What is a cohort in game analytics?

A cohort is a group of players who started in the same period, usually the same day or week. The cohort keeps its members for life. Each later period shows the share of that group who came back. Day 0 or week 0 is always 100%, because everyone in the cohort played then.

Totals hide change. If DAU stays flat while new players arrive from a sale, the old players may be leaving at the same rate. A cohort table separates the two: each row is one starting group, and you can see whether newer rows keep more players than older rows.

Sources: Manalens retention documentation; Mixpanel Retention documentation.

Which cohorts should a game developer compare?

Start with install date, then cut by build, storefront and source. Each cut answers a different question. Keep the entry event the same in every comparison, so that the rows mean the same thing.

Five ways to cut cohorts and how to set each one up.
Cohort byQuestion it answersHow to set it up
First launch day or weekDo newer players stay longer than older ones?Entry event game_open, Daily or Weekly period
Build or app versionDid the patch help?App filter: Build is 571
StorefrontDo Steam and itch.io players behave differently?App filter: Storefront is steam
Acquisition sourceWhich channel brings players who stay?Dashboard filter on channel, or a saved segment
First actionDo players who finish the tutorial stay longer?Entry event tutorial_complete instead of game_open

Sources: Manalens retention and app events documentation.

How do I read a cohort retention table row by row?

Read each row from left to right to follow one group of players, then compare the rows. The table below is an example with invented numbers. It is not a benchmark and not data from a real game. It shows weekly recurring retention: a player counts in a week only when they play in that week. In the example, patch 1.1 ships on Monday 19 October and a 20% sale runs in the week of 26 October.

Next, read down a column to find where players stop. In every row, the largest loss is between week 0 and week 1. That points to the first sessions, not to late content. Last, look for a flat tail. When a row stops falling, you have found the players who stay. A bigger flat tail is worth more than a higher week 1.

  • Row 1, week of 5 October: the baseline. 18% return in week 1, which is 72 of 400 players. Week 2 is 11%, and week 2 for this row is the patch week.
  • Row 2, week of 12 October: week 1 is 21%, above the baseline. Week 1 for this row is also the patch week, so the patch brought some older players back.
  • Row 3, week of 19 October: these players started on patch 1.1. Week 1 is 24% and week 2 is 13%, both above rows 1 and 2. This is the row that tests the patch.
  • Row 4, week of 26 October: the sale cohort is twice as large, and its week 1 is 15%. That is a lower rate but more players: 135 of 900, against 101 of 420 in row 3.
  • Row 5, week of 2 November: week 1 is 23%, close to row 3. The gain from the patch holds for players who arrive after the sale.
  • The dashes: those weeks are not complete, so the table shows no number for them.
Example weekly cohort retention (recurring). All numbers are invented. A dash is a week that is not complete yet.
Example cohortPlayersWeek 0Week 1Week 2Week 3Week 4
Week of 5 October400100%18%11%8%6%
Week of 12 October380100%21%10%7%5%
Week of 19 October (patch 1.1)420100%24%13%10%-
Week of 26 October (20% sale)900100%15%7%--
Week of 2 November410100%23%---

Sources: Manalens retention documentation; Mixpanel Retention documentation; GameAnalytics Retention documentation.

What does a patch do to a cohort table?

A patch can change two different things, and a cohort table keeps them apart. The first is the row effect: players who start on the new build get a new first session, so their whole row can move. In the example, row 3 is higher than rows 1 and 2 in every week.

The second is the calendar effect: a patch gives older players a reason to come back in one week. That shows as a diagonal, because one calendar week is week 2 for one row and week 1 for the next. In the example, the patch week is the 11% in row 1 and the 21% in row 2. GameAnalytics notes that PC players return for updates, expansions and events, so expect this on PC.

Judge a patch by the row effect. The calendar effect fades in a week or two. A chart of DAU mixes both, and a cohort table does not. Recurring retention shows the diagonal, because its curve can go down and come back up. Rolling retention hides it.

Sources: GameAnalytics Benchmarks Hub; Manalens retention documentation.

What does a sale do to a cohort?

A sale makes the cohort larger and changes who is in it. Steam sends a wishlist notification when a discount is 20% or more, and Valve notes that many users wait for a reason to buy, such as a discount. So a sale cohort holds more players who were less sure about the game.

Expect a lower retention rate and a higher count. In the example, the sale row keeps 15% in week 1, against 24% in the row before it, but 135 players return against 101. Do not read the lower rate as a worse game. Compare the sale cohort with an earlier sale cohort, or split it by source.

Refunds add noise in the same weeks. Steamworks says the weekly refund rate can be volatile after a launch and after discounts, and that you are likely to see higher rates after a discount, when sales have dropped back but some sale buyers ask for refunds.

Sources: Steamworks Wishlists documentation; Steamworks Developer Refund Reporting.

Should I use rolling or recurring retention?

Use recurring retention to see the habit, and rolling retention to see who is still around. Recurring retention counts a player in period N only when they return in exactly that period, so the curve can go down and come back up. Rolling retention counts a player who returns in period N or any later period, so the curve can only go down.

Rolling retention is always the same as or higher than recurring retention for the same players. For a premium game that people play on some weekends, weekly rolling retention gives a fairer picture. For D1, D7 and D30 benchmarks, use strict day N numbers. Other tools use other names: Amplitude calls the two Return On and Return On or After, and Mixpanel calls them On and On or After.

Sources: Manalens retention documentation; Amplitude retention documentation; Mixpanel Retention documentation.

How do I test whether a patch improved retention?

Compare the cohort that started on the new build with the cohort that started just before it, on the same day N, with enough players in each. Change one thing at a time when you can.

  1. Set the build option in the SDK init call, for example build 571, so every event carries the build.
  2. Ship the patch and note the date.
  3. Add a Retention chart with game_open as the entry event and session_start as the returning event, Daily period.
  4. Filter the chart by the old build, then duplicate it and filter by the new build.
  5. Wait until day 7 is complete for the new cohort. Incomplete periods show a dash, not a low number.
  6. Compare D1 and D7. Check that both cohorts have a few hundred players before you trust a small gap.
  7. Check the acquisition mix of both cohorts. A sale or a creator video can change who arrives, and that alone moves retention.
  8. Test the gap with the A/B test significance calculator before you call it a win.

Sources: Manalens retention guide.

How do I build a cohort table in Manalens?

Add a Retention chart to a dashboard and choose the entry event, the returning event, the retention type and the period. The Game KPIs template already has one: Install cohorts, with game_open as the entry, session_start as the return, a Daily period and recurring retention. The table lists what the chart does today and what it does not do.

Cohort retention in Manalens, checked against the product documentation on 19 November 2026.
What you wantSupportedDetail
Weekly cohortsYesWeeks 0 to 12. A week starts on Monday, in UTC.
Daily cohortsYesDays 0 to 30, which includes D1, D7 and D30.
Your own entry and returning eventsYesAny event, a page, or a custom event such as tutorial_complete.
Rolling and recurring retentionYesNew charts start as rolling. Choose Recurring for benchmarks.
Cohorts by build, storefront or platformYesThe App group in the filter menu.
Cohorts for one group of playersYesChoose a saved segment in the chart.
Incomplete periodsYesThey show a dash. Hover a cell for the player count.
Monthly cohortsNoOnly weekly and daily periods.
Periods after week 12 or day 30NoThe chart stops there.
Custom day brackets, such as days 1 to 3NoUse a weekly period.
A marker for a patch or a sale dateNoKeep the dates in your notes.
A significance test in the chartNoUse the A/B test significance calculator.

Sources: Manalens retention, dashboards and app events documentation; the Game KPIs board template.

How do I compare cohorts by acquisition source?

Filter the retention chart by channel, or by a saved segment of players from one campaign. Install attribution links a first launch to a recent website visit from the same network, and the Installs tab of the Attribution page shows the channel, campaign and creator. A player who opens your website from the game with a link_url() link is joined without a match.

Source cohorts answer the most expensive question in marketing: which channel brings players who stay. A creator video that brings 300 installs with high D7 can be worth more than an ad that brings 1,000 installs who leave on day 0.

Sources: Manalens install attribution and Godot SDK documentation.

What mistakes break cohort analysis?

Most errors come from mixed or tiny groups. Watch for these before you draw a conclusion.

  • Editor and test runs in the data. Set enabled to false in the editor, or use a separate test project.
  • Cohorts that are too small. One extra return in a cohort of 50 moves the rate by 2 points.
  • Periods that are not complete yet. A cohort from yesterday has no D7, and Unity notes that a D30 chart needs 30 days to be complete for an install date.
  • Two storefronts or two builds in one row. Filter by storefront and build.
  • A changed definition. If the entry event changes, old and new rows are not comparable.
  • A sale cohort compared with a normal week. Compare a sale with a sale.

Sources: Unity Analytics dashboards documentation; Manalens retention and Godot SDK documentation.

How do I build cohort charts for my game?

Manalens is the analytics and growth engine built for games. Install the Unity or Godot SDK, then add Retention charts with your own entry and returning events, and filter them by build, storefront, channel or segment.

Sources: Manalens product documentation.

Frequently asked questions

What is the difference between a cohort and a segment?

A cohort groups players by when they started, such as the week of their first launch. A segment groups players by what they are or did, such as players from Steam or players who bought DLC. You can combine them: a retention chart groups by cohort and keeps only the players in a segment.

How many players does a cohort need?

There is no fixed number, but small cohorts move a lot by chance. In a cohort of 50 players, one extra return changes the rate by 2 points. Use weekly cohorts when daily cohorts are small, and wait for a few hundred players before you trust a gap of one or two points.

Why does my newest cohort show a dash?

Because the period is not complete yet. Manalens shows a dash for a period that has not ended at the end of the date range, instead of a low number that looks like churn. A cohort that started yesterday has a D1 tomorrow and a D7 next week.

Why did old cohorts go up in the week of my patch?

Because the patch gave old players a reason to return in that calendar week. In a recurring retention table this shows as a diagonal: the same week is week 2 for one cohort and week 1 for the next. It is a calendar effect. To judge the patch, read the cohort that started on the new build.

Can I build cohorts from an event other than the first launch?

Yes. Choose any custom event as the entry event, such as tutorial_complete or level_complete. Then each cohort is the players who first did that event in the period. This shows whether players who reach a milestone stay longer than players who only launch the game.

Sources and implementation