How we count.
Most sites that cite a number of references type that number into a page. This one generates it at build from the same file the grades are authored in, and prints every population it distinguishes rather than a single figure with nothing behind it.
In August 2026 there were six different answers to "how many references does this site cite" in circulation at once, every one computed from the same underlying file on the same day. None was an arithmetic mistake. Each was a different piece of code walking a different set of fields, and nobody had written down which set was the right one.
Two of those figures were sitting in documents. That is the same defect this site was built to find in other people's material, which makes it the least acceptable place for it to happen. The fix was not to pick a winner. It was to write down what is being counted, generate it from one place, and publish the definition beside the number.
A PubMed identifier held in a declared citation field of assets/data.json: pubmed_ids, pmids, page_only_pmids or pmid. An identifier is a positive integer of up to eight digits, which includes the shorter ones carried by papers from the 1970s. Nothing is counted by pattern-matching the file, and nothing is counted by scanning the published pages: a number that looks like an identifier is not a citation, and a count taken from the page tree changes with whichever files happen to exist when it runs. Every figure below is written by the same module the retraction watch and the citation audit read, so this page cannot quietly disagree with them.
Distinct papers cited on a page a reader can open. This is the figure any public statement about the size of the reference base should quote, and the one every other count on this site is reconciled to.
Every distinct paper held in a declared citation field, including those supporting records whose page is not published yet. The retraction watch re-checks this larger set against PubMed on every build.
How many times a claim points at a paper. Higher than the number of papers because the same trial supports more than one entry, which is ordinary and not a defect.
Verified sources held for a page that is currently unpublished. Real, checked, and deliberately excluded from the rendered figure because a reader cannot reach them.
Rows in the citation registry, which records each paper's year, design and trial registration. Every one already appears elsewhere, so these are counted separately: adding an index to the thing it indexes inflates a citation figure with its own metadata.
Papers withdrawn after publication and formally retired here, with the date and the reason recorded against each. Not part of the reference base. A retirement is the process working, not a defect in it.
The machine-readable version of this table, including the count of citations held in each field of the underlying data, is at /assets/reference_counts.json. It is rewritten on every build and carries the same date shown above, October 2026.
It is not a quality measure. A longer reference list is not better evidence, and a page can cite forty papers and still rest on none of them. The grade on any entry here comes from what its trials actually tested and in whom, never from how many were attached to it. Counting references measures breadth, and breadth is the least interesting property evidence has.
It is not fixed. The figure moves whenever a paper is added, retired or superseded, which is why every statement of it here carries the date it was generated. Two documents written a week apart may honestly carry different numbers, and saying so is more accurate than forcing them to agree.
It is not a claim that every paper was read the same way. What each citation was checked for, and what remains unverified, is set out separately in the citation audit, which lists every defect found in this site's own citations rather than only the ones found in other people's.