The checker said 90% of my citations failed. The checker was wrong.

A doctoral candidate in the health sciences ran a completed thesis — years of work, a couple of hundred references — through a citation check and got a number that would make anyone's stomach drop: roughly nine in every ten citations could not be matched to the reference list. The true number of problems in that thesis was two.

This is a story about how a checking tool manufactures a catastrophe out of one broken assumption, why the shape of a failure report tells you more than its size, and the sanity questions worth asking before you spend a weekend "fixing" citations that were never broken.

What 90% actually looks like

The thesis used Vancouver referencing — the numbered style standard in medicine and much of the health sciences. In the text, citations appear as bracketed numbers. At the back, the reference list gives each source once, numbered in order of first appearance, authors written surname-first with initials and no comma play-acting as punctuation of any other style:

…shown to improve adherence in older cohorts [12], though replication
outside hospital settings remains limited [13,14].

12. Okafor T, Lindqvist M, Duarte P. Medication adherence in
    community settings. …; 2021.

The report the candidate received said, in effect: citation [12] has no matching reference. Citation [13] has no matching reference. And so on, page after page — the overwhelming majority of the document flagged as cited but not in the bibliography. A handful of items passed, seemingly at random, which somehow made the report feel more credible rather than less.

A report that says "you made two hundred mistakes" is almost never describing two hundred decisions a human got wrong. Humans don't fail that consistently.

The real defect: a reference list that parsed to nothing

Here is what had actually happened, and it happened before a single citation was examined. The tool read the reference list expecting author–date entries — the APA-style shape, "Surname, A. B. (2021)." — because that is what most checking tools silently assume. Vancouver entries don't have parenthesised years. They don't put a comma after every surname. Their numbers live at the front of the line.

Faced with two hundred entries in a shape it wasn't expecting, the parser didn't extract two hundred slightly-wrong records. It extracted nothing. An empty reference list. And against an empty list, every citation in the document is, by definition, unmatched. The 90% wasn't measuring the thesis at all. It was measuring the tool's assumption, applied two hundred times.

1. Tool assumes author–date entries
2. Vancouver list parses to 0 references
3. Every citation compared against an empty list
4. Every comparison fails → "90% missing"
5. The report looks like a thesis problem.
   It is a parser problem.

Notice what makes this failure so convincing: each individual line of the report is locally plausible. "Citation [12] not found in reference list" is exactly what a real missing reference looks like. Only the totality is absurd — and the totality is the thing a worried student is least equipped to step back and see at two in the morning, a week before submission.

The style assumption is everywhere

Why would software assume author–date? Because most of the documents most tools were built on use it. APA and its cousins dominate the social sciences, education, business, psychology. If your test corpus is drawn from there, author–date stops looking like one style among many and starts looking like what a reference is. Then a medical thesis arrives, written in the style that medicine has used for decades, and the software treats the document as malformed rather than the assumption.

This is why we treat style as the first thing to establish, not the last. Before any matching happens, the document itself gets to answer: are these citations numbered or named? Do the reference entries lead with numbers or with surnames? A thesis is usually consistent about this — overwhelmingly one style, with stray exceptions — and that dominant style then governs how everything downstream is read. (What we do with the stray exceptions is its own story.)

The sanity rule that catches this class of failure

There is a simple bookkeeping principle that makes a report like "90% failed" impossible to ship without scrutiny: every item the document contains must end up somewhere, and the somewheres must add up. Citations found, references found, matches made, orphans, misses, and things discarded as not-actually-citations — the totals on each side have to balance, at every stage.

Run that ledger against the broken report and the anomaly announces itself instantly: two hundred citations entered, zero references entered. A document with two hundred citations and a forty-page reference section that contributes zero parsed entries is not a document with missing references. It is a document the software failed to read — and the honest output at that point is "we could not read your reference list," not a page of accusations. We wrote about this principle in Nothing gets lost: the conservation ledger.

So what was actually wrong with the thesis?

Once the reference list was read as Vancouver and the matching re-run, the two hundred alarms collapsed into two findings:

• One orphan — a reference sitting in the list that no citation in the text ever pointed to. Common, easy to resolve: cite it or cut it.
• One genuinely missing reference — a citation in the text with no entry in the list. The only real instance of the thing the first report had alleged two hundred times.

Two findings, both fixable in minutes, both the kind of tidy-up examiners expect a final pass to catch. That is what a healthy thesis looks like: not zero findings — real documents are rarely perfect — but a small number of specific, explainable ones. We've written more about reading a healthy report, because knowing what normal looks like is half the defence against reports that aren't.

How to read your own failure report

Whatever tool produces it, put a citation report through these four questions before acting on a single line of it:

1. Does the failure rate pass the human test? Anything above about a quarter of the document means either a systematic problem or a systematic misreading. Individual mistakes cluster in the tens, not the hundreds.
2. How many references did it actually parse? If the tool won't tell you, that silence is itself the finding. A near-zero parsed count with a thick reference section means the list wasn't read.
3. Do the failures share a shape? Two hundred different mistakes look varied. One broken assumption looks uniform — every failure worded identically, spread evenly across the whole document.
4. Was your style detected or assumed? A tool that never names the referencing style it thinks you used has assumed one. If you wrote in Vancouver and it assumed author–date, no number it prints afterwards means anything.

And one thing this story is not: a reason to distrust checking in general. The candidate's instinct to verify before submission was exactly right — a real missing reference did exist, and it was found. The lesson is narrower and more useful: trust reports that show their working, and interrogate reports that only show their conclusions.

One more thing worth checking once your citations all resolve: whether the references themselves are still current — superseded editions, preprints that have since been published, even quietly retracted papers. That's a different job from matching, and it's what our sister tool CitationUpdater does before submission.