Cohorts and retention curves

Play this lesson

A flat line that hides trouble

Shard Garden has had about 20,000 daily players for two months. Kai is happy. Theo, the studio's data analyst, is not sure yet.

A 2010 paper by Google researchers Kerry Rodden, Hilary Hutchinson and Xin Fu points out that a count of weekly active users doesn't separate new players from returning ones. In theory, it could still rise with 100% of players replaced every week.

Joining and coming back

A cohort is a group of players who started at the same time, like everyone who installed in the first week of September. Google Analytics' cohort explorations ask for two rules:

Weekly cohorts run Sunday to Saturday, not any 7 days in a row. Days run midnight to midnight in the property's time zone.

Quick check

Theo wants "came back" to mean the player really played. Which return rule fits best?

  1. Defended a garden
  2. Opened the app
  3. Received a notification
Show the answer

Defended a garden

Opening the app can be a mistaken tap, and receiving a notification takes no action at all. Defending a garden means the player actually played.

Three ways to count a return

These are Google Analytics' names. The same cohort gives three different numbers, so a retention figure always says which one it is.

Quick check

For the same cohort and week, the rolling number can never be higher than the standard one.

Show the answer

True

Rolling counts only players who came back that week and in every week before it. They are a part of the standard group for that week, so the number can only be the same or smaller.

Reading the curve

Theo draws a retention curve: the share of a cohort still playing, week by week since install.

Week 1: 40%. Week 2: 28%. Week 4: 20%. Week 8: 18%. Week 12: 17%.

It falls fast, then flattens. The flat part is players who made Shard Garden part of their week. A curve that keeps falling towards zero means nobody stays for long, whatever the daily total says.

Read the numbers

retention, by weeks since install
Cohort       1     2     4Sep 1      40%   28%   20%Sep 29     44%   33%   26%

The phone co-op invite was fixed on 22 September. What does this table suggest?

  1. The newer cohort keeps more players at every week so far
  2. The September 1 cohort is doing better
  3. Nothing, because the newer cohort has no week 8 yet
Show the answer

The newer cohort keeps more players at every week so far

Cohort tables compare players of the same age, so week 2 is compared with week 2. Something else could have changed too, like school holidays. An A/B test would show whether the fix caused it.

Age, not dates

In a cohort table, each row is a cohort and each column is an age: weeks since install. So one calendar week runs diagonally across the table, hitting a different age in every row.

When the servers struggled one weekend, every cohort dipped on that diagonal. A change in the game shows up down a column instead: every cohort that met it at that age behaves differently.

Did the fix cause it?

Cohorts show that something changed. To prove a change caused it, the studio needs two groups of players, alike in every way except one.

Bolt says: Next: designing an A/B test.

Fact-checked 11 October 2026