ClaimReview Markup Turns Human Fact-Checks Into Structured Data for Search and AI

The outcome of a digital fact-check is not determined when the journalist publishes the article, nor when a reader clicks the link. It is determined the moment a content management system injects a block of JSON-LD code into the page's invisible header. That block, known as ClaimReview markup, is the only language search engines and artificial intelligence agents actually understand. Without it, a meticulously researched debunking is just another page of unstructured text; with it, the article becomes a structured data object that machines can parse, index, and deploy.[1][4]

Created in 2015 through a collaboration between the Duke Reporters' Lab, Google, and the schema.org standards group, ClaimReview was designed to solve a specific machine-vision problem. Search crawlers cannot read a 1,200-word political analysis and reliably extract the core verdict. The markup forces the publisher to map their journalism to three rigid fields: the `claimReviewed` (what was said), the `author` (who said it), and the `reviewRating` (the standardized verdict).[4]

For nearly 10 years, this structured data powered the consumer-facing features that tech platforms touted as their primary defense against misinformation. By 2025, more than 250,000 fact-checks had been tagged with the markup, generating an estimated 4 billion annual impressions across Google's services.

When a user searched for a viral rumor, the search engine would read the ClaimReview data and generate a prominent badge directly on the results page. The platforms marketed this as elevating truth, but the capability was entirely dependent on human journalists manually filling out web forms to feed the platforms' databases.[2]

The consumer-facing utility of the markup, however, was abruptly curtailed. In July 2025, Google quietly retired the visual ClaimReview badges from its primary search results, stating that removing them would "help streamline the results page and focus on other experiences." The decision removed the immediate visual cue that users relied on, leaving fact-checking organizations frustrated that their structured data was no longer being surfaced to the public in the way it was originally promised.[2]

Yet the deprecation of the search badge did not render the markup obsolete; it simply shifted its primary audience from human readers to machine-learning models. While Google phased out the visual labels in Search, it maintained the Fact Check Explorer tool, which relies entirely on the ClaimReview data feed hosted on DataCommons.org. The massive, standardized archive of tagged claims has instead become foundational training data.[1][3]

Artificial intelligence developers now utilize the dataset to ground large language models and retrieval-augmented generation (RAG) systems. Because the markup explicitly links a false claim to a verified correction, it provides a clean signal that helps prevent AI agents from repeating debunked rumors. As Full Fact head of AI Andrew Dudfield noted during a May 28, 2026 summit of 85 researchers, the tool is now highly optimized for machines. "ClaimReview can be especially useful for communicating with non-humans," Dudfield said.[3]

The ecosystem is also expanding to address new formats. The Duke Reporters' Lab and its partners have developed MediaReview, a sibling taxonomy designed specifically to tag manipulated images and videos. As generative AI makes synthetic media cheaper to produce, the ability to structurally tag a specific video file as a manipulated deepfake in a way that platforms can automatically read is becoming the next critical layer of digital provenance.

The immediate future of the standard now depends on reducing the friction of data entry. Lead Stories cofounder Maarten Schenk estimates that manually encoding the schema takes about 30 seconds per article, a cumulative burden that falls entirely on the publishers. To sustain the database that currently anchors AI fact-checking models, the International Fact-Checking Network is pushing for deeper content management system integrations that generate the JSON-LD payload automatically upon publication, removing the manual step entirely.[3][4]