The report that changed its mind on every refresh

The report that changed its mind on every refresh

A high-traffic publisher kept getting different revenue-by-source numbers each time they reran the same report. The data wasn't broken. It was sampled.

When a date range is huge, many tools estimate from a slice of sessions instead of counting them all, so rare, high-value segments wobble between refreshes. The clue: a tiny but lucrative affiliate source jumped 40% between two identical queries.

The change: we exported the heavy report through an unsampled path, narrowing the date range and pulling the raw figures rather than the estimate.

The affiliate source settled at a stable value, and the variance between reruns collapsed from 40% to under 2%.

Sampled reports are fast and fine for big rounded trends. Unsampled is slow and necessary when a small segment carries real money. If your numbers change on refresh, you're reading an estimate and treating it like a count.