Series: Measuring what matters
- Part 1: proxies and their failure modes
- Part 2: session length is a trap
Session length is the classic inverted proxy: when a task gets harder, sessions get longer. Averages make this worse, because the distribution is heavily skewed.

Look at the shape first
6.4
mean minutes
4.9
median minutes
23.1
99th percentile
A handful of very long sessions drag the mean upward. When those sessions turn out to be people stuck in a broken checkout, “engagement is up” is the opposite of good news.
Better questions
- Did people complete what they started?
- How long did completion take, for those who completed?
- Where did the long sessions spend their time?