The link works, the page is real, and the sentence is still wrong.
You pasted an AI-assisted paragraph, clicked the citation it handed you, and shipped it. The page loads, the topic matches, and it never states the thing your sentence claims. Every broken-link sweep and green status your stack runs waved that citation through, because each one only checks whether the page is there. Whether it backs your claim, none of them ask. The check you trust most misses the failure your AI draft is most likely to produce.
Claim: In a 125-chain audit, 30% of citation chains pointed to a page that loaded cleanly while the number it was cited for was no longer on it. Source: LiquiChart citation provenance study, pages where the data had been removed. Verified: 2026-05-29.
Why a Green Link Check Misses the Most Common AI Failure
A hallucinated citation is a reference to a page that loads, sits on the right topic, and still does not state the claim attached to it. Link checks pass, because the link works and the page is real. Only reading the page against your specific sentence reveals that the support is missing or reworded.
The word "hallucinated citation" used to mean a fabricated URL: a model invented a link, you clicked it, and the page did not exist. That version is easy to catch. The page 404s, a link checker flags it, you delete the sentence.
The 2026 version is harder. The model stops inventing the URL and starts inventing the relationship between a real page and your sentence. The link resolves, the topic matches, and the value you asserted is not on the page, or sits there as a different number. The hallucination moved from the link to the meaning. No link checker can see that.
Even the layer beneath it is not clean. Of the SaaS citations that link out, almost one in four resolves to something dead, gated, or broken. A broken-link sweep flags the outright dead ones. The semantic check sits a layer above even that. It assumes the link resolved and the page loaded, then asks the one thing the sweep never does: does the page say what your sentence says?
Borrowed numbers are where the mismatch concentrates. The same corpus work shows 73% of third-party claims are borrowed data, lifted from someone else's research and re-cited. A borrowed number loses precision in transit. Your sentence rounds, the source qualifies, an intermediate post restates, and the figure that lands in your draft sits close to the page without matching it.
How to Read a Page Against a Single Claim
The check is one sentence and one source. You take the claim your draft makes, open the page it cites, and look for the line that states it. Either the page carries a sentence that supports your figure, one you can quote back word for word, or it does not, and the citation has nothing under it. There is no third move.
You are not reading the whole post, and you are not following the link to wherever it points next. One claim, one page, the words on it.
The test is literal. If your sentence says 50% and the page says 48%, the page does not support your sentence, even though the topic is identical and the gap is small. Claim verification asks whether the page states the value you wrote, not whether you could derive it from what the page shows. You never normalize, round, or reconcile the two numbers in your head.
Reading a page against one sentence runs opposite to auditing a post. The Content Health Scanner takes a whole URL and works outward, pulling every statistic off the page and scoring each one for staleness. Claim verification works inward: one claim, the one source it leans on, and the question of whether that page states it. Outward auditing finds the claims you forgot you made. The inward read settles the one you are about to publish.
Doing it by hand is slow, which is why it gets skipped. Under deadline, the cheap version wins: click the link, see a real page, move on. The habit holds right until the page on the right topic does not carry your number. By then the citation has shipped.
Each of these checks tells you something different about the page you cited. Which one do you actually run before publishing?
Only one of these reads the page against the specific claim. The rest pass a citation that can still be wrong.
Every habit above stops at a different point on the way to the page. Trusting the link, skimming the topic, searching for the number: each one settles before it reaches the words and asks whether they back your sentence. That last step is the inward read, and the gap before it is exactly where a citation that points at the right page for the wrong figure slips through.
The same read works whether the paragraph came from a model or a junior writer. But AI drafts produce the mismatch at a volume hand-writing never did, which is why the manual check is worth turning into something you can run in seconds.
Run One Claim Through the AI Citation Checker
This is claim verification made runnable: the read you just did by hand, in seconds. Paste the source URL and the claim into the AI Citation Checker below. It reads the live page and tells you whether that page states what your sentence says, quoting the supporting line back word for word. Free, no account, ten checks a day from a single browser.
Claim verification has a deliberately narrow job. It does not rewrite your prose. It does not watch the source for you after today. Reading one claim against one page, once, before you publish, is the whole of it. Every check lands on one of five outcomes, and the two that matter most are the two a link check can never reach.
When the page does state your claim, the checker returns the supporting sentence verbatim, lifted from the live page. That quote is the proof, and it is guarded: the engine verifies every returned line as a literal substring of the page before showing it to you, so it cannot invent a sentence that flatters your claim. A quote you can read on the page beats a green status that only confirms the page answered. This is the spine of claim verification, and the reason it holds where a working link does not.
The Two Verdicts That Survive a Link Check: Reworded and Wrong Page
Two of the five outcomes are the whole argument, because both sail through every link check you own.
The first is reworded. Your draft says a tactic lifts conversion by 50%. The cited page is the right study, on the right topic, hosted on the right domain, and the link is healthy.
You read it against your sentence and the page says 48%. Not close enough to wave through as roughly half. The page states 48%, your draft states 50%, and those are different claims.
The checker returns the verdict reworded and quotes the page's own line back, so you see the 48% sitting where your 50% should be. The page stated 48% before you ever cited it. Your draft stated 50% the whole time. A broken-link sweep had nothing to flag, because nothing was broken. Only reading the words caught it.
The second is wrong_page. The link resolves to a real, authoritative page about something else. Your sentence cites a figure on email open rates; the link lands on a well-maintained page about email deliverability. Same domain, adjacent topic, healthy URL. A reader skimming for the general subject would nod and keep going. The page never makes your claim, because it covers the room next door.
Once the support check comes back negative, the checker asks a second question: is this page even on the subject? A page that fails on a subject it never covers reads as wrong_page, not a plain absence. The citation looks careful. It points at the wrong room.
Both verdicts share one property: the URL returns 200, the page renders, the topic looks right, and the claim has no support. That is the cell every automated check leaves empty. It is the cell an AI draft fills most often.
When the Page Is Honest and the Checker Is Honest: Not Found and Could Not Verify
The third outcome is the page being straight with you. not_found means the page is on the right topic, the link is healthy, and the specific claim is simply not stated anywhere on it. That is the honest result for a citation that points at a relevant page that happens not to carry your number. It is the verdict you will see most once you start reading sources against sentences instead of skimming for the gist.
The fourth outcome is the checker being straight with you. couldnt_verify is a real category, not a failure hiding behind a polite label. When a source is dead, sits behind a login or paywall, or stays an empty shell that yields no readable text even after the checker tries to render it, the page genuinely cannot be read.
Instead of guessing, the checker tells you it could not see the page, and why. A couldnt_verify does not mean your claim is wrong. It means the verdict is unavailable and the judgment is back in your hands. A check that admits what it could not read is worth more than one that fills the gap with a confident guess.
The fifth outcome, supported, is the one you are hoping for: the page states your claim and the checker quotes the line that proves it. That is the citation you keep.
A Source That Supports You Today Can Change Tomorrow
The check is one-shot by design. You verify the page today, the sentence holds, you publish. But the page underneath keeps moving, and the version that backed your claim is not the version a reader lands on next quarter.
Closing that gap is a different job from the pre-publish read. It means watching the sources you cite and getting told when one of them changes what it says. That is what a Monitored Page does, and the waitlist puts your cited sources under that standing watch. For the corpus-wide version of the problem, whether a whole chain of citations still reaches its source over time, see citation chain monitoring.
This is where claim verification stops being a single read and becomes content maintenance infrastructure. The pre-publish check is the gate. The standing watch keeps it shut after the draft is no longer yours to edit.
What You Stop Shipping When You Check Before Publish
The most dangerous citation in your back-catalog is the one that passes. It resolves to a credible page on exactly the right subject. The number it carries was never on that page. A broken citation announces itself: a sweep finds it, a reader hits a 404, someone fixes it. The one that passes does none of that, because every check you run stops at whether the page loaded.
You are the last reader before the rest of them. A figure that claims 50% over a page that says 48% does not get more true after you publish, and nothing downstream of the editor will catch it. It travels into the board deck. Into the competitor's rebuttal. Into the reader who cites you in good faith, carrying your name on a claim the source never made.