Where every answer on this site comes from, how it is checked, and what "last checked" on a page actually means.
Last updated
Every page on this site carries a line saying when its data was last checked. This page explains what happens between a source publishing a fact and that line appearing.
GeoNames is the spine. It supplies the hierarchy of countries and their divisions, the names in English and in the local languages, the coordinates, the population figures and, most importantly, a stable numeric identifier for every place. We key everything to that identifier rather than to a name, because names change and identifiers do not. Kazakhstan's capital has been called Akmolinsk, Tselinograd, Akmola, Astana and Nur-Sultan within a century, and it is the same city with the same GeoNames id throughout.
Divisions are loaded to four levels where GeoNames models them. A country with only two levels of administration gets two.
Wikidata holds its own record of which city is the capital of which place, maintained by a different community with different sources. Every capital relation the pipeline loads is compared against it.
Where the two agree, the relation is published. Where they disagree, the conflict is written to a table with both values and the page waits for a human decision. If more than one relation in twenty disagrees, the whole run is refused rather than published, because that rate means something has gone wrong upstream rather than in twenty separate places.
Decisions made by a human are recorded as overrides, with the reason and the date, and survive the next run. That is how a page can say something different from both of its sources: because someone checked, and the record says who and why.
The shapes on the maps come from Natural Earth for countries and most first-level divisions, and from geoBoundaries where Natural Earth has no match. Both are simplified before publishing, at three levels of detail, so a page carries the smallest outline that still looks like the place.
We only use a boundary file whose licence asks for attribution and nothing more. That rules out a lot of good data. Sweden has no municipal boundaries on this site because the available file is share-alike; Canada's provinces and some of the United Kingdom's outlines are excluded on the same ground. It is better to show no boundary than to publish one we cannot license onward to the people who use our dataset.
Photographs come from the image recorded on a place's Wikidata item, and from a search of Wikimedia Commons for cities. Both carry machine-readable licence metadata for every file, which is why they were chosen first: a candidate can be judged rather than guessed at.
The licence rule is strict and it is enforced by the database, not by a person remembering. A photograph can only become the picture on a page if it is public domain, CC0, or CC BY at any version. Share-alike is excluded. Attribution explains why at length.
A photograph of people, an occasion, or a satellite view is never used as the main image of a place, whatever its licence. The first two are pictures of an event rather than of the place, and the third is a map, and the page already has one.
It is the date the pipeline last read the place from its sources and confirmed the values, not the date the fact became true. A capital that moved in 1991 and was verified last month shows last month.
Boundaries, photographs and descriptions carry their own dates, and the run report behind each publish records what changed.
Tell us. The contact page takes corrections, and a correction with a source attached is acted on much faster than one without, because the research only has to happen once. Corrections are reviewed by a person and, when accepted, recorded as overrides so the next pipeline run does not undo them.
Every source, with the licence it is served under and what we use it for, is on the attribution page.