AI Citation Checker: Are Your ChatGPT References Real?
AI tools such as ChatGPT, Gemini and Copilot often write references that look perfect and do not exist.
- Free, no sign-up, no limits
- Checks 5 free scholarly databases
- Nothing stored by LetRelay
Word (.docx), .bib, .ris and .txt files work. Each reference is looked up in free public databases; nothing is stored by LetRelay.
How to use it
Paste the AI's references
Copy the reference list from ChatGPT, Gemini, Copilot or Claude and paste it in. Any style.
See which are real
Each one is looked up. Invented ones come back not found; a DOI that does not exist is flagged.
Keep only real sources
Copy the verified list, or replace the invented ones with real papers before you cite them.
Paste them here and each one is looked up in Crossref, OpenAlex, PubMed and two other free databases: real ones come back verified, invented ones come back not found, and a real title given the wrong authors, year or DOI is shown with the true details.
How each reference is checked
Most reference checkers either only look at punctuation, or send your list to an AI that can itself make things up. This one looks every reference up in the published record, and only calls it found when a record agrees on both the title and the authors. The databases, all free and public, are asked in this order, cheapest and most exact first:
| Database | What it holds | Used for |
|---|---|---|
| doi.org | Every DOI ever registered, with any agency | Does the DOI in your reference exist at all? |
| Crossref | 188 million records: journal articles, books, chapters and conference papers with a DOI | The DOI's own record, and matching the whole reference as written |
| Europe PMC | 49 million records, including all of PubMed and PubMed Central, and preprints | Medicine and life sciences, also papers without a DOI |
| DataCite | arXiv, Zenodo, figshare, Dryad and university repositories | Preprints, datasets and theses |
| Open Library | Books, with every edition and printing | Books that have no DOI |
| OpenAlex | 318 million works from all of the above and more | Anything the others missed, and retraction flags |
When a record matches, the year, DOI, journal, volume and pages are compared with what you wrote. A different year is only called wrong when the journal, volume or pages confirm it is the same publication, and for books it is never called wrong, because editions and reprints carry their own years. Matched papers are checked for retractions with the Retraction Watch database (part of Crossref since 2023).
What the results mean
- Verified: a published record agrees with the reference on title, authors and year. Nothing to change.
- Found, with fixes: the work exists, but something in your reference is wrong (year, DOI, first author, journal, volume or pages). The correct value is shown next to yours, and the finished list uses it.
- Retracted: the work exists but has been retracted, withdrawn or given an expression of concern. Do not rely on it.
- Does not match: the title exists under other authors, or the DOI belongs to a different work. This is how AI-made references often look.
- Check this one: a close record was found but something important differs. Open it and say yes or no.
- Not found: no database has it. It may be invented, or a source they do not cover (an unpublished thesis, a report, a web page, a paper from the last few days). Search Google Scholar with one click before you keep it.
What it cannot do
No free database holds every source, so the tool never calls a reference fake: it says not found and tells you where to look. It cannot see whether a paper says what you cite it for; it proves the paper exists and is cited correctly, so read the part you rely on. Web pages, laws, standards and most reports are not in paper databases, so those are labelled rather than checked. Titles are put in sentence case for APA, Vancouver and IEEE; a program cannot always tell a name from an ordinary word, so check capitals in names of places, methods and software. Your own titles keep your capitals unless you ask otherwise.
How we tested it
Before release it was run against 56 references in APA, Harvard, Vancouver, IEEE, Chicago and MLA, from biology, medicine, psychology, economics, computer science and sociology: 40 real journal articles, conference papers, books, chapters and preprints; two well-known retracted papers; four real papers with one deliberate error each (a year, two DOIs, an author list); and ten references invented the way AI chatbots invent them. 38 of the 40 real works were found and verified. The other two: a WHO report, labelled as a report because paper databases do not list reports, and the 2014 conference paper that introduced GANs, which no free database lists under its published title (it came back “not found”, or “check this one” when OpenAlex was available, which is why not found never means fake). Both retractions were flagged, all four deliberate errors were caught, and none of the ten invented references was accepted. The formatting of every style was compared with the official CSL style files run by citeproc-js, the engine inside Zotero and Mendeley: 189 style and record pairs come out identical, and the rest differ only where we follow the style guide on purpose.
More free reference tools
The same engine, set up for a different job: Reference checker, Citation checker, DOI to citation generator, DOI to BibTeX converter, BibTeX to APA converter, APA to Vancouver converter, BibTeX to RIS converter and RIS to BibTeX converter. Writing a thesis? The dissertation word count breakdown splits your word limit by chapter, and the PhD Gantt chart maker plans the time.
Frequently asked questions
How do I check if ChatGPT references are real?
AI chatbots write references by predicting text, so they produce ones that look right but do not exist, or attach a real title to the wrong authors, year or DOI. Paste the list here: every reference is looked up, invented ones come back as not found, a DOI that does not exist or belongs to another paper is flagged, and a real title under the wrong authors is shown with its real authors.
Why does ChatGPT make up references?
A chatbot writes text that is likely, not text it has looked up. A reference is a pattern it has seen millions of times (names, a year, a title, a journal, numbers), so it can write one that fits the pattern without a real paper behind it. Newer chatbots that search the web make fewer, but still some, especially for page numbers, years and DOIs.
What are the signs of a fake reference?
A DOI that does not exist, or that opens a different paper; a real author with a title they never wrote; a journal that never published that volume; a title that sounds like a summary of your question. This tool checks the DOI with doi.org, and checks that the title and the authors belong together in the published record.
What does "not found" mean?
It means none of the databases has a work that matches both the title and the authors. That is how an invented reference looks, but some real sources are not in these databases: theses that were never put in a repository, government and NGO reports, web pages, very new papers (a few days old) and some regional journals. So the tool says "not found" and never "fake": search the title on Google Scholar or your library catalogue, and if you cannot find it there either, do not cite it.
Can I trust a reference marked verified?
Verified means a published record agrees with the reference on its title and authors, and on the year. It proves the work exists and is cited correctly. It does not prove the work says what the AI claimed it says: open it and read the part you cite.
How accurate is it?
We ran it on 56 test references in six styles (7 November 2026, live databases): 40 real works, two retracted papers, four real papers with one wrong year, DOI or author list, and ten invented the way AI tools invent them. 38 of the 40 real works were found and verified; the other two were a WHO report (labelled as a report, which paper databases do not list) and a 2014 conference paper that no free database lists under its published title. Both retractions were flagged, all four planted errors were caught, and none of the ten invented references was accepted as real. A record must agree on title and authors before it counts, so a match is not a guess.
Is it free and private?
Yes. There is no sign-up, no daily limit, no watermark and no paid tier. The databases it asks are free public services, called straight from your browser, so checking costs LetRelay nothing and there is nothing to charge for. Your list goes from your browser to the public databases (Crossref, OpenAlex, Europe PMC, DataCite, Open Library), one reference at a time, the same as typing each one into their search. LetRelay does not store it, and there is no account. Results are kept in your own browser for a week so checking the same list again is instant.