How to read a retention curve

What is a retention curve?

A retention curve is one row of a cohort table plotted as a line, with Day N on the horizontal axis and the share of the cohort counted as retained on the vertical axis.

Cohort: users grouped by install day. The install day is Day 0, and Day 0 activity is not a return.

Retention curve: the cohort's Day 1, Day 2, Day 3 and later values drawn as one line, or several cohorts pooled into one line. There is a short definition in this glossary entry on the retention curve.

If you are not yet comfortable with the table the curve comes from, start with how to read a cohort table, then come back.

The three parts of a classic curve

Most classic retention curves have the same three parts.

1. The early drop. A large share of users never return after Day 0, so Day 1 is already well below 100%, and the next few days fall further.
2. The bend. The fall slows as the users who were never going to stay drop out.
3. The flattening point. The curve levels off. The users left are the ones who have built a habit, and this level is what drives long-term value.

In the sample app on the home page, a made-up app, pooled classic retention falls from about 35% on Day 1 to about 18% on Day 7 and about 16% on Day 14. Most of the loss happens in the first week, and the curve is close to flat by the second. The flattening section of the reading guide shows how to spot this in the table itself.

Why does a retention curve go up again?

A curve that rises after falling, often called a smile curve, usually means something happened on a calendar date, such as a release, a push notification or a reactivation campaign, rather than users suddenly liking the app more. Check the diagonal of the cohort table to confirm.

A calendar event touches every cohort on the same date, which is a different Day N for each cohort. In the cohort table it shows up as a diagonal stripe. In the sample app, the 3 Sep cohort shows about 38% classic retention on Day 7, against roughly 13% to 20% for its neighbours. That Day 7 falls on the same date as a made-up push, so the lift belongs to the push, not to a better cohort. The diagonals section walks through the same example.

A genuine, lasting rise in the tail is possible, for example when a product change makes old users come back. Treat it as a hypothesis until you have seen it across several cohorts and dates.

The definition decides the shape

The same users produce three different curve shapes depending on which retention definition you plot.

• Classic counts users active on exactly Day N, so it can fall or wobble.
• Cumulative counts users active at least once from Day 1 to Day N. Each day can only add users, so the line never falls. A rising cumulative curve is not a sign of improving engagement.
• Rolling counts users active on Day N or any later observed day. On the last observed day there is only one day left to count, so rolling equals classic there and the tail collapses onto the classic line.

If your curve looks wrong, check the definition first. The cohort explorer lets you switch the same sample data between all three to see the shapes change.

Comparing curves fairly

Only compare mature days

A cohort has a Day N value only once Day N has fully passed for it. A cohort that installed last week has no Day 14, and its rolling values are still rising. Compare two curves only up to the last day both have reached.

Small cohorts wobble

A cohort of a few dozen users can swing several points from one day to the next because one or two users came back. Read small cohorts as directional, pool several days of cohorts where you can, and see the sample size notes before reading meaning into a bump.

Use the same definition and day boundary

Published benchmarks usually use classic, exact-day retention, so compare like with like. Check that both curves use the same definition, the same time zone for day boundaries and the same event for "active".

For a longer walkthrough with more curve shapes, see this guide to reading a retention curve. If you are also tracking what users do on Day 0, this overview of app onboarding metrics pairs well with the early part of the curve.