CitationLab — REFerence [IN]tegrity for universities and researchers

Every in-text citation in your thesis should resolve to an entry in your reference list, and every entry in that list should actually be cited. CitationLab checks both — deterministically, against the document your examiners will read — and certifies the result with a Ref[In] Score.

---

The gap it fills

A university’s quality stack covers originality, argument, method and presentation. A similarity checker looks for overlap with other texts. None of it measures whether the references hold together — and in a 300-page thesis with 250 references accumulated over four years, across chapter drafts merged and re-merged, they frequently do not.

Because nothing routinely looks at it, the problem surfaces at the worst possible moment: in the viva, in the hands of an external examiner reading closely. It is rarely misconduct. It is almost always administrative drift — and it is entirely preventable.

How it works

1 · Start at the document, not the reference manager

Your reference manager holds what you meant to cite. Your thesis holds what you actually cited. Only one of them is going to be examined. Upload DOCX, ODT, LaTeX, PDF or TXT up to 100 MB and the engine reads all of it — body, appendices, and any chapter-end reference lists after the main bibliography. Word reads best, and the upload screen says so before you choose.

2 · Cross-check deterministically

Every citation is paired against the reference list parsed from your own document, on a fixed backbone applied in a fixed order of confidence: shared identifiers first (DOI, PMID, legal instrument, standard number), then exact author-year, then the reference line as written.

No model decides whether a citation matches a reference. That is why the cross-check is free and can be re-run as often as you like — it is a rule, and rules give the same answer every time. Every citation lands in one of four bands:

• Matched — citation and reference agree.
• Questionable — a pair resting on weak evidence: a bare year marker, two different styles, a contested reference.
• Missing — cited in the text, absent from the list. The reader is pointed at a source the document never identifies.
• Orphan — listed in the bibliography, never cited. Almost nobody catches these by eye.

3 · Find what is genuinely missing

Six databases — Crossref, OpenAlex, Semantic Scholar, PubMed, Google Scholar and Google Books — establish that a reference is genuinely missing rather than merely misspelled, hyphenated or accented differently, before any AI judgement is invited. PubMed is only consulted where a citation reads as biomedical. Your own uploaded library of PDFs, DOCX or BibTeX is searched first, because the file on your disk is the edition you actually read.

4 · Update, format and export

Ageing references can be reviewed and replaced against a threshold you choose, annotation builds in order, and export supports fifteen citation styles — drawn from the Citation Style Language project and rendered through Citation.js.

---

Reading the Ref[In] Score

The arithmetic is published rather than proprietary, because a score that governs decisions has to be legible. It has two halves, deliberately weighted:

• Citation resolution — the share of in-text citations that resolve to a reference. Counts triple.
• Reference utilisation — the share of reference-list entries actually cited. Counts once.

Nothing else moves the number. A citation with no reference makes a claim the reader cannot verify; an uncited entry is untidy but harmless. Identifier coverage, styles in use and completeness are all reported as evidence but do not affect the score — so a law or humanities thesis is not marked down for a field that rarely publishes DOIs.

Bands run from Excellent to Critical on a fixed scale that never adapts to the document or the cohort — which is what makes two reports comparable, and lets a department read a distribution rather than a pile of unrelated numbers. The same report issues at three levels: Level 1 organising, Level 2 reference checking, Level 3 final export. The level records how far the work had gone, not how good it is; any level can come out green, amber or red.

Where the AI sits — and where it does not

A language model asked to eyeball a bibliography will produce something impressive-looking and occasionally fictional. Fabricated citations are a well-documented failure mode, and a checker that can invent a reference is worse than no checker at all — it converts a visible gap into an invisible fabrication.

So AI here is advisory only: it tidies messy in-text forms, proposes corrections on weak pairs, and helps locate references the databases have already confirmed are missing. Every suggestion is approved or dismissed by a human, and nothing in your thesis is rewritten automatically. Each run closes with a conservation ledger reconciling what entered against what left — and where a check has not run, it reports itself as unaudited rather than claiming a balance it has not verified.

---

See it for yourself

• Walk every screen in the product tour, or read the full Product Guide (PDF).
• Try it on your own document with the live demo.
• Credit costs are on the pricing page; method questions are answered in the FAQ.
• Read our own writing on it: reference integrity as institutional risk, a full cross-check walkthrough, and why a citation checker should not be an LLM.