Sheet 3 · Edition 2026-08-30 02:52 UTC

Open data file formats, counted across 31 government catalogues

Every entry in a government catalogue advertises a format for each downloadable resource. Read across these catalogues, that is 3,828,453 resources naming their format in 3,547 different strings. One string per resource, and the counts sum to the resource total exactly. This sheet is the whole of that column, sorted. If you are about to filter one of these catalogues by format, it is what you would be filtering against. One further catalogue is registered and has not been opened, and nothing from it is in these figures.

3,828,453

Resources counted

3,547

Distinct format strings

300

Strings published in full

31

Catalogues read

The 20 most common values, verbatim

Exactly as the portals wrote them, lower-cased and trimmed and nothing else. No two of these rows have been merged, which is why the same format appears more than once in the list.

• html562,391 · 14.7%
• csv517,427 · 13.5%
• pdf266,761 · 7.0%
• xls232,438 · 6.1%
• xlsx212,942 · 5.6%
• zip179,866 · 4.7%
• json131,056 · 3.4%
• http://publications.europa.eu/resource/authority/file-type/csv129,342 · 3.4%
• unknown122,438 · 3.2%
• xml116,718 · 3.0%
• http://publications.europa.eu/resource/authority/file-type/netcdf108,072 · 2.8%
• application/x-netcdf107,261 · 2.8%
• wms80,991 · 2.1%
• http://publications.europa.eu/resource/authority/file-type/html70,900 · 1.9%
• png59,144 · 1.5%
• wfs49,792 · 1.3%
• tiff31,854 · 0.8%
• http://publications.europa.eu/resource/authority/file-type/xml31,382 · 0.8%
• http://publications.europa.eu/resource/authority/file-type/pdf30,965 · 0.8%
• geojson30,435 · 0.8%

Below the published 300, a further 3,247 distinct strings cover 15,712 resources, 0.4% of the corpus. They are counted but not listed: portals declare formats as free text. The export records only how many strings and how many resources, so the most that can be said about their spread is the average, 4.84 resources per string.

HTML leads on single values, CSV once spellings are added up

The single most declared value in the corpus is html, on 562,391 resources. A further 7 strings say the same thing in different words, and together the 8 of them account for 635,794 resources, 16.6% of everything catalogued. The two rankings disagree, and the disagreement is the finding. Added up across their spellings, CSV totals 650,304 resources and HTML totals 635,794, so which format is the most common depends entirely on whether the spellings of one format are counted as one format. Neither figure is adjusted, and both are in the table below. A resource declared as HTML is a link to a page. It may be a page with a table on it, or a landing page with a download button, or a page that no longer exists. What it is not is a file, and no catalogue field records which of those it is.

• html562,391
• http://publications.europa.eu/resource/authority/file-type/html70,900
• text/html; charset=utf-8931
• https://data-lra-cha.opendata.arcgis.com/api/feed/dcat-ap/ftype/html617
• html_simpl518
• file:///srv/udata/ftype/html179
• shtml159
• htm99

One format, several spellings

Grouping these is a judgement, so it is shown as one. Each row prints the exact strings it is the sum of, and the total beside it is the sum of those strings and of nothing else. No row claims to be every occurrence of its format: only the published 300 strings were searched, so the 3,247-string tail is not in any of these figures.

15 formats, 3,139,344 resources, 82.0% of the corpus, spread across 89 different spellings. A compressed or profiled variant is never folded in: csv.gz is an archive and kmz is a zipped KML, so neither counts as its parent format here.

Four things that break a format filter

These overlap the groups above and each other on purpose: one string can be both a spelling of CSV and a URI. Nothing here is added to anything there.

How much of this has been checked against the file

A declared format is a claim the portal makes in its catalogue. Everything above counts those claims; it does not verify them. This harvester has fetched 6,540 of 3,828,453 resources, 0.2% of the corpus, and the download phase is what the rest is waiting on. Of the ones fetched, 74 returned a web page in place of the file they advertised and 4 came back empty. Those are counts of what happened, not a rate: the fetch order is not a random sample of the corpus, so nothing here can be scaled up to it. The per-format byte figures for what was mirrored are on the harvest status sheet.

Where the per-catalogue figures are

This sheet is the corpus total. Every catalogue also has its own declared-format breakdown, on its register entry, and the same figures are in the JSON behind this page: formats.json for the corpus and portals/<portal_id>.json for one catalogue. Both are linked, with their terms, from the export. How the harvester reads a catalogue in the first place is on the method sheet.