Cohort analysis: Cohort analysis.

Overall numbers like DAU and revenue mix new and long-time players together. A big ad campaign can make them rise even while the game is getting worse. Cohort analysis separates players by when they arrived, so you compare like with like.

What is a cohort?

A cohort is a group of users who share a starting event in the same period. In mobile games the starting event is almost always the install, and the period is a day or a week. "The 1 March cohort" means everyone who installed in the week starting 1 March.

You can also build cohorts by:

• Acquisition source: players from one ad network, campaign or creative.
• Country or platform: iOS vs Android, US vs Brazil.
• Behaviour: players who finished the tutorial, made a purchase, or joined a clan.
• App version: players who installed after a particular update.

How to read a cohort table

Each row is a cohort. Each column is a number of days since install. Each cell shows the share of that cohort still active on that day (or their cumulative revenue, if you are tracking money). This snapshot was taken on 7 April:

Install week	Installs	D1	D7	D14	D30
1 Mar	2,000	32%	14%	9%	5%
8 Mar	2,400	35%	16%	10%	–
15 Mar	1,800	30%	12%	8%	–
22 Mar	2,100	38%	17%	–	–
29 Mar	2,200	36%	–	–	–

The empty cells are days those cohorts have not reached yet, which gives the table its triangle shape.

Working the first row. The 1 March cohort had 2,000 installs. 640 came back the day after installing: 640 ÷ 2,000 = 32% D1 retention. 280 were active on day 7: 280 ÷ 2,000 = 14%. 100 were still active on day 30: 100 ÷ 2,000 = 5%.

Reading down the columns. D1 moved from 32% to 35%, dipped to 30%, then rose to 38% and 36%. If an update shipped on 21 March, the two cohorts after it are your first evidence that it helped, and the 22 March cohort's D7 of 17% backs that up. The 15 March dip is worth checking too: maybe a new ad campaign brought in lower-intent players that week.

Cohort analysis for revenue

The same table can show cumulative revenue per install instead of retention. That is how teams measure LTV and ROAS:

Cohort revenue per install on day N = total revenue from the cohort up to day N ÷ installs in the cohort

If the 1 March cohort has earned $1,300 by day 30, its D30 revenue per install is $1,300 ÷ 2,000 = $0.65. Compare that with what you paid per install (eCPI) for the same cohort to see whether the campaign is paying back.

Why cohorts beat totals

• They remove growth effects. Total DAU can rise just because you bought more installs. Cohorts show whether each player is staying longer.
• They show cause and effect. Changes line up with the cohorts that experienced them.
• They make channels comparable. Two ad networks with the same install cost can produce very different cohorts.
• They feed forecasts. Predicted LTV is built from early cohort curves.

Common mistakes

• Cohorts too small. A cohort of 80 players swings wildly. Group by week, or wait for more installs, before drawing conclusions.
• Mixing classic and rolling retention. Pick one definition and keep it. See retention.
• Comparing cohorts of different ages. A week-old cohort has no D30 data yet. Compare at the same day since install.
• Ignoring the mix. A cohort full of players from a cheap, low-intent channel will look worse even if the game did not change. Split by source when you can.

Tools for cohort analysis

Most game analytics tools, product analytics tools and MMPs build cohort tables for you. The important part is sending clean events from the game: install, session start, level progress, ad impressions and purchases, each with a timestamp and a user ID.