Methodology

A snapshot, not a feed

The data here was captured on 2026-09-24 and does not update itself. Populations are restated, borders and capitals change, cities twin and quietly untwin. Nothing on this site re-reads its sources; every figure is what that source said on the day it was read, and the gap only widens.

Treat it as a printed reference with a date on the spine. For a current figure, go to the source named in the table below.

What this is

An atlas of 662 cities, chosen by two rules: every city with more than a million inhabitants, and the capital of every sovereign state whatever its size. That is 556 cities over a million and 202 capitals — the two sets overlap, since most capitals are large, so the union is 662 rather than their sum.

The second rule is what brings in Vaduz, Funafuti and Vatican City, and what takes country coverage to 198.

The pages are organised around what cities share rather than what they are. A city page is mostly a set of doors: to the parallel it sits on, the offset it keeps, the place on the far side of the Earth, the cities that have formally twinned with it.

Sources

How cities were matched

Each city was matched to a Wikidata entity in four passes, widening from an exact name-and-country match to a coordinate search. A match was only accepted if the candidate sat within a tolerance of the gazetteer coordinates and carried at least two properties a city would have.

That last condition is the important one. The risk in this kind of matching is not the city you miss; it is the plausible wrong answer you accept and then publish with confidence. 662 of 667 candidates matched. The six that did not are districts and neighbourhoods the gazetteer counts as cities; they are excluded rather than guessed at.

What not to trust

Coordinates are given to two decimal places, roughly a kilometre, and refer to a point the gazetteer considers the city centre rather than any administrative boundary.

Population comes from a gazetteer mixing censuses with estimates of different vintages. A noticeable minority are round numbers: multiples of a thousand are about eleven times more common than chance would give, so apparent precision is often rounding. Figures are not comparable between cities as though they were measured the same way. Where Wikidata disagrees, both are shown rather than one being silently preferred.

Twinning is recorded as Wikidata holds it, and the relation is often one-sided: a city lists a partner that does not list it back. Those are shown as stated rather than reconciled.

Generated text and images

452 pages carry a short written passage, and 662 cities have an illustration. Both were generated by a language model and an image model respectively, and both are labelled as such where they appear.

The passages were produced from the data on the page and checked against it: any output introducing a figure not present in the source was rejected and regenerated, so the prose cannot assert a number the data does not support. It can still be flat or obvious, which is a different problem.

The illustrations are paintings, not photographs, and were generated from written research into each city's setting, architecture and materials rather than from any source image. Every one carries a caption saying so, and a block listing what guided it. That block is headed "what guided this image" rather than "what is shown", because a prompt records what was asked for, not what came back.

Corrections

Errors here are most likely to be an inherited wrong figure or a confident wrong match. If you find one, the useful thing to say is which page, which figure, and what the source you trust says instead.