Journal · 2 April 2026

Retention curves that tell the truth

A retention curve is a confession. Most of the ones I am handed in United Kingdom product reviews are dressed for a board: a steep first day, a polite flattening, a title that says “healthy.” Cohort Signal Lab exists because that drawing is often a collage of novelty, crashed session clocks, and a notification that woke people who had no job to do.

Truthful curves start with a job, not with “opened.” Plot the share of an install cohort that completed the same human task on day 1, 7, and 30. Then plot the same cohort with the vendor’s default “any event” retention beside it. The gap is the story. If the gap is small, your product may actually be used. If the gap is a canyon, you have been reporting curiosity.

Censoring and late packets

People who installed yesterday cannot have a D30. That is not pessimism; it is censoring. Curves that treat the incomplete as churned look dramatic and wrong. Late events — a phone that uploads a queue when it finds wifi — make D1 look luckier than it was. In the lab we mark a 36-hour late window and refuse to close D1 before it.

What not to smooth

Moving averages hide store-feature spikes and outages. Leave the dents. Annotate them. A truthful curve is a little ugly. The memo can still be short: three numbered remaining rates, the sample size, and one sentence about whether the product still deserves the next sprint.

If you want the lattice around this drawing, sit Cohort Signal Lab. The journal will not replace week four.