How to Calculate User Retention: Cohorts, Day-N Rates, and Time Windows
Published:
Last Updated:
Category: Marketing Glossary
Authors: Shusaku Yosa
User retention is the proportion of a defined starting cohort that performs a specified return action within a specified observation window. Freeze the entry condition, return event, identity rule, and time window before calculating it. A label such as “seven-day retention” is incomplete unless readers know which returns qualify.
Define the behavior you want to measure
Returning after registration and receiving recurring product value are different questions. A login may be enough for one analysis; completing a creation, reading, or purchase action may suit another. Choose an event that matches the question and the product's natural usage cycle. A service used monthly should not automatically treat the absence of daily activity as failure.
Specification | Decision | Fictional example |
|---|---|---|
Entry condition | First or repeated entry; eligible population | First completed creation |
Return action | The event that demonstrates return | Another completed creation |
Identity | User, account, or another unit | A unique user ID |
Time boundary | Day 0, calendar boundaries, time zone | Entry date is Day 0; consistent calendar days |
Cutoff | How far each user has been observed | Every user observed through Day 60 |
If you want to track valid contracts instead, use customer retention based on an opening customer cohort. Missing an activity event and ending a contract are not the same state.
Compare exact-day, interval, and later returns
Amplitude's retention calculation documentation distinguishes Return On from Return On or After. Rather than relying on labels such as classic or rolling, record the actual time window used by your tool.
The following fictional example follows 100 users through Day 60. Each enters once, the return event is fixed, and returners are counted once within each applicable window.
Definition | Qualifying window | Unique returners | Result |
|---|---|---|---|
Exact Day 7 | Day 7 only | 30 | 30 / 100 = 30% |
Days 7–30 | At least one return from Day 7 through Day 30 | 45 | 45 / 100 = 45% |
On or after Day 7 | At least one return from Day 7 through the Day-60 cutoff | 50 | 50 / 100 = 50% |
These are nested sets in this example: the 30 exact-day returners are included in the 45 interval returners, who are included in the 50 later returners. “On or after” is not a measurement of an infinite future. Extending the cutoff can change its result.
Align Day 0 and day boundaries
A rolling 24-hour window differs from a calendar day. Someone who starts at night and returns the next morning has crossed a date boundary without waiting 24 hours. Similarly, a 30-day interval and a calendar month are different periods.
Amplitude's time-window documentation describes how elapsed windows, calendar dates, and project time zones affect its calculations. These are that product's rules, not universal defaults. Check repeated entry events, timestamp rounding, and week boundaries before comparing reports.
Do not count immature cohorts as zero retention
Suppose fictional cohort A has 100 users observed through Day 7, with 30 returning that day. Cohort B has 80 users observed only through Day 3. The available Day-7 result is 30 / 100 = 30%. Calculating 30 / 180 = 16.67% incorrectly treats users who have not reached Day 7 as non-returners.
Distinguish observed cells with zero returners from cells that are not yet observable. Give cohorts equal observation opportunities when comparing interval retention. For on-or-after retention, older cohorts also have more time to return later; define a common follow-up horizon or clearly disclose the difference.
Aggregate people, not percentages
In a separate fully observed fictional example, one cohort has 30 returners out of 100 and another has 80 out of 200. Combined retention is 110 / 300 = 36.67%, not the unweighted average of 30% and 40%, which is 35%. Sum eligible users and returners first.
- Confirm identical entry, return, and identity definitions.
- Align elapsed observation time and check missing events or consent-related measurement gaps.
- Separate changes in plan, acquisition source, or release exposure.
- Investigate reasons for non-return, then specify a hypothesis, intervention, cost, and review horizon.
Exact-day usage retention does not automatically equal 100% minus subscription churn, and a higher rate alone does not establish satisfaction or a causal intervention effect. Use KPI analysis and review procedures to connect the measurement to a decision.
