actually use it
Organisations that produce fact-checks as their core output — the members of the International Fact-Checking Network among them — publish ClaimReview markup on every verdict they issue, because it is how that verdict reaches the platforms and models that might otherwise never see a single fact-checking article among the billions of pages available to retrieve. The markup travels with the content; a crawler encountering the page picks up both the human-readable verdict and its structured counterpart in the same pass.
The practical effect is visibility disproportionate to any single article's traffic. A fact-check that few humans read directly can still shape what a search engine — or a model trained on a corpus that includes it — surfaces when the underlying claim resurfaces somewhere else entirely.
Why models look for structured signals at all
Plain prose is expensive for a machine to evaluate for credibility; structured data is comparatively cheap. A sentence buried in paragraph four of an article, hedged with qualifiers, requires real interpretation to extract a verdict from. A ClaimReview block states the claim, the rater, and the rating in a fixed, parseable shape — no interpretation required, no risk of a summarisation step dropping the caveat that mattered.
That is not unique to ClaimReview. Any structured signal — a dated Organization schema with a stable identity, an author with a verifiable track record, a page's last-modified timestamp, a citation graph pointing back to primary sources — reduces the interpretive burden on a system that has to make a retrieval or weighting decision in a fraction of a second. Structure, in other words, is a form of legibility, and legibility is what gets rewarded when a system has to choose fast.
ClaimReview alongside other credibility signals
ClaimReview addresses one narrow case — a specific claim, formally reviewed. It sits alongside a broader family of structured signals worth distinguishing:
- Organization and sameAs markup — establishes a stable, cross-referenced identity for an entity, so a model can tell your official presence apart from an impersonator or a namesake.
- Author and byline metadata — attaches a traceable, accountable identity to a piece of content, rather than leaving it anonymous.
- dateModified and versioning — signals whether a page reflects current or stale information, which matters enormously given how retrieval timing works.
- ClaimReview itself — the narrowest and most explicit signal: a specific verdict, on a specific claim, from a named reviewer.
None of these substitutes for the others. A well-marked-up organisation with no ClaimReview coverage is still vulnerable to a specific false claim circulating unchecked; a page full of ClaimReview blocks but no stable entity identity still leaves a model guessing who, exactly, is being discussed.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
What structured markup cannot do
Markup is a signal, not a guarantee. A model is not obligated to weight ClaimReview data over conflicting prose it retrieved elsewhere, and plenty of systems — particularly those relying on parametric memory rather than live retrieval — never encounter the markup at all, because it was published after their training cutoff, or on a page their crawler never reached. Structured data raises the odds of being read correctly; it does not force the outcome.
There is also a volume problem. ClaimReview coverage concentrates on claims fact-checking organisations have chosen to review, which tends to mean claims with existing public visibility — political statements, viral rumours, widely shared health claims. A quieter factual error about a mid-sized company's leadership history is unlikely to ever receive a formal ClaimReview verdict, however wrong it is.
Neither problem is a reason to skip markup — it is a reason not to treat it as the whole strategy. Structured data is the layer you control directly; the layer a fact-checking organisation chooses to review is not, and building a credibility strategy that depends entirely on the second one happening is building on a step you cannot take yourself. The practical response is to treat markup as one input alongside the others covered above, not as a silver bullet that makes the rest of the work optional.
Putting structured signals to work
For most organisations, the realistic move is not to become a fact-checking publisher yourself, but to make every other signal as strong as it can be: a consistent, cross-referenced Organization identity across your properties; authored, dated content with real bylines; a stable canonical page for every fact a model might need to retrieve; and, where a genuine third-party fact-check exists — favourable or not — no attempt to suppress or obscure it, since a documented review, even a critical one, is itself a credibility signal about how seriously the underlying claim was investigated.
The goal is not to game a ranking signal. It is to remove the ambiguity that forces a model to guess, so that when it does retrieve something about you, the structure around that content already points toward the answer you would give it yourself.
Audit these signals the same way you would audit any other exposure: pick a handful of pages a model is likely to retrieve, check whether the Organization markup is present and consistent, confirm bylines and dates are actually populated rather than templated placeholders, and note anywhere a stale or duplicate entity record might confuse a crawler about who you are. This kind of audit rarely takes long, and it usually surfaces at least one inconsistency nobody had noticed because no single team owns every page where a given fact appears.
Structured signals inside SetRecord's method
When we score how exposed an organisation is inside AI-generated answers, credibility signals are one input among several — alongside provenance, source diversity and the other variables behind our risk scoring. A well-marked, well-documented factual record does not eliminate exposure, but it changes what a model has to work with when it decides what to say about you next.
In practice, this means our audits flag missing or inconsistent markup alongside factual gaps, rather than treating the two as separate workstreams — a false claim and an unmarked, ambiguous entity record often trace back to the same underlying gap in how an organisation's public information is structured, and fixing one without the other tends to leave the exposure only partly addressed.