lgelauff/wikipedia-image-requests

Daily image-request counts for Wikipedia article images

Daily counts of how often each of ~1.19 million Wikipedia article images was requested from Wikimedia's image servers, attributed to the wiki whose page the request came from, together with the article each image appears on.

406,263,728 daily observations covering 1,186,150 image files across 534,067 articles in nine Wikipedias, from 2025-09-01 to 2026-09-05 (370 days).

Every file in the set appears on exactly one main-namespace article of its wiki. That property is what the set is selected for, and it is what makes per-article attribution possible at all — see Why the article column exists.

Files & schema

file	rows	images	articles	wikis
file-requests_en.wikipedia.tsv.gz	345,986,855	1,005,803	505,698	English
file-requests_8-wikis.tsv.gz	60,276,873	180,347	28,369	ar, bn, he, hi, hu, ro, simple, uk
positions_en.wikipedia.tsv.gz	9,468,892	—	2,989,544	English
units_en.wikipedia.tsv.gz	1,012,102	—	—	the exact file list collected
units_8-wikis.tsv.gz	192,563	—	—	the exact file list collected

Request series — tab-separated, no header, multi-member gzip (decompress with gzip -dc, not zcat, which returns zero rows without error on some platforms):

referer        file                                     article     date      requests
en.wikipedia   Si_Woo_Kim_Travelers-012_(cropped).jpg   Kim_Si-woo  20250901  42

column	meaning
referer	the wiki whose page the request came from, e.g. en.wikipedia
file	file name, underscored, no namespace prefix
article	main-namespace article the file appears on, underscored
date	YYYYMMDD
requests	count of HTTP 200 responses for that file, that day, from agent=user

Positions — tab-separated with a header: article, file, k, k_any, char_offset, page_chars, n_images. k is the index of the enclosing top-level section (0 = before the first heading); k_any counts every heading level regardless of depth. Derived from the article wikitext of the same dump date.

Why the article column exists, and what it is

The upstream service does not report a page, and the raw request does not contain one.

Wikipedia serves a referrer policy of origin / origin-when-cross-origin, and images are served from a different origin than any article host. So the Referer accompanying an image request carries the site and nothing more — https://en.wikipedia.org/ — never the article path. The published statistics reflect that: the served dimensions are referer × media-type × agent, and no endpoint takes or returns a page.

⚠ This limitation is not stated in the API documentation — the served specification describes the endpoint only as returning "requests for a media file", and mentions pages or articles nowhere. The single hint is that the referer parameter is documented as the "Origin of the request", which is precisely accurate and easy to read past.

⇒ this dataset supplies article attribution by construction: the file was used on exactly one main-namespace article, so that article is the only article it could have been rendered from.

⚠ "Only article" is not "only page." The eligibility rule considers main-namespace usage only. Measured on 120 randomly sampled English files: 88% are used on exactly one page of any kind, and 12% also appear on a non-main-namespace page — user sandboxes (15 occurrences), portals (5), project pages (3), templates (2), talk pages (2). Requests generated by those pages cannot be distinguished. Templates are the case to weigh: a file used in a template loads on every page that transcludes it.

Which files are included

A file is eligible when it is used on exactly one main-namespace page of the wiki, as of the 2026-09-01 dump. Cross-wiki reuse is irrelevant, because each request is already attributed to the wiki it came from — and 10,119 files in this set are single-use on more than one wiki, appearing once per wiki.

From the eligible set, this release covers articles that appear in a fixed panel of pages exceeding 150 views on at least one day (2017–2026) with such a day in 2025 or 2026; carry at least two eligible files of type JPEG, PNG or SVG; and have at least one such file before the first section heading and at least one after it. For English Wikipedia that is every article meeting those conditions — 505,698 of them, two images each. The eight other wikis carry all eligible images per article rather than two, averaging 6.8.

How it is built

1. Eligibility from the MediaWiki SQL dumps (imagelinks, linktarget, page), 2026-09-01. No API can answer "which files across a whole wiki have property X", so this step needs a dump.
2. Positions from the pages-articles XML dump of the same date, so structure and eligibility describe the same instant. Validated against rendered HTML: agreement on the section index in 608 of 609 sampled cases.
3. Counts from the Wikimedia Analytics Query Service, metrics/mediarequests/per-file/{referer}/user/{path}/daily/{start}/{end}.

Collected at 27–30 requests/second from Wikimedia Cloud Services with an adaptive rate limiter; zero rate-limit responses across 1.2 million calls, and zero incomplete units.

Coverage & caveats

• Only HTTP 200 is counted. A 304 Not Modified — a browser revalidating a cached image — is not counted upstream. Files differ in how often they are already cached, so counts are not a direct measure of how often an image was displayed.
• Days with no requests are absent, not zero. Zero-fill within a file's valid span at analysis time; a missing date does not mean the file did not exist.
• 6,299 English units (0.62%) returned no requests at all across the whole 370 days, and 353 articles have no data for either image. The file exists and the query was correct; the answer was nothing.
• agent=user only. No spider/crawler series is included.
• Per-file requests carry no device dimension. There is no desktop/mobile split for media requests from any public source.
• Requests are not page views. A page view may fetch zero, one, or many images depending on client, viewport, and whether images below the fold are scrolled into view. Daily page views for the same 506,051 English articles — per access channel, redirects included, 2025-09-01 to 2026-09-05 — are published separately in the Hugging Face bucket Schroedinger1000/Wikidatavector-derivatives, folder enwiki-pv/20250901-20260905/, for use as the denominator when relating image requests to article traffic.
• Positions are read from wikitext, so an image reaching a page only through a template body is not located. Roughly 16% of files are unplaced; they are absent from positions_en.wikipedia.tsv.gz rather than recorded with a wrong position.
• Eligibility is measured on one date. A file single-use on 2026-09-01 may not have been throughout the window; membership drift is not corrected here.

Provenance & license

Derived entirely from public Wikimedia sources: the MediaWiki SQL and XML dumps, and the Wikimedia Analytics Query Service. No non-public data is used and no personal data is present — the counts are aggregate, per file and per day.

Article and file titles originate in Wikipedia and Wikimedia Commons and are licensed CC BY-SA 4.0; the aggregate counts come from Wikimedia's published analytics service. The dataset as a whole is released under CC BY-SA 4.0.

Collected 2026-09-08/09 against dumps dated 2026-09-01.