You pasted the paragraph your AI draft produced, clicked the source it attached, watched a real page load on a relevant topic, and hit publish. The number in your sentence was supposed to be somewhere on that page. It never was. Every automated check in your stack passed that citation anyway, because every one of them stops at the same question: did the URL resolve?
Hallucinated Citations After the Fake URL
A hallucinated citation used to mean a fabricated URL. The model invented a link, you clicked it, the page 404ed, and any link checker on earth flagged it. You deleted the sentence and moved on.
The 2026 version keeps the URL real. The model picks a genuine page on the right subject and invents the relationship between that page and your sentence. The figure you asserted sits on the page at a different value, or on a different page, or nowhere at all. The hallucination moved from the address to the meaning, and a tool that only tests addresses has no way to see it.
The address layer has its own rot. Almost one in five outbound SaaS citations resolves to something dead, gated, or broken, and a broken-link sweep earns its keep on those. Claim verification starts where that sweep finishes. The page loaded, the words rendered, and the question becomes whether any of those words state your claim.
Borrowed numbers concentrate the risk. In the same corpus, 65.5% of third-party claims are borrowed data, statistics lifted from someone else's research and re-cited. A number loses precision every time it is passed along. The source qualifies it, an intermediate post restates it, your draft rounds it, and the figure that reaches your paragraph sits near the page's figure without matching it.
Reading One Claim Against One Page
The check itself is small. Take the claim your draft makes, open the page it cites, and look for the line that states it. If the page carries a sentence that supports your figure, a sentence you could quote back word for word, the citation holds. If no such line exists, the citation is decoration.
You read for one claim only. The rest of the post can wait, and so can whatever the page links to next.
The test stays literal the whole way. Your sentence says 50%, the page says 48%, and the page fails your sentence even with an identical topic and a two-point gap. You never round, average, or reconcile the two numbers in your head to make them agree. Either the page states the value you wrote, or your citation points at a page that states something else.
This read runs in the opposite direction from a page audit. The Content Health Scanner takes a URL and pulls every statistic off the page, scoring each one for staleness. Auditing sweeps a published page for every claim on it; claim verification reads a single unpublished sentence against the one source it leans on. The sweep finds the claims you forgot you made, and the inward read settles the one you're about to ship.
By hand, the read is slow enough that deadlines kill it. I've taken the shortcut myself: click the link, see a real page on the right subject, move on. The habit works until the day the right subject arrives without your number, and by then the citation has shipped.
Everyone who publishes from AI drafts has a default version of this check. Which one do you actually run?
One of those habits reaches the page's actual words and asks whether they back the sentence. Each of the others can clear a citation that carries the wrong number.
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 inward read applies to a paragraph from a junior writer or an old draft of your own. AI drafts simply produce the mismatch at a volume nobody ever hand-wrote, which is what makes a seconds-long version of the check worth having.
AI Citation Checker in Practice
The AI Citation Checker below runs the manual read for you. Paste the source URL and the claim, and it fetches the live page, reads it against your sentence, and reports whether the page states what you wrote. Free, no account, 10 checks a day from one browser.
Its job is deliberately narrow: one claim, one page, one read, before you publish. It leaves your prose alone, and it stops watching the source the moment the verdict lands. Every check resolves to one of five outcomes.
When the page supports your claim, the checker returns the supporting sentence verbatim from the live page. Before showing you that line, the engine confirms it exists as a literal substring of the page, so a flattering quote can never be invented on your behalf. That quoted line is the whole basis of claim verification: proof you can go read with your own eyes.
Reworded and Wrong Page
Two of the five outcomes carry the argument, because both describe citations that pass every automated check you own.
"Reworded" is the first. Your draft says a tactic lifts conversion by 50%. The cited page is the right study on the right domain, and the link is healthy. Read against your sentence, the page says 48%. The page said 48% before you ever cited it; the 50% appeared somewhere between the source and your draft. The checker flags the claim as reworded and quotes the page's own line, so you can see the 48% sitting where your 50% should be. Nothing there would ever trip a broken-link sweep.
"Wrong page" is the second. The link resolves to a real, well-maintained page about the subject next door. Your sentence cites a figure on email open rates, and the page covers email deliverability. Same domain, healthy URL, plausible at a skim. When the support check comes back empty, the checker asks whether the page is even on your claim's subject, and a page that fails on a subject it never covers gets the wrong-page verdict instead of a plain miss. The citation looks careful and points at the wrong room.
In both cases the URL returns 200, the page renders, and the topic reads right. That combination is invisible to every automated check a publishing stack runs today, and it's the combination an AI draft produces most often. This is the hallucinated citation in its current form, and reading the words against the sentence is the one check that catches it.
Not Found and Could Not Verify
The third outcome is a page being straight with you. "Not found" means the link is healthy, the topic is right, and the specific claim appears nowhere on the page. Once I started reading sources against sentences in our own drafts, this became the verdict I saw most. A relevant page that happens to lack your exact number turns out to be a very common thing to cite.
The fourth is the checker being straight with you. "Could not verify" covers the pages the engine genuinely cannot read: dead, behind a login or paywall, or an empty shell that yields no text even after rendering. The checker reports that it could not see the page and tells you why, and the judgment returns to your hands. An honest gap beats a confident guess.
The fifth outcome, "supported," arrives with the quoted line that proves it. That citation you keep.
When a Verified Source Changes
The check runs once, on purpose. You verify today, the sentence holds, you publish. The page underneath keeps its own schedule, and the version that backed your claim may differ from the version a reader loads next quarter.
Watching for that is a separate job from the pre-publish read. A Monitored Page puts an external URL under a standing watch and tells you when what it says changes, and the waitlist extends that watch to the sources you cite. For the corpus-scale version of the question, whether an entire chain of citations still reaches its origin over time, see citation chain monitoring.
Together the two form content maintenance infrastructure. The pre-publish read gates what goes out, and the standing watch covers the years after the draft leaves your hands.
The Citation That Passes Every Check
The most dangerous citation in your back-catalog resolves to a credible page on exactly the right subject, carrying a number that was never there. A broken citation gets found: a sweep flags it, a reader hits the 404, somebody fixes it. The one that passes stays wrong for as long as nobody reads the words.
You're the last reader before all the others. A sentence claiming 50% over a page stating 48% doesn't get truer after publish, and nothing downstream of your editor will catch it. It travels into the board deck and the competitor's rebuttal, and it ends up under the byline of a reader who cited you in good faith and now carries your number with your name on it.