Research

Reports

How these reports are produced

The query is published before the number is

One read-only SQL file produces every figure, it runs inside a transaction the database refuses to let write, and the file itself is downloadable beside the data. The window, the exclusions and both suppression floors are applied in that query, not afterwards in a spreadsheet. The query also emits its own reconciliation checks, and the site refuses to build if the arithmetic a reader would do by hand does not come out.

Nothing that identifies a person or an app leaves the database

No app names, no bundle identifiers, no accounts, no per-account rows. Countries are published as a count of countries, never as a table a single build could be picked out of.

Small buckets are suppressed, not rounded

A published bucket has to clear two floors, not one: at least 50 rows and at least 5 distinct owners, where an owner is the account behind an upload or, for an upload made without one, the device that made it. A bucket that fails either is absent and its residual is stated instead, which is why those breakdowns often do not sum to the population total, and each chart says so where the reader is looking. There are no exemptions: the 2026 report withholds its upload-source split in full, because publishing the half that clears the floors would recover the other half by subtraction.

The population is what happened, not what survived

A report counts everything that occurred inside its window, whether or not the row was still live when the query ran, and whether or not the person who made it had signed in. Edition 2026-09 broke that rule in both directions and was corrected; the correction is marked on the report itself rather than folded in quietly.

A number that cannot be published honestly is withheld

Where a measurement describes the instrument rather than the world, it is left out and the reason is printed in its place. The Android tester OS split in the 2026 report is the worked example.

The figures are free to reuse with attribution under CC BY 4.0. Link to the report you are citing, and if a figure looks wrong, say so through the contact page.

Where the data comes from

BetaDrop distributes iOS and Android builds to testers over the air: upload an .ipa or an .apk, share one install link or a QR code, and testers install from a browser without accounts. The reports describe what happens on that service, which is also the boundary of what they can claim.