HeyMark
Content

Blog

Cohort analysis: how to read retention by group

7 min read

A cohort analysis groups customers or users by something they share, almost always the period in which they started, and follows each group over time to see how many are still active. Google Analytics uses the same idea: a cohort is a group of users who share a common characteristic, such as the same acquisition date.

Customer retention rate answers how many customers in one group remain at the end of a period. A cohort analysis puts several groups side by side. That shows whether the customers who joined this quarter stay more or less than earlier ones, something a single total blends together.

How to read a cohort table

Fictional case. A monthly subscription counts its customers by sign-up month, with data as of the end of April. “Active” means they paid that month. Month 1 is the first full month after sign-up.

Cohort Customers at start Month 1 Month 2 Month 3
January 200 60.0% (120) 48.0% (96) 44.0% (88)
February 400 55.0% (220) 42.0% (168) —
March 100 70.0% (70) — —
April 300 — — —

The dashes aren’t zeros. They are months that cohort hasn’t reached yet: the April cohort has just formed, and its month 1 will be May.

The table reads in three directions:

  • Along a row, one cohort ages. January goes from 60% to 48% to 44%: it loses 12 points between months 1 and 2, and only 4 between months 2 and 3. The drop slows down, which tells you a group of customers is sticking around.
  • Down a column, you compare cohorts at the same age. In month 1, March retains 70% against February’s 55%. Before crediting something that changed in March, look at the size: it’s 100 customers, and a small cohort varies more. Wait for the next one to confirm.
  • Along a diagonal, you see the same calendar month. January in month 3, February in month 2 and March in month 1 all fall in April. If all three cells dropped at once, the first hypothesis would be something that happened in April, such as a price change or a billing failure. Before blaming it, compare, for example, April’s failed payments with those of earlier months.

How to summarize a column without fooling yourself

Two common mistakes change the result.

Dividing by cohorts that haven’t reached that month yet. In month 2 there are 264 active customers (96 + 168). If you divide them by the 1,000 customers in all four cohorts, you get 26.4%. But March and April haven’t reached month 2. With the two cohorts that have, retention is 264 ÷ 600 = 44.0%.

Averaging percentages from cohorts of different sizes. The simple average for month 1 is (60 + 55 + 70) ÷ 3 = 61.7%. Weighted retention is 410 ÷ 700 = 58.6%. The simple average gives March’s 100 customers the same weight as February’s 400. We explain it in more detail in weighted average.

To know what share of your customers is still active each month, use the weighted figure. The simple average gives every cohort the same weight, so it can land above or below it depending on which cohorts retain more; here it lands above because March, the smallest, retains more. To see whether cohorts resemble each other, look at their rates row by row.

Cohort retention calculator

Enter how many customers started in each cohort and how many were still active each month. Leave blank any month a cohort hasn’t reached yet.

Customers by cohort and month
CohortCustomers at startMonth 1Month 2Month 3
January
February
March
April

Use whole numbers without thousands separators. Each row is a cohort, from oldest to newest; the names are the example’s.

Initial example result

CohortMonth 1Month 2Month 3
January60.0%48.0%44.0%
February55.0%42.0%—
March70.0%——
April———
Weighted retention58.6%44.0%44.0%
Simple average of percentages61.7%45.0%44.0%
Cohorts with data321

Month 1: weighted retention of 58.6% across 3 cohorts; the simple average gives 61.7%. If they differ, the cohorts have different sizes and the simple average gives each one the same weight.

Retention in a cell = active that month ÷ customers at start × 100. Weighted retention for a month = sum of active ÷ sum of customers at start, using only cohorts with data for that month.

Fictional subscription example. A blank cell isn’t a zero: it means the cohort hasn’t reached that month yet and leaves it out of the summary. Decide first what counts as active and apply the same rule to every cohort.

Define what counts as active before comparing

The same table changes with the rule. In a subscription, active can mean paid that month. In an app, opened it. In a store, bought again. Pick the rule based on the decision you want to make, and apply it the same way to every cohort.

In Google Analytics 4, the cohort exploration asks for that definition in three parts: the condition to join a cohort (for example, first visit), the return condition (for example, a purchase) and the calculation type. Standard counts those who met the condition in that period, regardless of the others. Rolling requires them to have met it in that period and in every previous one. Cumulative counts those who met it in any period up to that one. GA4 also builds cohorts from device data only and doesn’t consider User-ID, so one person using two devices can count twice. If you change the calculation type halfway through the analysis, the figures stop being comparable.

What to do with what you find

If one cohort retains worse, look for what was different about it: the channel it came from, the price it paid or the onboarding it got. Splitting cohorts by channel answers, for example, whether customers who came through a referral program stay longer than those from ads.

Cohort retention is also the basis for estimating customer lifetime value from observed data. A curve that flattens leaves a base of customers who keep paying; one that keeps falling calls for a look at the product or the offer before spending more on acquiring customers.

Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.