The average that hid a bimodal site

The average that hid a bimodal site

A content site reported a healthy 3-minute average session and patted itself on the back. The number was real and useless. The shape underneath was the story.

Averages get yanked around by a long tail. The clue appeared the moment we pulled the distribution instead of the mean: two clusters, not one. A big group bounced in under 15 seconds, a smaller group stayed eight minutes plus, and the average split the difference into a number nobody actually experienced.

The change: we segmented by landing template and found one programmatic page type generating most of the sub-15-second crowd.

Fixing that template's above-the-fold content lifted its median time-on-page from 19 seconds to 1 minute 50.

When behavior is bimodal, the average describes a visitor who doesn't exist. Pull the histogram before you trust the mean. The median would have warned you first.