CiteOwl

Blog/Getting citations right

How to check if a citation is real (in about five minutes)

The CiteOwl Team· Updated 20 September 2026· 10 min read

Six checks catch almost every fake, and every tool they need is free. A fabricated reference is not a broken one. It has a plausible author, a real journal, a well-formed DOI, and it resolves to nothing. Search the exact title, resolve the DOI, read the page it opens, confirm the authors exist, then check the details against the publisher's own record. Run the list through the free citation checker first, then open what survives.

The six checks, in order

The order matters more than it looks. Each check is cheaper than the one after it and closes a different kind of entry, so a reference that fails the title search never costs you the four minutes the rest would take. Work down the column and stop at the first failure.

Search the exact title in quotes

Copy the title and paste it into Google Scholar or plain Google with quotation marks around it. The quotation marks matter. Without them the search engine helpfully finds pages that are merely about your topic, which is the noise a fabricated title hides in. A real paper almost always turns up as the first hit, and this single check exposes most fabricated references in seconds.

If Scholar comes up empty, try one more index before you conclude anything. OpenAlex is a free, open catalogue of scholarly works, and its own API reported 327 million of them when we checked on 20 September 2026. A title missing from Scholar and from a catalogue that size is not a paper that is hard to find.

Before you delete it, rule out the handful of real papers that genuinely resist a title search. A work behind a paywall still has a landing page and a resolving DOI even when you cannot read the PDF. A preprint may sit outside the indexes you searched, and an older work or a book chapter may have no DOI at all. The test that separates real-but-buried from fabricated is simple. A genuine paper leaves a trail somewhere independent of the text that handed it to you. Find one trace and it is probably real. Find none and you have your answer.

Resolve the DOI, then read the page it opens

If the citation includes a DOI, a string like 10.1037/abc1234, paste it after https://doi.org/ in your browser. A real one takes you straight to the article's official page. A made-up one returns "DOI Not Found". Chatbots are good at producing DOIs that look correctly formatted, so the format proving nothing is exactly the point.

One trap here catches almost everyone, and it is why this step has two halves. A DOI that resolves is not the same as a DOI that resolves to the paper your citation describes. When researchers examined GPT-4o's fabricated citations in a 2025 JMIR Mental Health study, 64% of the DOIs attached to fake references loaded real but unrelated papers. In our own test of 218 model-written DOIs, the best model produced more DOIs that opened the wrong paper than dead ones. So read the title and the author list on the page you land on and confirm they are the ones in your reference. If they are not, the check has failed, not passed, and the DOI links to the wrong paper is the repair. If it will not resolve at all, five different things can cause that and only one of them means the paper is fake.

Find it on the publisher's own site

Locate the article on the journal's own website, not just on a results page or a PDF mirror. Real articles have a landing page on the publisher's domain with an abstract, the full author list and publication details. If the only trace of a paper is the citation itself, it probably is not a paper.

While you are there, glance at the journal. A real article in a venue nobody has heard of, with a homepage promising review in 72 hours and a fee schedule on the front page, is a different problem from a fabricated citation and still a weak thing to build an argument on. How to tell if a journal article is fake covers the venue check, which asks a register rather than the website, since a forger controls the website completely.

Cross-check the details against the record

Now compare what the citation claims against what you found. Authors, year, journal, volume, issue, page numbers. The common failure is not a wholly invented paper but a mangled one. Real authors stapled to a title they never wrote, or a real article credited to the wrong year and journal. The rule of thumb librarians use is that if two or more core fields disagree, the reference is unreliable even though some of it checks out. To pull the authoritative version in one go, paste the title or DOI into the free Crossref metadata search, which returns the authors, year and journal from the publisher's own deposit.

Do not skip this because the paper turned up. A chatbot reference that exists is still usually wrong somewhere in the details. One analysis of 115 ChatGPT references found that of the half pointing at real papers, almost all carried at least one error in the authors, year, volume or pages, leaving 7% both real and accurate.

There is one failure none of the six checks can see, and it is the one a marker is most likely to find. The paper is real, the DOI resolves, the authors are right, and it does not say what you cited it for. Existence is not support. Open the source, find the sentence that actually carries your claim, and keep it beside the citation. ChatGPT cited a real paper that says something else has the 90-second version, and if the borrowed words are inside quotation marks, check the quote too.

Confirm the authors actually exist

Search the lead author's name on Google Scholar and look them up on ORCID, the registry researchers use to list their own work. A genuine researcher leaves a trail. Other papers, an ORCID record, a university profile, a Scholar page in roughly the same field. An author who appears nowhere except in this one citation is a warning sign.

This check also settles the most convincing kind of fake, the one assembled from real parts. Search the author's name together with two or three words from the title. If the author is real and active and has never published anything close to it, the paper as described does not exist even though every piece of it does.

Check it hasn't been retracted

A real but retracted paper is its own problem, because citing it as sound evidence undermines the argument you built on it. The fastest check is the free Retraction Watch Database, which Crossref acquired in September 2023 and now distributes without a subscription, updated every working day. Paste in the title or the DOI.

This matters more than it used to. Nature reported that more than 10,000 papers were retracted in 2023, a record, so a citation that was sound a year ago may not be now. A clean result is not a clearance either, since a notice can exist without reaching the index you searched. Is the paper I cited retracted has the four places to look and how to read a notice when you find one.

When an entry fails

Do not try to salvage it by reformatting or guessing the right version. That buries the problem under tidier packaging. Delete it, then go back to the claim it was supporting and find a source that genuinely makes it, using your library's databases or Google Scholar directly.

Work claim by claim rather than citation by citation. Swapping one unverified reference for another plausible-looking one keeps the exact risk you were trying to remove. For each factual sentence, ask what would support it, find a real paper that does, read enough of it to be sure, and attach that. Sometimes the claim survives with a solid source behind it. Sometimes you find the point was overstated and needs softening. If you search honestly and find nothing at all, that is useful information, because the claim itself may not hold, which is far better to learn now than in feedback.

A few signs should stop you before you finish the six. No result for the exact title. A DOI that will not resolve. A paper that exists only on the page that cited it. Real authors attached to an unfamiliar title. Or a suspiciously perfect fit, where the paper says exactly what you needed it to say, in a journal you cannot quite find. Any one of those is reason enough to set the reference aside and start again on the claim.

Doing this for a whole reference list

Six checks are quick for a handful of sources. For a bibliography of forty, going one by one gets tedious, and tedium is where mistakes slip in. Two habits keep it manageable. Check a reference the moment you add it, because it costs seconds then and an evening at midnight. And run the list through a checker first so you only spend your own time on the entries that survived it. Our citation checker is free, takes a pasted list, and resolves each entry against CrossRef and OpenAlex, and the DOI lookup prints the registered record for a single identifier.

Be clear about what an automated pass can and cannot do. It confirms what is broken. It does not hand you a real source to use instead, and nothing automated can tell you whether a real paper supports your sentence. That last judgment is yours, on every entry, forever.

Which is why the durable fix is upstream. A reference you never have to check is one that was looked up before it was written, which is what a writing tool that retrieves papers first and cites only what it opened is for, and what separates the options in how CiteOwl compares. The settings and prompts that cut the fabrication rate in a general chatbot are in how to get ChatGPT to cite real sources, and the mechanism underneath all of it is why AI makes up citations. Aim for the plain thing at the end of all this. A draft whose citations resolve, and whose edits you read before you kept them.