Cohort analysis compares groups of customers or users who share a defined starting event, such as their first purchase or signup week, and follows the same groups over time. It helps a team see whether behavior is improving for newer groups without mixing them with older customers who have had more time to return, renew, or leave.
How do you define a useful cohort?
Begin with the decision and the event that puts someone into a group. A subscription business might group customers by first paid month; a marketplace might use first completed transaction; a product team might use first successful activation. The event needs a consistent definition. If some rows start at account creation and others start after purchase, the comparison will be misleading.
Choose a time unit that fits the natural behavior. Daily cohorts can help with a high-frequency product, while monthly cohorts may suit a longer purchase cycle. Google Analytics defines a cohort as users sharing a common characteristic and its cohort reports group users by a selection criterion, then follow them over daily, weekly, or monthly periods. A reporting tool does not choose the business question for you.
- Write the entry event and eligibility rule.
- Choose daily, weekly, or monthly intervals that match the customer cycle.
- Select one outcome, such as an active user, repeat purchase, or renewal.
- Keep the definition and observation window consistent across rows.
How do you read a cohort table?
Rows usually represent the cohort's start period and columns show the time elapsed since that start. The first column is the cohort size or its initial activity. Later columns show how many or what percentage met the selected return condition. Compare the same elapsed period across rows: week-four retention for a January cohort belongs beside week-four retention for a February cohort, not beside February's week one.
Recent cohorts have not had time to fill later columns, so blank cells are not zero retention. Use percentages when cohort sizes differ, and retain the underlying counts so a small group does not look more conclusive than it is. Investigate instrumentation, seasonality, acquisition source, and product changes before assigning a cause to a pattern.
- Compare the same cohort age, not the same calendar date alone.
- Show both the percentage and the number of eligible customers.
- Treat incomplete observation windows as unavailable, not failures.
- Use customer research or an experiment to test possible explanations.
What can cohort analysis reveal that an average cannot?
An overall repeat rate can rise because the business acquired more customers from an already strong segment, even if the experience did not improve. It can also fall when a large new cohort has had little time to repeat. Cohort analysis separates those timing effects. It can indicate where to investigate onboarding, channel quality, pricing, or product changes, but it does not prove which change caused the difference. Pair the finding with customer segmentation when groups have materially different needs, and record the interpretation and next test in a decision log.
A fictional week-four retention comparison
A service groups first-time customers by the week of their first completed order. Of 200 customers in the 1 June cohort, 50 place another order during week four, so week-four retention is 25%. Of 160 customers in the 8 June cohort, 56 return in week four, or 35%. The ten-percentage-point difference is worth investigating, but it is not automatically evidence that a checkout change launched on 8 June caused it. The second cohort may have come from a different channel or faced different seasonal demand. The team checks cohort composition and measurement, interviews customers, and tests the most credible explanation before scaling the change.