Paste a citation. The verifier resolves the identifier, compares the resolved title to the cited title, and flags the dominant fabrication pattern documented by Topaz et al. (Lancet 2026): a real, resolvable identifier paired with an invented title.
8 identifier types · 4 verdict outcomes · Field-level diff · Free, no signup
Paste the cited title and one identifier. Optional refinements tighten the match but never gate the verdict on their own. Calls POST /api/verify - no authentication required.
Or load one of the three Topaz et al. (Lancet 2026) illustrative cases - each prefills the form with a real, resolvable identifier paired with the fabricated title as published in Supplementary Appendix 2. Expected verdict for all three: mismatch.
This is a syntactic check: identifier resolution plus bibliographic field agreement (title, first author, year, container). It catches fabricated citations - real DOI + invented title, made-up identifiers, wrong identifier for a real paper. It does not verify that the resolved paper actually supports the claim you are citing it for. Claim-vs-content checking is a separate problem and is not solved here.
ambiguous) from real DOI + invented title (verdict: mismatch).Confidence is reported as high, medium, or low. Low-confidence mismatches are the bucket the opt-in LLM screen rescues - informal abbreviations, paraphrased titles - behind paid authentication.
One identifier required. Optional refinements - first-author family name, year, journal - tighten the match but never gate the verdict on their own.
This page calls POST /api/verify behind the scenes. If you want to script verification or run it inside an AI coding workflow, two other surfaces expose the same backend:
verifyCitation tool ships with the Scholar Sidekick MCP, callable from Claude Desktop, Cursor, or any MCP-aware clientREST API and MCP usage above the anonymous tier is metered on RapidAPI (free tier available; paid plans scale up). The web verifier on this page stays free at the anonymous tier.
Switch to Batch above to paste or upload a .bib / .ris / .csl-json file and verify up to 10 citations at once from one paste. For larger batches, script directly against POST /api/verify.
Three things, in order. (1) It resolves the identifier you supplied (DOI, PMID, PMCID, arXiv, ISBN, ISSN, ADS bibcode, or WHO IRIS URL) against the appropriate registry. (2) It compares the claimed title against the resolved title field-by-field, with similarity scores per field. (3) If the resolved title and the claimed title disagree, it title-searches the claim across Crossref, PubMed, and OpenAlex to see whether the cited title exists under a different identifier - distinguishing 'wrong identifier' (ambiguous) from 'fabricated title' (mismatch).
Any of: DOI, PubMed ID, PMCID, arXiv ID, ISBN, ISSN, NASA ADS bibcode, WHO IRIS URL. You only need one. Optional refinements - first-author family name, year, journal - are used to tighten the match but never gate the verdict on their own.
Yes - PMID and PMCID are first-class identifiers here. Paste either and the verifier resolves it against PubMed, then compares the claimed title the same way it does for a DOI. This matters for biomedical references, where a fabricated citation often carries a plausible-looking PMID; many citation checkers stop at DOI and arXiv identifiers and cannot validate a PubMed ID at all.
Yes - detecting fabricated citations, including AI-hallucinated references, is the core job. Two things separate it from most fake citation detectors. First, the verdict is not a binary real/fake: matched / mismatch / ambiguous / not found distinguish a fabricated title from a wrong-but-real identifier. Second, it shows its work - per-field similarity scores on every verdict, and a published validation set with confidence intervals at /citation-integrity instead of an unsourced accuracy claim.
Matched means the claimed title matches the resolved paper. Mismatch means the identifier resolves but the title disagrees - the Topaz et al. fabrication pattern. Ambiguous means the identifier resolves to one paper but the claimed title matches a different paper found via search, which usually indicates a wrong identifier for a real paper. Not found means neither the identifier nor the title resolves anywhere.
The web verifier is free at the anonymous tier with a published rate limit (see API documentation). No signup. Nothing you paste is stored beyond standard server logs. Programmatic access via the REST API and MCP server is metered on RapidAPI (free tier available; paid plans scale up). The opt-in LLM screen, which catches informal-abbreviation false positives, is gated to authenticated paid callers because each model call carries real per-call cost.
Yes - switch to Batch above to paste a BibTeX, RIS, or CSL-JSON bibliography (or upload a .bib / .ris / .json file). Up to 10 entries are verified at once and rendered as a per-row verdict table. For a whole reference list with retraction checks and a downloadable evidence report, use the Bibliography Audit (up to 25 entries per run). For scripted batches, call POST /api/verify or POST /api/audit directly, or use the verifyCitation MCP tool from Claude Desktop, Cursor, or any MCP-aware client.
Retraction asks 'has this real paper been retracted'. Verification asks 'does this citation actually correspond to a real paper at all'. A fabricated citation is not retracted because the paper never existed. Use the Retraction Checker after verification, on the resolved paper, not on the claim.
On a 1,402-entry blind holdout - drawn 2026-08-01 from a recorded seed after the code was frozen, hand spot-checked 100 of 100, then measured exactly once on 2026-08-02 - overall fabrication recall was 87.3%, with 4 of the 5 dominant patterns caught perfectly (real-DOI + fabricated title 35/35, wrong author 35/35, wrong DOI 35/35, PMID/DOI split 25/25). High-confidence false-accusations on correctly-cited papers were 0.87% (Wilson 95% CI 0.47-1.59%). That rate fell from 2.70% to 1.22% against the previous 1,395-entry holdout after five author-comparison fixes - a 55% relative drop, p = 0.0098 - while detection held. We also report a measured blind spot: single-word near-miss semantic flips, caught 10/35, with every miss returned at high confidence. Earlier 1,395-entry, 350-entry and discovery-run holdouts (the 5.3% -> 1.8% lineage) stay published alongside it. The fixtures, methodology, and downloadable receipts are at /citation-integrity.
verifyCitation tool for AI-coding workflowsThis page is a front end for the public Scholar Sidekick API — the same citation check runs over plain HTTP, so you can call it from a script, an AI agent, or your reference manager. Light use needs no key at all. A free account issues an ssk_ key that raises your rate limit.