Methodology

Two layers, clearly labeled

Historical layer. The complete final edition of the CIA World Factbook (data as of January 2026), preserved verbatim inside every country profile. It is a US Government work in the public domain (17 U.S.C. § 105). We label it as historical everywhere it appears and never silently edit it.

Live layer. Statistics fetched from primary sources and merged in on each rebuild. Every live figure carries its source and observation year inline, because a number without a year is a guess dressed up as a fact.

We run no refresh on a fixed timetable, and we would rather say so than publish a cadence we do not keep. In practice this matters less than it sounds: only one field here moves on its own schedule — heads of state and government, which change whenever a government does. The statistical sources publish rarely, the World Bank a few times a year, the IMF each April and October, the UN Population Division roughly every two years, Our World in Data annually, and area does not come from a live feed at all. So rather than ask you to trust a schedule, we date every figure and every page, and you can read the age off the site directly.

Do not take the schedule on trust: take the date. The "last data change" stamp above, the updated date on every country profile, and the lastmod in our sitemaps all come from the same content hash, and none of them moves unless the underlying data moved. If a page's date looks old, the data behind it has not changed since that date, or the refresh has not run. Either way the date is the truth and the schedule is only an intention.

The pipeline

1. 01 Fetch. A scheduled job pulls all live sources (35 HTTP requests for a clean run) and stores dated snapshots in our repository. A source that fails keeps its last good snapshot and the run exits with an error, so nothing downstream can publish a half-updated mix of vintages.
2. 02 Validate. Sanity gates run before any snapshot is accepted. Each indicator is measured against the previous snapshot and may not quietly lose more than 5% of its entities; the twenty largest populations may not move more than 5% between runs; and reference values are asserted outright (if India’s population comes back under 1.3 billion, the run stops). A gate that trips aborts the source rather than overwriting good data with bad.
3. 03 Normalize. CIA GEC codes are joined to ISO codes through a hand-verified 261-entity mapping (the CIA’s "ch" is China, not Switzerland). Every value is stored with its source, observation year, and retrieval date. Normalize refuses to run on a snapshot that is missing, hollow, or flagged as failed by the fetch step.
4. 04 Publish. The site rebuilds from the committed snapshots. A page’s "updated" date changes only when its underlying data actually changed, never because the site redeployed. Sitemap lastmod dates follow the same rule.

There is no human approval step in that chain. Once the gates pass, the refresh commits and the site rebuilds without anyone signing it off. We think that is worth stating plainly rather than implying an editorial desk that does not exist. What we do instead is make the machine's work inspectable: the run writes a per-source status file, and every leadership change it found in Wikidata is listed in the refresh job's log so a change of head of state is visible as an event rather than appearing silently. The gates above are the real defence against a bad upstream value, and the source-and-year label on every figure is what lets you check ours against the publisher's.

Known limitations

• 38 of our 254 countries and territories have no live source for population or GDP. Some (the Paracel Islands, Akrotiri, the Coral Sea Islands) have no ISO code and so no join at all; others (Jersey, Vatican City, Western Sahara, the Cook Islands) have a code but sit outside the World Bank's country list. A few of them still pick up UN population-structure or CO2 figures. Where nothing live exists, the profile is the archive and says so.
• Armed forces personnel data has not been updated globally since 2020; we show it with its observation year rather than hide it.
• Observation years differ across countries within the same table. We print per-row years instead of pretending a single "2026" vintage exists.
• Four Factbook entities are parts of a single ISO territory: the Gaza Strip and West Bank both map to PSE, Svalbard and Jan Mayen both map to SJM. Live feeds report the combined territory, so we use the archive's genuine per-part figures and omit these entities from tables where only a combined figure exists.
• Where no live source covers an entity, ranked tables carry its final Factbook value and mark the row archive. Defence spending in dollars has no archived equivalent, since the Factbook reported only a share of GDP, so those rows are computed from that share and marked derived.
• Some crisis-economy official statistics (inflation in particular) are contested; we show the international series and note disputes on field pages.

Found an error? Tell us. We fix what we can verify, and the fix reaches the site on the next refresh that runs after it lands.