Per-variant time series

Almost everything on a results screen is cumulative: one p-value, one lift, one interval, all computed over the whole run. None of them can show whether a result is settling down or still moving around, and that is often the thing you most want to know.

The time series gives you the daily counts per variant instead: sessions, conversions and revenue, day by day. Rates are deliberately not computed for you, because the raw counts are what let a cumulative view and a per-day view be built from the same data, and what let the evidence be recomputed at each point rather than only at the end.

It is also how a weekday effect becomes visible. A lift that exists on weekdays and reverses at weekends is a real finding about your customers, and it is completely invisible in a single pooled number.

One honest limitation the console states rather than hides: session history begins where the hourly rollup that feeds it began, because the underlying rows are deleted once processed and there is nothing to backfill. The response says when the series starts, so an empty early chart is explained rather than read as no traffic.