BlogFindings
Fact check schema markup: Google is phasing it out, and 7 of 45 fact-check pages still declared it
Google's fact check structured data documentation, which carries a last updated date of 2026-09-08, opens by saying that support for ClaimReview markup in Google Search is being phased out while the markup remains supported by the Factcheck Explorer Tool. Lantad asked all 1,419 hostnames in this repository's two committed corpus seed files for robots.txt and then the home page on 4 November 2026. 1,080 answered HTTP 200 with HTML, 622 of those carried a parseable JSON-LD block, and none declared ClaimReview. Following each home page's own fact-check links reached 46 pages on 8 hosts, 45 of which answered with HTML. 7 declared ClaimReview, every one of them on rappler.com, and all 7 carried the three properties Google documents as required.
That type is now being withdrawn from the place most publishers added it for. Google's fact check structured data documentation, read at source on 4 November 2026 and carrying a last updated date of 2026-09-08, opens with a sentence saying that it is phasing out support for ClaimReview markup in Google Search, and adds that the markup remains supported by the Factcheck Explorer Tool. So the field stays valid, the tool that reads it stays, and the search feature it was written for is going. This blog has measured the other half of that pattern before, when Google deleted the FAQ rich result documentation and the scanner still scored FAQPage: markup outlives the result it was added for, because nobody goes back and removes it.
Which raises a question worth a measurement rather than an opinion. If the one machine-readable carrier for a checked claim is leaving Google Search, how much of it was ever there? Lantad asked all 1,419 hostnames in its two committed corpus seed files for robots.txt and then the home page on 4 November 2026, and then followed those sites' own fact-check links to the pages where the markup would belong.
In short
- Google's fact check structured data page, read at source on 4 November 2026 and carrying a last updated date of 2026-09-08, states that it is phasing out support for ClaimReview markup in Google Search and that the markup remains supported by the Factcheck Explorer Tool.
- Fact check schema markup appeared on 7 of the 45 fact-check pages Lantad read on 4 November 2026, and all 7 were on one host, rappler.com. The other seven hosts published fact-check work with no ClaimReview node anywhere on the pages read.
- ClaimReview appeared on none of the 1,080 corpus home pages Lantad read on 4 November 2026, of which 622 carried at least one JSON-LD block that parsed. Neither did the Claim type, on any of them.
- Only 8 of those 1,080 readable home pages linked a fact-check section or article at all, 6 of them news sites and 2 healthcare bodies, against 88 readable home pages in the news and media-local strata.
- All 7 ClaimReview nodes carried claimReviewed, reviewRating and url, which Google documents as required, and all 7 rated the claim False. None carried appearance or firstAppearance, and 6 of the 7 claimReviewed values ran longer than the 75 characters Google's page asks authors to try to stay under.
| What was read | Pages | Carried parseable JSON-LD | Declared ClaimReview |
|---|---|---|---|
| Corpus home pages | 1,080 | 622 | 0 |
| Fact-check pages that answered with HTML | 45 | 43 | 7 |
| Fact-check pages on rappler.com, the only host declaring it | 13 | 13 | 7 |
What is fact check schema markup, and what does Google require?
It is one schema.org type doing one job. schema.org's ClaimReview definition, read on 4 November 2026, describes a fact-checking review of claims made or reported in some creative work, referenced through the itemReviewed property, and defines claimReviewed as a short summary of the specific claims reviewed. reviewRating holds a Rating, which the same page describes as the rating given in this review. The structure is deliberately plain: here is the claim, here is what we decided about it, and here is where the claim came from.
Google's page narrows that to a supported subset and states what is mandatory. The required properties are claimReviewed, a short summary of the claim being evaluated; reviewRating, the assessment of the claim; and url, a link to the full fact-check article on the same domain or a subdomain as the page. Inside the rating, alternateName carries the textual verdict, the word a reader would recognise as False or Misleading. The page also asks authors to try to keep claimReviewed under 75 characters, and gives the reason as minimising wrapping on a mobile device, which is a presentation concern rather than a validity one.
The eligibility conditions are stricter than the syntax, and they are the part that explains why this type was never going to be widely used. Google's page says a site must have several pages marked with ClaimReview structured data, must meet the standards for accountability, transparency, readability and site misrepresentation set out in its News General Guidelines, and must have a corrections policy or a mechanism for users to report errors. It states that websites for political entities such as campaigns, parties or elected officials are not eligible. On the technical side it says that to be eligible for the single fact check rich result a page must only have one ClaimReview element, and that adding several per page makes the page ineligible for it.
Read together, that is a type addressed to working newsrooms with a published corrections policy, not to sites in general, which is worth holding onto before reading any adoption figure as a failure. A florist has no claims to review. The useful question is not how many websites declare ClaimReview but how many of the sites doing fact-check work declare it, and that is a question about a much smaller population. It is also a question that matters beyond Google, because the same fields are what any system would read to tell a verdict from an assertion, which is the distinction that decides whether a checked claim can be cited as checked in answer engine optimization.
| Property | Google's classification | Present in the 7 nodes read |
|---|---|---|
| claimReviewed | Required | 7 of 7 |
| reviewRating | Required | 7 of 7 |
| url | Required | 7 of 7 |
| reviewRating.alternateName | Required inside Rating | 7 of 7, every one False |
| itemReviewed | Recommended | 7 of 7, each a CreativeWork |
| appearance or firstAppearance | Recommended, one not both | 0 of 7 |
| claimReviewed under 75 characters | Asked for, not required | 1 of 7 |
How many corpus home pages declared ClaimReview?
None of them. Lantad requested https://host/robots.txt once for each of the 1,419 hostnames in this repository's two committed corpus seed files on 4 November 2026, resolved the path / for the LantadBot token with the shipped parser and evaluator in core/src/robots.ts, and read the home page of every host its own file permitted. The two frames are worker/seeds/corpus-seeds-platform.json, holding 392 hostnames in ten platform strata, and worker/seeds/corpus-seeds-industry.json, holding 1,027 in eight industry strata, with no host in both. Every request went out as LantadBot/1.0 with redirects followed and no JavaScript executed, which is the single-request method set out in the method behind every score and on the scanner's own crawler page.
1,080 hosts answered HTTP 200 with an HTML content type. 622 of those 1,080 carried at least one script block of type application/ld+json whose contents parsed as JSON, and those 622 blocks were flattened with flattenDeep from core/src/jsonld.ts, which descends arrays, @graph and ordinary property values so that a node nested under a publisher or an author property is still seen. Across every typed node on every one of those pages, ClaimReview appeared zero times. So did Claim, the type Google's page recommends for describing the claim's own appearances.
Zero is the expected answer here and it is worth saying why before it is read as a finding about neglect. A home page is not where a fact check lives. ClaimReview belongs on the article that reviews one specific claim, so a sweep of home pages tests only whether any site treats the type as site-furniture, the way many treat Organization or WebSite nodes. None does, which is correct behaviour rather than a defect. The same sweep has shown the opposite pattern for types that genuinely are site-furniture, when 141 of 382 home pages carried no structured data at all in the raw HTML and the ones that did mostly described the site rather than its contents.
The home-page pass earned its place by producing the number that actually constrains this subject. Of the 1,080 readable home pages, only 8 linked anything whose path named a fact check, a debunk, a disinformation section or a truth-or-fake segment. Six of the eight are news sites: cnn.com, france24.com, grist.org, motherjones.com, news24.com and rappler.com. Two are healthcare bodies publishing myth-correction material, aap.org and cma.ca. Against 88 readable home pages in the news and media-local strata, that is 6 that advertise fact-check work from the front door. An earlier corpus reading found the news stratum distinctive in a different way, when 41 of 69 news robots.txt files disallowed GPTBot, so this is a sector that thinks about machine readers and still mostly does not route to a fact-check section where an AI crawler can find one.
Do the pages that publish fact checks declare the markup?
Mostly not, and the split is by publisher rather than by page. Following the fact-check links from those 8 home pages reached 14 pages, and following the same path pattern one level further from each of those reached another 32, for 46 pages in all. Every one of the 46 was permitted by its own site's robots.txt for the LantadBot token, resolved against that page's own path rather than the root. 45 answered HTTP 200 with HTML; the exception was aap.org, which answered 503 at its fact-checked news-room index. 43 of the 45 carried at least one parseable JSON-LD block, the two without being a cma.ca myth-correction article and a motherjones.com politics piece, both of which shipped no JSON-LD at all.
7 of the 45 declared ClaimReview, and all 7 were rappler.com. Of the 13 rappler.com pages read, 7 were individual fact-check articles and each carried exactly one ClaimReview node; the other 6 were a section index, a second index page, a methodology page, an FAQ about the fact-check project and two copies of its standards and corrections policy. The remaining seven hosts contributed 32 pages between them and not one ClaimReview node: cnn.com 11 pages, news24.com 11, france24.com 7, grist.org 1, motherjones.com 1, cma.ca 1 and aap.org 1.
What those 32 pages declared instead is the specific thing worth reporting, because it is not an absence of markup. All 11 cnn.com pages were ten fact-check articles plus a section index, and every article declared NewsArticle alongside Person, Organization, NewsMediaOrganization, ImageObject, WebPage and WebPageElement nodes. The france24.com Truth or Fake pages declared NewsArticle with TVEpisode and TVSeries nodes attached, which describes the segment as television rather than as a claim review. news24.com declared NewsArticle on its debunking articles and LiveBlogPosting on its election debunking live blog. So these are pages with careful, complete structured data that says what kind of content they are and who made it, and says nothing at all about the claim they assess or the verdict they reached.
That is the same shape this blog found on community pages, when zero of 88 declared discussion forum structured data while declaring WebSite and organisation nodes describing the site around them. The markup present describes the container; the markup absent describes the content. It is also consistent with the narrower mismatch the corpus has recorded inside article markup itself, where 27 of 165 article pages declared a headline their own h1 did not contain. A machine reading any of these 32 pages can tell it is a news article by a named author at a named publisher. It cannot tell that the article's purpose is to rule on a claim, or what the ruling was, without reading the prose.
| Host | Stratum | Pages read | Answered HTML | Carried JSON-LD | ClaimReview nodes |
|---|---|---|---|---|---|
| rappler.com | news | 13 | 13 | 13 | 7 |
| cnn.com | news | 11 | 11 | 11 | 0 |
| news24.com | news | 11 | 11 | 11 | 0 |
| france24.com | news | 7 | 7 | 7 | 0 |
| grist.org | news | 1 | 1 | 1 | 0 |
| motherjones.com | news | 1 | 1 | 0 | 0 |
| cma.ca | healthcare | 1 | 1 | 0 | 0 |
| aap.org | healthcare | 1 | 0, answered 503 | 0 | 0 |
What did the seven ClaimReview nodes actually contain?
They were correct, and they were correct in a way that is worth describing precisely, because a measurement of adoption that does not check quality tends to overstate both. Each of the 7 nodes arrived as a single JSON-LD object carrying a two-value @type array, Article and ClaimReview together, rather than as a separate ClaimReview node beside an article node. That is legal and it is also why a flattener that walks only @graph would find nothing here: the node sits under the page's graph as the article itself. Lantad's reading used flattenDeep and entityTypes from core/src/jsonld.ts, the latter of which normalises a compact or full IRI to the bare term, so a type written as schema:ClaimReview or as a schema.org URL is recognised the same way.
All 7 carried claimReviewed, reviewRating and url, which are the three properties Google's page lists as required, and all 7 carried alternateName inside the Rating. Every one of the 7 verdicts was the word False. All 7 carried itemReviewed as a CreativeWork with datePublished, dateModified and an author Organization naming where the claim had appeared, which in these cases were social and video accounts: State News Online, PINAS BALITA TV twice, Viral Philippines, Makata Ng Bayan, Tom Rushmaster Limpao and Birada Kagay-an. Each page carried exactly one ClaimReview element, which is the condition Google's technical guidelines set for the single fact check rich result, so these pages are inside that rule rather than tripping it.
Two deviations from Google's recommendations are worth recording, and neither is a validity failure. None of the 7 carried appearance or firstAppearance, the properties the page recommends for pointing at where the claim was made, and the page is explicit that an author should use one or the other rather than both. And on length, 6 of the 7 ran past the 75 characters Google's page asks authors to try to stay under: the claimReviewed values measured 48, 101, 108, 123, 170, 193 and 213 characters, so only one was inside the guidance. These are summaries of political claims about named officials, and compressing one into 75 characters without changing what was claimed is genuinely hard, which is a reasonable thing for a publisher to lose on deliberately.
The honest reading of 7 of 45 is therefore not that one publisher is careless and the rest are not. It is that one publisher has wired a fact-check template to emit the type and the others have not, and that where the type is emitted it is emitted properly. The corpus has found the reverse often enough to make the distinction real: 117 of 612 home pages with JSON-LD carried a defect, 33 of 615 declared a rating and 14 of those rated their own business, and 8 of 111 article pages named a brand as the person who wrote them. Declared and correct are different measurements, and on this type the small population that declared it got it right.
JSON-LD node, rappler.com fact-check article, read 4 November 2026
- "@type": ["Article", "ClaimReview"] both types, one node
- "claimReviewed": "Cagayan de Oro 1st Representative Lordan Suan threatened ..." 213 chars
- "reviewRating": { "@type": "Rating", "alternateName": "False" } required, present
- "url": "https://www.rappler.com/newsbreak/fact-check/..." required, present
- "itemReviewed": { "@type": "CreativeWork", "author": { "name": "State News Online" } } recommended, present
- "appearance" / "firstAppearance" absent
- ClaimReview elements on the page 1, inside Google's limit
Should a publisher still add it, now that Google is phasing it out?
For the Google Search rich result, the documentation answers that itself: it is being phased out, so adding the markup to win that result is spending work on a feature its owner says is going. Lantad has measured no effect of ClaimReview on any AI answer, has run no test of whether any answer engine reads the type, and knows of no published study that settles it, so nothing here supports the opposite recommendation either. What can be said from this sweep is narrower and it is about carriers rather than outcomes: on 38 of the 45 fact-check pages read, no node anywhere in the served markup declared ClaimReview or Claim, so nothing in the structured data said that a claim had been assessed or what the assessment was.
That gap is the part worth acting on, and it does not depend on any one vendor's feature surviving. Google's page says the markup remains supported by the Factcheck Explorer Tool, which is a separate consumer at toolbox.google.com/factcheck/explorer and is written as plain text here because that host is not a registered outbound source on this site. Beyond it, these fields are the schema.org vocabulary for a verdict, and an answer engine summarising a disputed claim has to get the verdict from somewhere. If it is in prose only, it is being inferred. The platform guides on getting cited by Google AI Overviews and by ChatGPT set out what each vendor documents about retrieval, and neither names this type, which is itself the state of the evidence.
Three things are cheaper than the markup question and bear on it directly. First, whether the fact-check section can be reached at all: only 8 of 1,080 readable home pages linked one, so the most common failure here is routing rather than markup, and a section an AI crawler never reaches cannot be read whatever it declares. Second, whether the crawler is permitted on that path: every one of the 46 pages here was allowed, which is worth confirming on your own site with the robots.txt tester rather than assumed, since a rule written for a directory can close a section the home page advertises. Third, whether the verdict is in the served HTML before any script runs, which is what the GPTBot view shows and what the corpus figures in the crawlability study are built from.
The broader point is the one this corpus keeps arriving at from different directions. Markup that describes the container is widely deployed and markup that describes the claim is not, and the same asymmetry showed up when 107 of 163 schema topic statements named the site itself and when 85 of 317 organizations named a reference entry. Fact check schema markup is now an unusually clear case of it: a type with a precise job, a documented syntax, a shrinking consumer, and 7 nodes on one host out of the 45 pages these publishers themselves filed under fact checking. For generative engine optimization the lesson is not to rush this field in. It is that the thing you most want a machine to repeat about your work, the verdict, is the thing you are least likely to have told it.
Flow: 1,080 readable home pages (not reachable) to 1,072 linked no fact-check section; 1,080 readable home pages to 8 linked one; 8 linked one to 46 fact-check pages read; 46 fact-check pages read to 45 answered with HTML; 45 answered with HTML (verdict in prose only) to 38 declared no ClaimReview; 45 answered with HTML (verdict in fields) to 7 declared ClaimReview.
Lantad
Published .
A fact check is a verdict about a claim, and a verdict is the one thing an answer engine most needs and can least infer. Structured data has a type built for carrying exactly that. schema.org defines ClaimReview as a fact-checking review of claims made or reported in some creative work, referenced through itemReviewed, and the whole point of the type is that the claim, the rating and the thing being rated arrive as separate machine-readable fields rather than as prose a reader has to interpret.
Common questions
What is fact check schema markup?
It is the schema.org ClaimReview type, which carries a reviewed claim, the verdict on it and a reference to where the claim was made. schema.org, read on 4 November 2026, defines it as a fact-checking review of claims made or reported in some creative work. Google's documentation lists claimReviewed, reviewRating and url as required properties, with alternateName inside the rating holding the textual verdict such as False.
Is Google still using ClaimReview markup?
Google's fact check structured data page, read at source on 4 November 2026 and carrying a last updated date of 2026-09-08, states that it is phasing out support for ClaimReview markup in Google Search, and that the markup remains supported by the Factcheck Explorer Tool. Lantad has not measured what any engine does with the type and reports only what that page says.
How many sites actually declare fact check schema markup?
In Lantad's reading of 4 November 2026, none of the 1,080 corpus home pages that answered with HTML declared ClaimReview, and 7 of the 45 fact-check pages that answered with HTML did. All 7 were on rappler.com. The other seven hosts whose home pages linked fact-check work contributed 32 pages with no ClaimReview node, declaring NewsArticle, LiveBlogPosting or TVEpisode markup instead.
Does ClaimReview markup need to be on every page of a site?
No, and Google's page sets the opposite constraint in two directions. It says a site must have several pages marked with ClaimReview structured data to be eligible for the fact check rich result, and that an individual page must only have one ClaimReview element to be eligible for the single fact check rich result. All 7 nodes Lantad read on 4 November 2026 sat one per page.
See what AI can read on your site
Run a free scan and get a graded report of exactly what AI crawlers can and cannot read, with ranked fixes.
Structured data for AI (schema markup)
Structured data is schema.org markup, usually written as JSON-LD, that labels a page's facts in a machine-readable format so AI systems can extract and attribute them without having to interpret the surrounding prose.
AI visibility
AI visibility is whether AI systems can access, read, and faithfully represent your web pages, the prerequisite for being cited in an AI answer.
How to get cited by Google AI Overviews
Google AI Overviews are built from Google's Search index, so getting cited starts with a page Googlebot can crawl, index, and read, and no site can promise the citation itself.
Google deleted the FAQ rich result documentation, and we still score FAQPage
Google's documentation changelog records the FAQ rich result leaving Search on 7 May 2026 and its documentation being removed on 15 June 2026. Read on 5 August 2026 the old URL answers 301, schema.org still defines FAQPage, and Lantad's schema check still counts it as a recognised business type.
Discussion forum structured data: zero of 88 community pages declared it, and 6 shut out GPTBot
Google documented the UGC Fresh Data Program on 8 October 2026 and named on-page SocialMediaPosting or DiscussionForumPosting markup with interactionStatistic sub-fields among its eligibility criteria. Lantad asked all 1,419 hostnames in this repository's two committed corpus seed files for robots.txt on 9 October 2026 and read every permitted home page with no JavaScript executed. 1,090 answered HTTP 200 with HTML and 626 carried a JSON-LD block that parsed. None declared either type. Following each site's own link to a community area reached 88 readable pages, and none of those declared either type either.
Schema markup validator: 117 of 612 home pages with JSON-LD carried a defect, and on 7 not one block parsed
Lantad requested the home page of all 1,419 hostnames in this repository's committed corpus on 21 September 2026 and read the raw bytes with no JavaScript executed. 1,072 answered HTTP 200 with HTML, and 612 of those carried at least one JSON-LD block, 1,143 blocks between them. 117 of the 612 carried a defect, 16 blocks never parsed as JSON, and on 7 sites not one block on the page did.
