Cohort Analysis
Cohort analysis groups users by a shared characteristic or start date and tracks how each group behaves over time, revealing patterns that aggregate metrics hide.
Key takeaways
- Time is re-indexed to periods since the defining event, not to calendar dates.
- Reading down a column compares cohorts observed at exactly the same age.
- Cohort width trades resolution against the noise of thin, small rows.
- Grouping by acquisition source separates channel quality from genuine product improvement.
- Recent cohorts are incomplete; their later columns rest on very few observations.
In depth
A cohort analysis fixes two things: the event that defines membership and the clock that starts at that event. Everyone who signed up in the same week forms one cohort, and their behaviour is then indexed by weeks since signup rather than by calendar date. That re-indexing is the whole trick, because it lines up people who joined at different times at the same point in their own lifecycle. The result is the familiar triangular table, each row a cohort, each column a period after acquisition.
Two choices decide what the table can tell you. Cohort width sets resolution: weekly cohorts catch the effect of a specific campaign but produce thin, noisy rows, while monthly cohorts are stable but blur anything that happened mid-month. The grouping dimension decides the question: cohorting by signup date answers whether the product is improving, while cohorting by acquisition source answers which channel brings people who stay. Cutting by both at once is where the insight usually is, and also where sample sizes collapse fastest.
Read the table in two directions. Down a column compares cohorts at the same age and shows whether recent changes helped; across a row shows how one cohort decays and where in the lifecycle people leave. In a quiz funnel, the score itself makes a useful cohort dimension: group leads by the tier the scorecard assigned them and follow each tier's conversion to customer. If the top tier does not outlast the bottom one, the scoring rules, not the traffic, are what need fixing.
The newest cohorts are always the least informative, because they have only lived through a few periods and the last column of the table is built from a handful of people. Cohorts are also not experiments: two cohorts differ in every way the world differed between them, including seasonality, pricing and competitor moves, so a difference between them suggests a hypothesis rather than proving a cause. And a cohort defined by an event most users never reach describes a self-selected group, not the base.
Example in practice
How to measure it
The core number in each cell is the share of the cohort still active in that period: active users in period N divided by the cohort's original size. Watch where the curve flattens, because that plateau is the share of users who stick, and it matters more than the steep early drop. Compare plateaus between cohorts rather than first-week numbers, which move with onboarding cosmetics.
Beyond retention, run the same table on revenue and on qualified conversion. Cumulative revenue per original cohort member, tracked by month since acquisition, shows whether payback is arriving faster in newer cohorts. For lead funnels, the equivalent is the share of each cohort that reaches a sales conversation, which exposes channels that deliver volume early and nothing afterwards.
Common mistakes
The most damaging error is comparing cohorts of unequal age in the same glance. A cohort three months old will always look worse on twelve-month retention than one from last year, simply because that column is empty. Truncate every comparison at the age the youngest cohort has actually reached, and mark incomplete cells as blank rather than zero, so nobody reads a missing period as a collapse in retention.
The second is attributing a cohort difference to the only change the team remembers making. A product release, a pricing test, a seasonal traffic shift and a new ad channel can all land in the same month, and the cohort table cannot separate them. Keep a dated log of releases and campaign changes next to the table, and where a difference matters enough to act on, confirm it with a controlled test rather than the cohort alone.
Frequently asked questions
What is the difference between cohort analysis and segmentation?
Segmentation groups users by attributes at a single point in time, while cohort analysis adds the dimension of time by tracking each group's behavior across later periods. Cohort analysis is essentially time-based segmentation focused on retention and lifecycle trends.
Which metrics work best in a cohort analysis?
Retention rate, lifetime value, and activation rate are the most common because they reveal whether engagement holds or decays over time. For lead funnels, lead-to-customer conversion by cohort is especially powerful.
How often should I run cohort analysis?
Most teams review cohorts monthly to catch trends early without reacting to short-term noise. After major product or funnel changes, run a focused cohort comparison to measure the real impact on retention.
What is the difference between a cohort and a segment?
A segment groups people by an attribute they have now, such as plan type or country, and its membership can change. A cohort groups people by something that happened at a fixed moment, usually acquisition, and membership never changes afterwards. That stability is what allows a cohort to be tracked over time; a segment measured over time silently swaps its members.
How large does a cohort need to be?
Large enough that a few individuals cannot swing the percentage. As a rule of thumb, cohorts below a few dozen members produce retention curves that jump around for reasons that have nothing to do with the product. If weekly cohorts are too small, widen to monthly or group several weeks of a single campaign together rather than reading noise as a trend.
Which event should define the cohort?
Pick the event closest to the decision you want to make. To judge acquisition channels, cohort on first touch or signup date. To judge onboarding, cohort on activation, the first moment of real value. To judge a pricing change, cohort on first payment. Using one cohort definition for every question is what makes a cohort report feel uninformative.
Can cohort analysis work for B2B lead funnels?
Yes, with a longer clock. Group leads by the month they entered the funnel and track how many reach a meeting, an opportunity and a closed deal in each subsequent month. Because B2B cycles run for weeks or months, most of the interesting movement sits in columns three to nine, so a table that stops at ninety days will show almost nothing.
Why do my newest cohorts always look best?
Usually because they are incomplete. Early periods have high retention by definition, so a young cohort sitting only in its first two columns compares favourably against older cohorts read at full length. Check that you are comparing the same column, not the same row length, and be alert to survivorship in any average that mixes cohorts of different ages.
How is cohort analysis different from a funnel report?
A funnel report shows how many people pass each stage; a cohort table shows when they do it and whether the pattern is changing between groups. A funnel can look healthy while every recent cohort is converting slower than the last, because the funnel blends them together. Use the funnel to find the leaky stage and cohorts to see whether it is getting worse.