The thirty days that are not thirty days
A statistics page here is not one table. It is twelve, one per question — countries, pages, browsers, hours, devices, operating systems, languages, regions, referrers, bots, screen resolutions, and the daily totals. You pick a period at the top, and every panel answers for that period.
Except the panels cannot all answer for the same period, because the tables behind them do not go back equally far. Counted on 26 August 2026:
| table | rows | earliest day | span |
| daily totals | 113,604 | 15 Jul 2025 | 408 days |
| countries | 318,161 | 1 Jan 2026 | 238 days |
| pages | 101,485 | 29 Jun 2026 | 59 days |
| browsers | 48,473 | 1 Jul 2026 | 57 days |
| hours, languages, systems | 85,699 | 9 Aug 2026 | 18 days |
| regions | 21,227 | 13 Aug 2026 | 14 days |
| devices, referrers, bots | 19,274 | 14 Aug 2026 | 13 days |
| screen resolutions | 0 | — | — |
The query behind each panel takes the range you selected and asks for it. The table answers with the days it happens to have. Nothing in between says the range was not covered. Ask the device panel for ninety days and it will answer with thirteen, confidently, with no footnote.
Why they differ
Not retention. The cleanup tool keeps aggregates for 400 days and only touches them when explicitly asked with a flag; the daily table currently holds 408 days, eight past that limit, untouched.
They differ because they were added at different times. Countries came in January. Pages and browsers came at the end of June. Hours, languages and systems arrived on 9 August; regions on the 13th; devices, referrers and bots on the 14th. Each was a separate small feature, each correct on its own, and none of them knew about the period selector sitting above all of them.
What this means when you read your own numbers
The practical rule is simple and it costs nothing to follow: never divide a number from one panel by a number from another. "Mobile share of my visitors" computed as device hits over total hits is not a share of anything — it is thirteen days over four hundred and eight. Ratios are only meaningful inside a single panel, where every number came from the same table over the same days.
The empty row at the bottom of that table is a different story, and a deliberate one: screen resolution is no longer collected at all. That panel will stay empty on purpose.
The honest summary is that a period selector spanning twelve tables of different ages promises more than it can deliver, and says nothing when it under-delivers. Knowing which panels are young is, for now, the whole defence.