How to read a cohort retention table

What do the rows and columns mean?

A cohort retention table puts install days down the side and days since install across the top. Each cell is the share of that row's users counted as retained on that day, under one definition.

• Across a row: one cohort ageing, from Day 1 onwards.
• Down a column: different cohorts at the same age, the fair comparison.
• The triangle: newer cohorts have not reached later days yet, so those cells are empty, not zero.

What does a bright diagonal mean?

Cells on a diagonal share one calendar date. In the sample app a feature release and a push notification went out on 10 Sep, and every cohort's classic retention jumps on whatever day 10 Sep was for it: the 3 Sep cohort on Day 7 (38%), the 9 Sep cohort on Day 1 (52%).

Classic shows calendar events most clearly. Cumulative absorbs them into every later day, and rolling spreads them backwards across earlier days, so check the classic view before you credit a cohort.

When does a retention curve flatten?

Most curves drop fast in the first days and then level off at the users who have made the app a habit. Where the curve flattens matters more than the first-day drop: a curve that keeps falling has no stable core yet.

In the sample, classic falls from about 35% on Day 1 to about 18% on Day 7, then barely moves. Note the later points rest on fewer cohorts, so they are less certain.

How should you compare paid and organic cohorts?

Split the same table by channel and compare on the same definition, the same Day N and mature cells only. In the sample app, organic users return more on every day: classic Day 1 is about 46% organic and 26% paid.

A gap is normal: organic users usually arrive with intent. What matters is whether the gap is stable, and whether paid users who do return behave like organic ones. The guide on why paid Day 1 is lower than organic lists what to check.

Why can blended retention fall when no channel got worse?

From 6 Sep the sample app scaled up paid installs. Each channel's classic Day 1 stayed flat, but because paid users return less and now make up most of the installs, the blended number dropped. This is the mix-shift trap, a form of Simpson's paradox.

How much can you trust a small cohort?

Small cohorts swing. In the sample, paid cohorts before 6 Sep had about 50 installs each, and their classic Day 7 ranges from 2% to 24% from one day to the next with nothing changing underneath. After 6 Sep, with about three times the installs, the paid range narrows to 9% to 14%.

Pool several days into weekly cohorts, or wait for more installs, before reading single cells. If one returning user would change a cell by several points, treat it as directional.

Why do newer cohorts look worse on rolling retention?

Rolling counts every day from Day N to the last day in the data, so older cohorts have more chances to be counted. In the sample, rolling Day 1 is 73% for the 1 Sep cohort and 51% for the 10 Sep cohort, while classic Day 1 for those two is 34% and 31%.