What Is Cohort Analysis? How to Build and Read a Cohort Table

Cohort analysis groups customers by a shared starting point, such as their signup month or the quarter their deal closed, and follows each group separately over time. Instead of one churn rate for the whole business, you see how the January customers are doing in their third month next to how the March customers are doing in theirs.

That matters most when the overall number looks fine. A steady flow of new customers can keep total revenue climbing while each new group leaves faster than the one before, and cohort analysis catches that before the totals do.

Where the term comes from

A cohort was originally a unit of a Roman legion, and later any group of people who share something. Medical researchers use it for groups followed over time: a cohort study picks a group before an outcome appears and keeps checking on it, as the British Doctors Study did with smokers and non-smokers from 1951. Business cohort analysis borrows the idea, not the method. It runs on records you already have, and Wikipedia's entry says the two should not be treated as the same thing. A cohort table shows that a group behaved differently. It does not prove why.

How cohort analysis works

Every cohort analysis comes down to five choices. Write them next to the chart, because two people can build different tables from the same data and both call it "retention."

1. The starting event. What puts someone in a cohort: first purchase, account created, deal marked closed won. Google Analytics calls this the inclusion criterion.
2. The outcome you track. Still paying, still active, revenue, number of orders.
3. The period size. Days, weeks, months or quarters, matched to how often customers actually buy or use the product.
4. How a period is counted. Active in that exact period, in every period so far, or in any period so far (more on this below).
5. The table. Cohorts down the side, periods since the start across the top, the outcome in each cell. Newer cohorts have fewer filled cells, so the table forms a triangle.

A cohort table worked through

Suppose you sell a monthly subscription. You group customers by signup month and count how many are still paying in each month after they joined. The numbers are made up to show the method.

January's 88% in month 1 means 44 of the 50 January customers paid again. The table can be read three ways, and each answers a different question.

Across a row: how one group ages

January goes 88%, 80%, 76%. The losses shrink each month, which suggests a core of customers who are likely to stay. A row that keeps falling at the same rate, with no sign of leveling off, is the worrying shape.

Down a column: are newer groups better or worse?

Month 1 is 88% for January, 85% for February and 71% for March. March customers left far faster. Look at what changed for people who signed up in March: a new lead source, a discount that drew bargain hunters, a new onboarding process, a different rep team closing deals. The table doesn't say which. It tells you where to look.

Along a diagonal: did something happen on the calendar?

Cells on one diagonal fall in the same calendar month: January's month 2, February's month 1 and March's month 0 are all March. If every cohort dips on the same diagonal, the cause probably hit everyone at once, such as an outage or a price increase, rather than anything about who signed up.

Counting customers versus counting revenue

Say the January cohort started at $10,000 in monthly recurring revenue. By month 3, the 12 customers who left took $2,000 with them, but customers who stayed upgraded and added $3,000. The cohort now brings in $11,000: it kept 76% of its customers and 110% of its starting revenue.

The second figure is that cohort's net revenue retention, which Stripe's guide defines as starting recurring revenue, minus revenue lost to churn and downgrades, plus revenue from upgrades, divided by starting revenue, with new customers excluded. Above 100% means the group grows on its own.

Run both views side by side. Customer counts show whether the deals being closed are deals that last; revenue shows whether account management is growing the ones that do. Falling counts with rising revenue often means a few large accounts are covering many small losses.

Time cohorts and behavior cohorts

Most cohort tables group people by when they started. The other kind groups them by what they did: customers who used a particular feature in their first week, or accounts that had a kickoff call within 10 days of signing, compared against everyone else. Treat a gap as a lead to test, not proof. The customers who act early may simply be the more motivated ones.

Why the same data gives different percentages

Tools count a "retained" customer in different ways. Google Analytics offers three calculations:

• Standard: people who returned in that specific period.
• Rolling: people who returned in that period and in every period before it.
• Cumulative: a running total of people who returned in any period up to that point.

Amplitude instead offers N-day retention (came back on one specific day) and unbounded retention (came back on that day or any day after). Check that two charts were counted the same way before comparing them.

Common mistakes

• Reading the newest cohort as final. The latest cell in a row is often a period that hasn't ended, so it looks worse than it will. With calendar months, someone who signed up on the 28th has had only a few days of month 0.
• Cohorts too small to mean anything. In a cohort of eight customers, one cancellation moves the number by more than 12 points. Use quarters instead of months until each cohort is big enough.
• Explaining the cause from the table alone. The table shows which cohort changed. The reason comes from what was different for it: deal notes, lead sources, pricing, onboarding.
• Changing the definition midway. If "active" meant "logged in" last quarter and "used a paid feature" this quarter, the trend compares two different things.

When it doesn't help much

With a few dozen customers in total, reading the account list one by one tells you more than a table of percentages. Cohort analysis also says little about a business where customers buy once and don't come back. And it can't replace talking to customers: it shows that the March group left early, not what they were thinking.

Cohort analysis vs. segmentation

Segmentation splits customers by a trait at one moment, such as plan, industry or region. Cohort analysis groups them by a shared start and follows them forward. The two combine well: split each signup cohort by plan to see whether one plan's customers drop off faster.