How to check whether a reference is real

A reference can look perfect and still point at nothing. The authors are real people, the journal exists, the title sounds like the field, the year is plausible, and there is even a DOI. Then a marker looks it up, and no such paper was ever published.

This is what language models do when they are asked for sources. They do not look anything up. They write text that has the shape of a reference, and sometimes the shape is all there is. Walters and Wilder (2023) asked ChatGPT for 84 short literature reviews and checked every one of the 636 references it gave: GPT-3.5 had fabricated 55% of its references, GPT-4 18% of its own. The newer model did much better, and the problem had not gone away. Why AI tools invent references explains the mechanism behind invented references.

A reference list with one of these in it is a serious problem, whoever typed it, because a reference is a claim: this exists, and I read it. Checking is not hard. It is just tedious, which is why it gets skipped. Below is the check by hand, one reference at a time, and then the quicker route: EdCitation's free Verify references runs the same lookups against the publisher's record for a whole list at once.

How do I check one reference by hand?

Follow the DOI if there is one, search the exact title if there is not, and compare every detail with the record you find. In order:

1. Start with the DOI, if there is one. Put it after https://doi.org/ in your browser. A real DOI lands on the publisher's page for that paper. Then compare: is the title the same? Are the authors the same? A common failure is a DOI that is real but belongs to a different paper. The model borrowed it. To see that case at a glance, paste the DOI into EdCitation's Cite a source: it builds the reference from the record the DOI points at, so the title it returns is the paper that DOI really belongs to. If DOIs are new to you, read what a DOI is.
2. No DOI? Search the exact title. Put the full title in quotation marks in Google Scholar, or search it in Crossref or PubMed. Crossref's search is free and needs no registration; its status page showed 187,980,352 records on 21 September 2026. PubMed holds more than 40 million citations in biomedicine and health, and links to the full text where it can. A published paper's exact title nearly always turns up. If you get nothing, or only near misses with different words, be suspicious.
3. Check the journal's own pages. A reference gives a volume, an issue and page numbers. Open that issue of the journal and look at its table of contents. If the pages belong to another article, the reference was built around a real journal.
4. Check the first author. Open their publication list on the journal's site, on PubMed or on their university page. A real researcher with a title they never wrote is the classic mix.
5. Read the part you are citing. A genuine paper that does not say what you cite it for is a broken reference all the same, and it is the failure a marker notices second.

The convincing mix: real authors, real journal, a title nobody wrote

The hardest fakes are assembled from real parts, so checking one part is not checking the reference. Gravel et al. (2023) asked ChatGPT 20 medical questions and found 41 of its 59 references fabricated, and yet 56 of those 59 references named authors who really had published on a related topic, and 71% of the fabricated ones were attributed to a known medical journal with a year, volume and page range that fitted that journal but belonged to a different article.

That is why the last comparison in the list above matters more than the first. Every part checks out on its own. Only the paper does not exist.

What if the DOI does not resolve?

A DOI that fails is a reason to look further, not a verdict on its own. Work through the ordinary causes before you conclude that a paper was invented.

• A typo. DOIs are copied by hand more often than anyone admits. Search the title before you condemn the reference.
• The wrong version. Preprints and published articles have separate DOIs. A reference can carry the preprint's DOI and the journal's details, or the reverse.
• A newly registered record. A DOI registered in the last day or two can fail to resolve for a short time.
• A borrowed DOI. The worst case, and the most common in AI-written lists: the DOI resolves perfectly, to something else. Linardon et al. (2025) found that of the 33 fabricated references that carried a DOI, 21 had valid DOIs leading to unrelated articles and 12 did not work at all.
• Reference rot, in older work. Klein et al. (2014) studied over 3.5 million articles published from 1997 to 2012 and found that one article in five was affected by reference rot: a cited web address that has stopped resolving, or that now shows something other than what was cited. A dead link in a 2005 reference list is ordinary. A dead DOI on a paper said to have appeared last year is not.

The DOI Foundation is clear that a DOI is a persistent identifier that redirects you to the thing it points at, and that persistence rests on "organizations, not technology". Nothing about it certifies the reference typed beside it.

How do I check a book, a report or a web page?

Check it where it was published, because the indexes that cover journal articles do not cover these well. Each kind has its own record.

• A book or a book chapter. Search the ISBN in Open Library or WorldCat, or your library's catalogue, and confirm the edition and the year as well as the title. Chapters are the weak point of AI-written lists: Walters and Wilder (2023) found GPT-4 had invented 70% of the book chapters it cited, while its rate across all its references was 18%.
• A report or a policy document. Go to the organisation's own website and find the publication. Check the title, the date and the publisher. Many are not in Crossref at all.
• A thesis. Look in the university's repository or the national thesis service.
• A conference paper. Check the conference's own proceedings, which are often outside the main indexes.
• A web page. Open it. If it has gone, look for an archived copy and cite the archived version with the date you saw it.

What does "not found" mean, and what does it not mean?

"Not found" means only that an index has no record, and whether that matters depends on whether the index ought to have one. Not every real source is in every index. Reports from government bodies, older papers from small journals, theses, and work published outside the large publishers can be missing from Crossref and PubMed and still be perfectly real.

Two results that must never be confused

• No record found in any index that ought to have it. For a recent article said to be in a well-known journal, with a DOI that leads nowhere and a title that Crossref, PubMed and Google Scholar all fail to match, that is strong evidence that the paper does not exist.
• Could not be checked. The source is of a kind the indexes do not cover. Here you verify at the source itself: the organisation's own website, the library's catalogue, the repository.

A good checker tells you which of the two it is. A bad one reports both as "not found", which either frightens you about a real report or reassures you about a fake paper.

Google Scholar deserves its own warning. It does not publish a list of what it covers, so a blank result there is weaker evidence than a blank result in Crossref or PubMed, which do say what they index.

How do I check whether a paper has been retracted?

Open the publisher's page for the paper, where a retraction notice is shown, and look the paper up in PubMed if it is biomedical. A paper can be real and still be one you should not rely on. Journals withdraw papers, for honest error and for misconduct, and the paper usually stays online with a notice on it. PubMed does not remove a retracted paper: it labels it "Retracted Publication" and links it to the notice. For articles indexed in MEDLINE, the ICMJE names PubMed as the authoritative place to learn of a retraction, and it puts the duty on authors to make sure no reference in their list is a retracted article, unless the retraction is what they are writing about. EdCitation screens for this as well: Cite a source checks for a retraction while it builds the reference, and Verify references flags a retracted paper anywhere in a list. The full method is in what a retraction is, and how to check your reference list for one.

How do I check a whole reference list at once?

Hand it to a tool that looks references up in the publisher's record instead of guessing at them. Doing all of this by hand for thirty references takes an evening. It is the kind of work a machine should do, provided it looks things up.

EdCitation's Verify references is the best tool for this job, for the reason this guide keeps coming back to: it never writes a reference, it looks each entry up in the publisher's record, and it compares the paper it finds with the whole entry, not with one part of it. Gravel et al. (2023) found real, relevant authors on 56 of the 59 references ChatGPT gave them, although 41 of the 59 did not exist, so a check that stops at "the authors are real" waves nearly every fake through. A chatbot asked to check its own list is writing again, not looking up, and cannot close that gap. Upload your paper or paste your reference list, and every entry gets one of three results: verified; "check this", meaning the match is doubtful; or not found. Retracted papers are flagged. A source the indexes do not cover is reported as "could not check", which is a different thing from "not found". Upload the whole paper rather than the list and it also pairs the citations in your text with the entries in your list, in both directions. There is no charge and no account.

The paid plans build on the same lookup. References from a file, part of Pro at $8 a month, checks every reference in a file and matches every in-text citation to the list; Max, at $24 a month, adds Theoretics QA and the Library. Both are on the pricing page, and paying for a year at once saves 25%.

Crossref also runs a Simple Text Query form, which takes a pasted reference list of up to 1,000 references and tries to return one DOI for each. It is a useful second opinion: a reference that Crossref cannot match to any DOI is worth looking at by hand.

What should I do if a reference does not check out?

Work out what happened before you change anything. Do not delete it and hope.

• If you read a real source and the details are wrong, find the source again and cite it from its own record. EdCitation's Cite a source rebuilds it from the record, given its DOI, title, PubMed ID or ISBN, in any of six styles.
• If the reference came from an AI tool and you never read it, the claim it was holding up now has no support. Find a real source for the claim (EdCitation's search takes the sentence itself and returns published work that could support it), read it, and cite that. If you cannot find one, the claim may not be true, and that is worth knowing before your marker tells you.
• If it is a wrong page range or a wrong year on a paper that exists, fix the detail. That is a referencing error, not a fabricated source, and the two are treated very differently.

Quick questions

Can a fake reference have a real DOI?

Yes. A fabricated reference often carries a DOI that exists but belongs to a different paper. In one 2025 test, 21 of the 33 DOIs attached to fabricated references resolved to unrelated articles. Follow the DOI and compare the title and the authors on the publisher's page with the reference in front of you.

Does a reference being missing from Google Scholar mean it is fake?

Not by itself. Google Scholar does not publish a list of what it covers, and reports, theses and older articles can be missing from any index. For a recent article in a well-known journal, though, a missing record in Crossref, PubMed and Google Scholar together is strong evidence that the article does not exist.

How long does it take to check a reference list by hand?

Allow a few minutes for each reference: following the DOI or searching the title, then comparing the authors, the year, the journal and the pages. A list of thirty takes an evening. EdCitation's Verify references runs those lookups for every entry in one pass, free, and leaves you only the ones marked "check this" or not found to open by hand.

Do markers and editors really check references?

Assume that someone will: a single click on a DOI is enough to expose a reference that does not exist. For journals, the International Committee of Medical Journal Editors puts the accuracy of references on the authors, who should be able to vouch that every reference backs the statement it sits beside.

Is it enough that the authors and the journal are real?

No. The most convincing fabricated references are built from real authors and a real journal with a title nobody wrote. Only finding the paper itself, with the same title, authors, year and pages, shows that the reference is real.

References

• Crossref. (n.d.). Simple Text Query. https://www.crossref.org/documentation/retrieve-metadata/simple-text-query/
• DOI Foundation. (n.d.). What is a DOI? https://www.doi.org/the-identifier/what-is-a-doi/
• Gravel, J., D'Amours-Gravel, M., & Osmanlliu, E. (2023). Learning to fake it: Limited responses and fabricated references provided by ChatGPT for medical questions. Mayo Clinic Proceedings: Digital Health, 1(3), 226-234. https://doi.org/10.1016/j.mcpdig.2023.05.004
• International Committee of Medical Journal Editors. (n.d.). Preparing a manuscript for submission to a medical journal: References. https://www.icmje.org/recommendations/browse/manuscript-preparation/preparing-for-submission.html
• Klein, M., Van de Sompel, H., Sanderson, R., Shankar, H., Balakireva, L., Zhou, K., & Tobin, R. (2014). Scholarly context not found: One in five articles suffers from reference rot. PLOS ONE, 9(12), Article e115253. https://doi.org/10.1371/journal.pone.0115253
• Linardon, J., Jarman, H. K., McClure, Z., Anderson, C., Liu, C., & Messer, M. (2025). Influence of topic familiarity and prompt specificity on citation fabrication in mental health research using large language models. JMIR Mental Health, 12, Article e80371. https://doi.org/10.2196/80371
• National Library of Medicine. (n.d.-a). About PubMed. https://pubmed.ncbi.nlm.nih.gov/about/
• National Library of Medicine. (n.d.-b). Errata, retractions, and other linked citations in PubMed. https://www.nlm.nih.gov/bsd/policy/errata.html
• Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5