How complete Wikipedia's citations are

A meta-bibliometric portrait of the fields the corpus's own references carry

Wikipedia is a citation dataset of its own: 29.4 million reference records whose bibliographic fields can be audited like any library catalog's. The audit has a shape. Almost every record says what it cites; most say where it was published; only half name who wrote it. That is the difference between findable and attributable — the corpus can overwhelmingly be looked up long before it can be checked.

Figure 2. The completeness portrait of release 2606's citations: of six bibliographic fields, title is near-universal (99.7%), publisher (86.6%) and URL (78.8%) usual, author a coin flip (50.0%), and ISBN (7.6%) and DOI (5.4%) rare — and the decade rows show author, DOI, and ISBN coverage peaking with the scholarly catalogs of the 1980s–1990s (author 76.4%) while URL coverage climbs to 93.2% in the 2020s. Denominator: corpus-wide rates and the by-medium block use all 29 395 357 citation records; the decade columns and the decade × medium panel condition on the 21 936 052 records (74.62%) with a parseable four-digit cited year — the 7 459 305-record no-year group stays visible, and the 91 sparse decades outside the 1800s–2020s window (107 185 records, 0.5% of the year-parseable subset) are omitted from the grid.

Source: K2-citeyears receipt (A3 completeness portrait) over the enwiki-20260601 citation store, via the dataset page.

How to read the portrait

Each row is a bibliographic field, ordered best to worst, with its corpus-wide rate in cyan. The wide block shows coverage per cited-year decade; the narrow block re-cuts the same rates by the citation's medium; the lower panel crosses the two, at full decade × medium resolution, for the two fields the receipt carries that deep — author and DOI. Decade-conditioned cells count only records whose cited year parsed to four digits; the no-year column keeps that conditioning visible rather than hiding it. Hover or focus any cell for its rate and record count.

What the gaps track

The decade columns tell an era story. Author coverage peaks with the scholarly citation habits of the 1980s–1990s (73.4%, then 76.4%) and slides to 54.8% in the 2020s; DOI and ISBN crest alongside (11.8% in the 2000s and 28.5% in the 1990s) and fall to 4.2% and 2.0%; URL coverage does the opposite, climbing from roughly 40% through the mid-century decades to 93.2%. Title never moves (99.9–100%). Citations with no parseable year — 7.5 million records, mostly bare web references — are the least complete of all: 16.7% name an author, though 93.9% carry a URL.

The medium rows tell a division-of-labor story. Journal citations are the scholarly ideal: authors for 92.1% and DOIs for 63.7% of them, against web citations that carry URLs 98.4% of the time but authors only 31.4%. The lower panel shows the decade decline is dilution, not decay: among 2020s journal citations, author presence is 96.6% and DOI presence 82.0% (against 0.07% on the web) — the shortfall lives in the web's growing share. Book citations are the only place ISBN lives (57.5%), and publishers are present nearly everywhere except news (66.4%). None of this is an error to be bulk-fixed — it is which evidence each medium physically carries, and it is where the citation parser's author-recovery work has the most headroom.

Author, title, and publisher rates use the receipt's raw (pre-normalization) columns. The full receipted distribution — overall, by decade, by medium, and the author/DOI decade × medium matrices — lives in the figure's data payload, described on the dataset page.