Is That Citation Real? How to Check a Source Before You Cite It

A citation looks the same whether it is solid or not. A title, a few authors, a year, a journal: the format carries authority all on its own, which is exactly why a bad one slips through. It might come from a chatbot, from another paper's reference list, or from a colleague's slides, and once it is in your bibliography it carries your name.

The fix is a short routine: four questions you ask of any source before you cite it. Most of them take a minute or two. Here is what each one catches, and why it is worth the time.

Why this is worth two minutes

AI tools are the newest source of bad citations⁠[1]. When researchers asked chatbots to find the papers for the same questions as 11 published systematic reviews, 39.6% of the references from GPT-3.5, 28.6% from GPT-4 and 91.4% from Bard were hallucinated⁠[2], meaning at least two of the title, first author and year were wrong. Newer models still do it. In a 2025 study on mental health topics, GPT-4o fabricated 35 of the 176 citations it produced⁠[3], with no identifiable source behind them, and 45.4% of the real ones contained errors.

But this is not only an AI problem, and it would be a mistake to check only what a chatbot hands you. A review of 28 studies of quotation accuracy in medical journals estimated that about a quarter of references were wrong or problematic⁠[4] as support for the statement they were cited for, and that roughly one in eight or nine did not support the claim at all. Those references were written by people.

The consequences can be public. In June 2023 a federal judge in New York ordered two lawyers and their law firm to pay a $5,000 penalty after they submitted court opinions that did not exist⁠[5], with fake quotes and citations created by ChatGPT, and kept standing by them after the court questioned whether they existed. The judge was careful to add that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." The failure was not using the tool. It was not checking.

1. Does it exist?

Search for the exact title in a scholarly database, such as Google Scholar, PubMed, Crossref, or your library's search, rather than a general web search. A real paper usually turns up in more than one of them, with the same authors, year, and journal you were given. If the title appears nowhere except where you found it, treat that as a warning.

If there is a DOI, resolve it at doi.org, and then look at where it lands, because a working DOI is not proof. In the same 2025 study, when GPT-4o supplied a DOI for a citation it had invented, 21 of the 33 DOIs worked but led to unrelated articles, and the rest did not work at all. Make sure the page you land on has the same title, authors, and year as the reference in front of you.

2. Is it still standing?

A paper can be real and still not be fit to cite as evidence, because it has been retracted. Retracted papers keep collecting citations long after the notice goes up. One study of biomedical papers found that of 13,252 passages citing a paper after its retraction, only 5.4% mentioned the retraction⁠[6]. In one striking case, a clinical trial report retracted in 2008 for falsifying data⁠[7] was still being cited positively eleven years later, with the retraction unmentioned in 96% of the citations the researchers could examine.

So check. The publisher's page for the article usually carries any retraction notice. PubMed labels retracted articles⁠[8] with the publication type "Retracted Publication" and links them to the notice. And since 2023, the Retraction Watch database, which tracks tens of thousands of retracted papers, has been owned by Crossref and made a public resource⁠[9]. If a paper carries real weight in your argument, check more than one of these.

3. Does it say what you are citing it for?

This is the check that is easiest to skip. A real paper that has never been retracted can still be cited for something it does not say, and that is the kind of error the quotation accuracy study above was counting. It is also easy to inherit: if you cite a paper because another author did, without opening it, you take on their reading of it, right or wrong.

So open the paper and find the passage. Read the sentence that supports your claim, and check the details that tend to drift in retelling: the population, the size of the effect, and whether the authors reported a finding or only speculated about one in the discussion. If you cannot find the passage, do not cite the paper for that claim.

4. Is it the version you read?

A preprint, a conference paper, and a journal article can share a title and still differ in their numbers. Cite the version you actually read, and if a preprint you rely on has since been published, check whether the published version changed the finding you are using.

Make checking the default

The best time to check a citation is the moment it enters your work⁠[10], because that is when the paper is easiest to open. Researcher+ is built around that habit: answers come only from the papers you upload, every claim cites the exact page, and clicking a citation opens the paper at that page and, where it can, highlights the supporting passage. Inside a paper, the References tab marks any work in its reference list that has been retracted.

Whatever tool you use, the rule is the same. A citation you have not opened is a guess. Four questions, a minute or two each, and every source in your bibliography is one you can stand behind.