How to Check If a Citation Is Real (Step-by-Step)
Whether you're reviewing a manuscript, checking your own AI-assisted draft, or vetting a paper you found online, the question is the same: does this reference point to a real publication? Here's a complete workflow, from 30-second manual checks to verifying a hundred references in one pass.
Why bother checking at all
Fabricated citations used to be rare enough to be newsworthy. With LLM-assisted writing they're now routine: language models produce references that look exactly right — real journal, plausible title, well-known authors — for papers that don't exist. Reviewers, editors, and increasingly universities screen for them. If a fake reference is in your bibliography, the responsibility is yours, whatever tool drafted it.
The manual checks (one reference, ~2 minutes)
1. Resolve the DOI
If the citation has a DOI, go to doi.org/<the-doi>. Three outcomes:
- Resolves to the right paper — compare title, authors, and year anyway; a real DOI wrapped in wrong metadata is a common AI failure mode.
- Resolves to a different paper — the citation is wrong even though the DOI is real.
- Doesn't resolve — the DOI was never registered. Strong fabrication signal.
2. Search the exact title
Paste the title in quotes into Google Scholar, CrossRef (search.crossref.org), or OpenAlex. Real papers almost always surface. Watch for near-misses: a real paper with a similar-but-different title is the classic chimera — check the authors match too.
3. Check the authors and venue
Do the named authors actually publish in that journal? Does the volume/issue exist for that year? Is the page range plausible (a 3-page article cited as pages 140–190 is wrong)? Author-venue mismatches are where fabricated references usually crack.
4. Check preprint servers for very recent work
A 2026 paper may legitimately exist only on arXiv, bioRxiv, or SSRN so far. Search there before concluding fabrication — but confirm the preprint title and author list match the citation exactly.
Red flags that justify a closer look
- Round, tidy metadata: volume 1, page 1, exactly year 2000.
- A very famous author on a suspiciously generic title.
- DOIs that follow the publisher's format but with sequential-looking suffixes.
- Journals you can't find in any index — check the ISSN portal.
- The paper's claims in the citing text don't match anything the real paper says.
The automated path (whole bibliography, ~2 minutes)
Manual checking doesn't scale past a handful of references. The automated equivalent of everything above:
- Paste the reference list — or upload the PDF — into the EvalCite checker.
- Each reference is parsed and looked up concurrently in 8 scholarly databases (CrossRef, OpenAlex, Semantic Scholar, DBLP, arXiv, OpenLibrary, Google Books, plus IEEE Xplore and Scopus with keys).
- Read the verdicts: verified means a real record matches; mismatch means a similar paper exists but specific fields disagree (the diff shows which); not found means nothing close exists anywhere — treat as fabricated until proven otherwise with the manual checks above.
- Download the corrected .bib to rebuild verified entries from the authoritative records.
One caution in both directions: an automated not found is not proof of fabrication (very recent papers and non-scholarly sources can be absent from every index), and an automated verified doesn't check whether the citation actually supports the sentence it's attached to. Verification answers "does this paper exist with these details" — relevance is still on you.
The 60-second checklist
- DOI resolves to the same paper? ☐
- Exact title found in a scholarly index? ☐
- Authors real and publishing in that venue? ☐
- Year, volume, pages consistent with the journal's archive? ☐
- If very recent: preprint exists with matching title/authors? ☐
Anything unchecked is worth two minutes before it costs you a retraction notice.