What is cohort analysis?
Cohort analysis follows groups that share a defined characteristic and examines their behavior over time. For a signup cohort, group membership begins with a qualifying signup. Compare a later action at the same elapsed age, using an explicitly stated population and observation window.
Define membership and the action you will count
Write the inclusion rule before building the table. Identify the unit, such as registered users or customer accounts, the qualifying event and the signup period. Decide whether someone can enter more than once. A user-count report and a revenue report require different numerators, even if their rows use the same cohort names.
Next, define the later action: a return visit, another completed purchase or continued paid status. Keep that condition with the report. Returning to an app during a week does not establish that a customer renewed a subscription.
Cohorts and segments are not exclusive categories. You can follow a channel segment over time or split a signup cohort by a recorded attribute. State when the attribute was measured and whether membership stays fixed, rather than relying on the label to explain the analysis.
A hypothetical return-rate table
Imagine two signup cohorts with invented counts. Each registered user enters once at their first qualifying signup. Assume stable unique identifiers and complete event records within the observed window. A return means at least one chosen activity during a column's interval. Count each original user once per cell, regardless of repeated events.
This worksheet uses elapsed 24-hour days from each user's signup. “Days 7–13” starts after 7 full days and ends before day 14; “Days 14–20” starts at day 14 and ends before day 21. These are this example's interval rules, rather than a reproduction of a provider's calendar report.
Assume every member of the older cohort has 21 full days of follow-up, while every member of the newer cohort has 14. This worksheet withholds a whole cell until all its members have completed that interval. Divide the returning-user count by the original cohort size to calculate each displayed percentage.
| Signup cohort | Starting users | Days 7–13 | Days 14–20 |
|---|---|---|---|
| Older | 100 | 35 / 100 = 35% | 20 / 100 = 20% |
| Newer | 80 | 32 / 80 = 40% | Not yet complete |
Read across the older row to compare return activity at different ages. Read down the Days 7–13 column to compare cohorts at the same age: the hypothetical difference is 5 percentage points. A person can appear in both columns, so adding the row's percentages would not give a distinct-user total.
Keep incomplete follow-up separate from zero
The newer cohort's final cell is unfinished. Treating it as zero would turn unobserved follow-up into an apparent lack of returns. Wait for the interval to finish under the worksheet's rule. A fully observed interval with no qualifying returns is a different state; an empty starting cohort has no calculable return rate.
Real reports also need their own completeness rules. Amplitude's retention FAQ explains that recent incomplete periods receive an asterisk and distinguishes them from completed observations. Check how your selected report handles those periods before interpreting its totals or averages.
Check which clock and population a tool uses
Google's GA4 cohort documentation describes inclusion and return criteria. Its daily buckets use the property timezone; its weeks run Sunday to Saturday. It also says cohorts use device data without User-ID. Those conditions differ from the registered-user, elapsed-time convention in the fictional worksheet.
Amplitude's timing documentation distinguishes rolling 24-hour windows from strict calendar dates. Record the convention used in your export before comparing it with another tool. Matching column labels alone does not establish matching observation periods.
Also check whether a cell requires an action in that interval, in every preceding interval, or at any time up to it. Keep the counting rule explicit. The fictional table above uses an independent return condition for each interval; someone can miss an earlier interval and return in a later one.
Use a difference to choose the next investigation
A cohort comparison can identify a pattern worth examining. It does not isolate the effect of a product release or marketing channel. Check changes in audience mix, signup conditions, tracking and follow-up before crediting a new onboarding flow with the difference. A release annotation is context for that investigation, not proof of its result.
Keep counts beside rates so readers can see the population behind a comparison. Avoid replacing that context with a universal “good” curve or minimum-size rule. Decide what uncertainty the decision can tolerate and what additional evidence would change the action.
Distinguish activity from customer loss and value
For customer lifetime value, distinguish observed revenue within a cohort's available history from a forecast beyond it. A younger group has had less time to accumulate recorded purchases. Compare compatible ages and value definitions before using that difference in a budget discussion.
Keep the measurement rules with the handoff
Save the inclusion rule, identity unit, return event, time convention, original sizes, observation cutoff and treatment of unfinished cells. Name the person responsible for checking event collection. Give the next reviewer the underlying counts and a specific question to investigate, rather than a conclusion that a launch already worked.
Sources and scope
Google documents its own cohort exploration. Amplitude documents its retention report's timing and incomplete-data behavior. The numerical worksheet uses fictional inputs and its own declared rules; it is not an exported tool report.