Free citation checker
Paste up to 50 references and every one is looked up in CrossRef and OpenAlex, in seconds. Free, no account, nothing stored. Fake references, dead DOIs and records that disagree with what you wrote are named separately.
Your reference list
What came back
Nothing checked yet. The rows arrive when a list does.
How the check works
Your list is split into individual references, and each one takes the shortest honest route to an answer. If a reference carries a DOI, that DOI goes to CrossRef, the registry publishers deposit their metadata with. CrossRef either hands back the registered record or it hands back nothing, and nothing is already a finding. Everything else, references with no DOI and references whose DOI came back empty, is searched in OpenAlex by title and author.
Then comes the part most checkers skip. Finding a record is not the same as the record agreeing with you, so the title, the author list and the year in your reference are compared against the ones the publisher registered. A citation can point at a real, resolvable, perfectly respectable paper and still be wrong about which paper it is.
Both phases run four lookups at a time, so a list finishes in seconds rather than in one round trip per entry. We timed our own checker against the live registries on 11 August 2026: twenty references took 3.3 seconds, and ten references that all carried a DOI took 1.3, since a DOI is one CrossRef call while a bare title has to be searched. Your own connection adds to that. The cap is 50 references in a paste, which covers a normal essay bibliography and most thesis chapters; a longer list goes in two goes.
Nothing is stored. There is no account, no sign-in and no card, and your list is not kept once the answer comes back. Both registries behind it are free public sources, so there is nothing here to charge you for.
What each verdict means
Verified. The work was found in CrossRef or OpenAlex, and what you wrote lines up with the registered record. Nothing to do.
Record does not match. A record exists, but it disagrees with your reference, and the fields that disagree are named for you. Two shapes turn up. The DOI resolves, but to a different paper than the one your citation describes, which is what a fabricated reference built around a borrowed identifier looks like. Or the paper is real and the DOI bolted onto it leads nowhere. Either way the reference as written is wrong, even though clicking the link opens something.
Not found. No record in either index. If the DOI is dead as well, that pairing is the strongest fabrication signal you can get without reading the paper, and it is worth checking the rest of the list, because invented references arrive in clusters rather than alone.
Dead DOIs are counted on their own, separately from the three verdicts, because a DOI that is perfectly formatted and registered to nothing is the failure that fools people hardest. There is no checksum in a DOI and no rule that makes an invented one look wrong. A DOI is real only because somebody registered it, which is a whole diagnosis of its own.
What this cannot tell you
It reads two indexes. Between them CrossRef and OpenAlex cover journal articles well and everything else less well, so a book, a book chapter, a thesis, a conference paper, a government report or an article in a small regional journal can be entirely legitimate and still come back not found. So can a paper published in the last few weeks, before its registration has propagated. Not found means the check could not confirm it, not that it is fake. Those are the ones to look up by hand, and the manual routine takes about a minute.
Here is one from our own test run, because it is the failure worth knowing about. We fed the checker ten references on 11 August 2026, one of which was Attention is all you need, cited the way it usually is, as a NeurIPS proceedings paper with no DOI. It came back not found. The paper is one of the most cited in machine learning; the reference simply had no identifier to resolve and the conference record did not match well enough on title and author alone. Meanwhile the one reference in the list we had invented, complete with a plausible DOI, was correctly flagged as unresolved. That is the shape of this tool's errors: it is good at catching an identifier that leads nowhere and weaker on real work that never got one.
It also cannot tell you whether a real paper supports the sentence you attached it to. That is the more common failure, no lookup catches it, and the only fix is opening the paper and reading the passage. It makes no judgment about quality either: whether the venue is a predatory or hijacked journal, or whether the paper has since been retracted. Those are separate checks, and both matter before you submit.
Why AI reference lists fail this check
A language model produces a reference the same way it produces any other sentence, by predicting what usually comes next. Author names, journal titles and DOI suffixes all follow strong patterns, so a model can assemble something that looks exactly right and corresponds to nothing. It is not lying and it has no registry to consult, which is the mechanism behind fabricated citations and the reason asking the chatbot to double check its own work never helps. Run the list through a registry instead. That is what this page does.