Best Tools to Check If Your References Are Real (2026): What Each One Actually Verifies

Reference checking means confirming two separate things about every entry in a bibliography: that the work exists as described, and that the details recorded for it are correct. A fabricated citation describes a paper that was never published. An erroneous citation points at a real paper but gets the author, year, journal, volume or page numbers wrong. Automated tools are good at the first problem and patchier at the second.

One declaration before the list: I built one of the four, so weigh what I say about it accordingly. Every row below sticks to what each tool’s own pages state, including ours.

What is the fastest way to check whether a reference is real?

Paste the reference list into a tool that matches each entry against Crossref, which is the registry publishers deposit DOIs into. Crossref’s Simple Text Query accepts up to 1,000 references per submission and returns the DOI it matched to each one, and Crossref’s instructions tell you to "click on the link that is returned with the reference" to check that match. Litrevu’s reference checker takes up to 30 references per check, needs no account, and checks each entry against Crossref, OpenAlex and doi.org in one pass.

A reference that comes back with no match is not automatically fake. Book chapters, theses, conference papers and older articles often have no DOI at all, so a no match result on those is a signal to search by hand, not a verdict. A reference that comes back matched to a different paper than the one you typed is the result worth slowing down for, because that usually means the author, year or title in your bibliography is wrong.

Which tools actually check a reference list, and what does each one verify?

Five free tools cover the reference checking job between them, and they do not overlap as much as their descriptions suggest. Each row below reflects what that tool’s own documentation states, read in October 2026.

Scite sits outside that table on purpose. Scite is strongest at citation context, meaning whether later published papers support or contrast a claim, which is a different question from whether your bibliography is accurate. It also has no free plan, where the five above are free, so it belongs in the next section rather than this one.

Why do real looking references turn out not to exist?

Because a language model producing a citation from memory is assembling a plausible string, not retrieving a record. Walters and Wilder measured this directly in Scientific Reports (2023, volume 13, article 14045, doi 10.1038/s41598-023-41032-5), collecting the 636 bibliographic citations found in 84 short literature reviews across 42 topics, and reporting that "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated".

Two cautions about carrying those percentages into 2026. Walters and Wilder tested GPT-3.5 and GPT-4 in 2023, and the models students use now are different, so the rates are not current rates. The finding that has held up is the one about appearance: a fabricated citation is formatted exactly as well as a real one, so nothing in the look of a reference list tells you which entries will survive a check.

Is it enough to check that the DOI resolves?

No. A resolving DOI proves the paper exists, and says nothing about whether your entry describes it correctly. Walters and Wilder, examining 636 citations from 84 AI generated literature reviews in Scientific Reports (2023, volume 13, article 14045), found that among the citations pointing at genuine works, "43% of the real (non-fabricated) GPT-3.5 citations but just 24% of the real GPT-4 citations include substantive citation errors". A reference can resolve to a real DOI and still carry the wrong year, the wrong journal or the wrong author order. Worse, a real DOI can be attached to the wrong paper altogether, which is why our checker compares the record the DOI points at with the title you wrote, rather than stopping at "the DOI works".

The standard that journals apply goes further than existence. The ICMJE Recommendations state that "Authors are responsible for citing references accurately and should be able to attest that the references cited support the associated statement", and that "To minimize bibliographic citation errors, references should be verified using either a bibliographic source, such as PubMed, or original sources". Attesting that a reference supports your statement is not something a DOI lookup can do for you. It means opening the paper.

How do I check whether something I cited has been retracted?

Check it against the Retraction Watch database, which Crossref now carries and which is "updated every working day by Retraction Watch" according to Crossref’s own documentation. Crossref exposes it two ways: a REST API filter, documented as api.crossref.org/v1/works?filter=update-type:retraction, and a CSV file published in a git repository on GitLab and updated once per working day.

If your reading sits in Zotero already, Zotero does this check for you. Zotero has flagged retracted items since version 5.0.67, released on 14 June 2019, using Retraction Watch data. Zotero flags a matching item in the items list, shows the retraction details in the item pane, and warns you if you try to cite a flagged item through the word processor plugin. The catch is identifier coverage: Zotero’s 2019 announcement limited the feature to items with a DOI or PMID, which it put at about three quarters of Retraction Watch data, so a saved item with neither may not be flagged. Litrevu’s reference checker reads the same Retraction Watch data through Crossref, and surfaces retractions and expressions of concern next to the reference, which covers entries you never put in a reference manager. There is a fuller walkthrough in the Litrevu guide on how to check whether a paper you cited has been retracted.

Citing a retracted paper is treated as an error you are responsible for, not an accident. The ICMJE Recommendations put it plainly: "Authors are responsible for checking that none of the references cite retracted articles except in the context of referring to the retraction."

What do universities expect you to have checked?

They expect you to have verified every reference yourself, and some of them say so in detail. The University of Pretoria’s Artificial Intelligence Guidelines tell students that "AI frequently generates plausible-looking citations to non-existent papers. Never cite a source without verifying its existence and relevance", and set out the method: "For every reference AI provides, search UP Library Catalogue, Google Scholar, or Web of Science using the complete citation (author, year, title)." That is South African guidance, from the UP library’s own guide.

In the United Kingdom, the University of Oxford’s guidance on safe and responsible use of GenAI is shorter and broader. It warns that "all tools can generate outputs that contain inaccuracies and fabrications", and that "You should always cross-check AI generated outputs against established sources to verify accuracy and identify erroneous information." Rules differ by institution and by country, so the version that matters for your submission is your own faculty’s, and the safest assumption while you work is the stricter of the two: verify the existence of every reference, and check the detail of every reference against the source document.

Can an AI research tool check its own citations?

Partly, and the distinction is in where the tool gets its citations from. A tool that searches a real index returns records that exist, because the record came out of a corpus rather than out of a model’s memory. Elicit searches "more than 138 million papers" on its free Basic plan, Consensus lets you "Keep using basic paper search for free", and Research Rabbit’s free plan covers "Unlimited searches across 310+ million articles", all as those pricing pages read in October 2026. The existence problem mostly disappears with tools like those. What does not disappear is the sentence level question, which is whether the claim written next to the citation is the claim the paper makes.

Litrevu is an AI literature review assistant that turns the papers a researcher has already gathered into a cited first draft. Its citations point at papers in the researcher’s own library, and each one opens to the passage it rests on, which makes the sentence level check possible instead of theoretical: you open the passage and read the line. Litrevu checks that every citation points to a passage it actually read, and that is not the same as the passage fully supporting the sentence, so the read is still yours to do. It is also a narrower job than searching the literature, and Elicit, Consensus, Scite and Research Rabbit each beat Litrevu at their own part of the work. I would rather say that here than have you find it out halfway through a chapter.

Paid plans change often, so each row links to the vendor’s own plans page rather than carrying a figure that may be stale by the time you read this, and the Litrevu guide to the cheapest literature review software for postgraduate students covers the cost side.

What does a checked reference list look like?

A checked list is one where every entry carries a status, so you can see at a glance which entries need you to open the paper. Litrevu’s reference checker runs four checks per entry, on up to 30 entries per check: whether the work exists, whether the year, first author, volume, first page and DOI agree with the record, whether the DOI belongs to that paper, and whether a retraction or expression of concern is on record. The block below is a format illustration, not a real result: it uses the checker’s own status labels, the tokens stand in for whatever your own list contains, and no author, title, year or finding in it is real.

Notice the wording on the fourth line. When we built the checker we made it say "We could not find this" and never "fake", because books, theses, reports and some regional journals are missing from these indexes, and a real source should not be called invented just because an index lacks it.

Statuses are not verdicts on your writing. A "Found and matches" status says the paper exists and the details line up, and nothing more. The sentence you wrote next to that reference is still unverified until you read the paper, which is the check the Litrevu guide on how to spot fake AI citations before you submit sets out by hand.

How long does checking a 60 item reference list actually take?

Two passes, and only one of them is quick. The automated pass is a single sitting: Crossref’s Simple Text Query takes a 60 item list in one submission, since its stated limit is 1,000 references per submission, and Litrevu’s reference checker needs two goes at its limit of 30 references per check.

The manual pass is everything the automated pass hands back. Searching one flagged reference the way the University of Pretoria’s Artificial Intelligence Guidelines describe, by its complete citation of author, year and title, in a library catalogue, Google Scholar or Web of Science, is slower work, one reference at a time. Then there are the claims: every sentence in your review that carries a citation needs the cited paper open beside it at least once, which is what the ICMJE Recommendations mean by being able to "attest that the references cited support the associated statement". No checker shortens that part. Formatting the verified entries is a separate job again, covered in the Litrevu guides to Harvard referencing and APA citation format.

Frequently asked questions