Methodology

What citation verification means here

CitationCheck answers one narrow question: does this citation appear to correspond to a real scholarly record, and do its bibliographic details agree with that record? It does not judge quality, correctness, retraction status, or whether the work supports any claim.

The evidence hierarchy

Checks proceed from strongest evidence to weakest: parse the citation into structured fields; detect identifiers (DOI, PMID, arXiv ID); resolve identifiers before any fuzzy search; compare returned records field by field; search configured providers when identifiers are absent or fail; classify using documented thresholds; show the evidence and limitations.

Identifier resolution

A resolved identifier anchors the check, but it never auto-passes it. If the identifier resolves and the title substantially agrees (normalized similarity of at least 75%), the result is Verified. If the identifier resolves but the title disagrees, the result is Partial match — a real DOI on the wrong paper is a known fabrication pattern as well as a common honest error.

Field matching

Titles and venues are normalized before comparison: lowercased, accents stripped, punctuation unified, whitespace collapsed. Author comparison uses family-name overlap, tolerant of initials and name order. Years may differ by ±1 and still count as agreeing (±2 where a preprint is involved), because online-first publication and indexing genuinely shift years.

Scoring and thresholds

When no identifier resolves, compared fields combine into a weighted agreement score: title 45%, authors 25%, year 15%, venue 15%. Fields that could not be parsed are excluded and the weights renormalized. A score of at least 85% classifies as Verified; at least 55% as Partial match; below 45% the candidates are treated as inadequate and the result is Not found (with the weak candidates still listed); the range between is Needs review. The score describes field agreement among compared fields. It is not a probability that a source is real, and it is never shown without the underlying comparison.

Preprints and versions

An arXiv preprint and its published journal version are related but different records. When a preprint is involved, venue mismatches are softened and year tolerance widens — but the result card shows both versions' details so you can see which one you are actually citing.

Provider coverage and failure

Providers fail, time out, and rate-limit. A provider failure is reported as Service unavailable — it is never converted into Not found. A Not found result requires at least 2 providers to have responded successfully; with fewer, the result is Needs review with the incomplete coverage stated.

Why "Not found" is not proof of fabrication

The enabled sources have real gaps: regional publishers, books, theses, reports, non-English venues, very recent works, and non-scholarly sources (news articles, webpages, standards) may not appear. Misspellings and incomplete details also defeat search. Not found means "we could not find an adequate record through these sources" — nothing stronger.

Why a real citation may still be a poor source

Verification is bibliographic. A real, correctly cited paper can be methodologically weak, superseded, corrected, retracted, or simply irrelevant to the claim it is attached to. Reading the source remains your job; this tool just makes sure you're reading the right record.

Is AI involved?

No. Parsing, normalization, identifier resolution, matching, and classification are deterministic code. No language model generates, corrects, or fills in bibliographic facts anywhere in the pipeline, and suggested records come only from data actually returned by providers.

How thresholds are determined

Thresholds are set in configuration, exercised against a fixture suite of representative citations (correct, malformed, mismatched, fabricated-looking, preprint/published pairs), and adjusted when incorrect-result reports show systematic error. They are judgment calls made inspectable — not claims of statistical optimality.

Reporting an incorrect result

Email the support address in the footer with the citation you checked and, if possible, the DOI or a link to the record you believe is correct. Reports of wrong classifications directly drive threshold and parser fixes.

Last substantive methodology update: August 7, 2026.