Cohort Analysis: Retention Explained for Designers
Cohort analysis groups people by signup date and tracks their retention. How to read the table, what the curve shape says, and which mistakes mislead you.
What is a cohort?
A cohort is a group of people who share a trait and a starting point, for example everyone who signed up in the first week of March. Amplitude describes it as grouping users by shared characteristics and following each group over time (Amplitude, “The complete guide to cohort analysis”).
A cohort’s retention is the percentage of its members who are still active in each period. If forty of a hundred people who signed up in March come back in the second week, the retention for that week is 40%. Koji frames it the same way: a group with a common starting point and the percentage still active (Koji, “Cohort analysis guide”).
Why should a designer care?
Because single-day metrics can mislead. A month with many signups can look excellent even if people leave within a week. Cohort analysis separates growth from retention: it shows whether the people arriving today stay better than the ones who arrived three months ago.
That lets you test changes with judgment. If you redesigned onboarding, the question isn’t whether signups went up, but whether the cohorts after the change retain more. And if the redesign helps, you should see it in the month-one column, not only on day zero.
How do you read the table?
The table has cohorts in rows and periods since signup in columns. Each cell shows the percentage still active. There are three ways to read it:
- Across a row, to see how one specific cohort decays over time.
- Down a column, to compare newer cohorts with older ones at the same age. This reading answers whether the product is getting better.
- Diagonally, to compare different cohorts at the same point in the calendar, for example whether a March campaign brought people who left sooner.
Amplitude recommends looking for sharp drop-offs, unusual cohorts, and trends, and warns that the chart by itself changes nothing unless it is tied to an action.
What does the shape of the curve say?
CRV describes four typical shapes:
- Decline to zero. Users leave and don’t come back. The product doesn’t yet have a use that lasts.
- Shallow decline to a low floor. There is some retention, but not much. Check how long it takes a person to get value.
- Plateau. The curve stabilizes at a level. That height is the product’s real retention, and a sign of product-market fit.
- Smile. The curve rises again, often because of a campaign or a launch.
Koji describes three forms: a decline to zero, a plateau above zero, and a rising curve. The key reading is the plateau: if it appears, a group of users is staying.
Example
Illustrative example: a habit-tracking app signs up 1,000 people in January and 1,200 in February. January’s cohorts retain 22% in the second week; February’s retain 31%. The team credits the new three-step onboarding. Before celebrating, it checks two things: that February’s campaign didn’t bring a different kind of user, and that retention keeps rising in month one. If month one doesn’t improve, the difference might be only in the first week. The decision is made with both columns in front of the team, not just the first one.
What mistakes make the reading misleading?
- Averaging cohorts. An overall average can hide that recent cohorts drop faster. CRV warns that blended numbers flatten out the differences.
- Changing the retention definition. Using a trivial action, such as opening a notification, inflates the curve. The retention action should be the one that measures value.
- Comparing cohorts of different ages. A cohort from a month ago has had less time to show its retention than one from a year ago.
- Using very small cohorts. With few accounts, a small variation looks like a trend. CRV suggests distrusting groups of fewer than about 30 accounts.
- Mixing channels. People who arrive through ads and people who arrive through referrals don’t retain the same way. Segment before drawing conclusions.
Benchmarks: use them with caution
Benchmark reports are useful for context, but they shouldn’t be the target. According to Amplitude’s product benchmark report, the median product retains only 3.8% of its users at month three, with large differences by sector: in travel and hospitality the top performers retain 25.6%, and in healthcare the 90th percentile reaches 10.6% (Amplitude, “The complete guide to cohort analysis”). It is a reference from an analytics vendor, and a figure like this depends on the sector and on how “active” is defined. The most valuable comparison is with yourself, cohort against cohort.
When NOT to use cohorts
- With a low volume of signups. If few people join each month, every cohort is noise.
- To measure satisfaction. Retention says someone comes back, not that they are happy. Ask if you want to know the latter.
- As a substitute for understanding why. The table shows how long people stay, not why they leave. For that, combine it with user interviews or session recordings.
- If the cohort definition changes every week. Without a stable definition, there is no curve to compare.
Common mistakes
- Looking only at day-one retention and celebrating.
- Changing the window or the retention event without saying so.
- Reading a plateau as success without looking at its height.
- Comparing cohorts with different filters.
- Measuring retention for a product that doesn’t yet have a clear moment of value.
How to start tomorrow
Define your product’s activation event, the action that shows someone got value. Group users by signup week in your tool, for example Amplitude or Mixpanel, and measure how many return in weeks one, two, and four. Compare the last four cohorts with each other, not against a benchmark. If retention improves after a design change, note it alongside the funnel analysis of that flow to see whether the step you changed explains the difference. And check the team’s KPIs to make sure retention is among them.