ES

KPIs · Chapter 4 · Concept lesson · 7 min read

Retention is measured per cohort, at the same age

Retention asks what share of a group came back, and it only means something if every group has had the same time to do it. Churn without subscriptions adds a second decision: how long is silence?

The idea

Retention asks a simple question about a group: of the people who started in the same period, how many are still buying? The group is a cohort: customers who share a start date, such as the quarter of their first booked order. Each cohort is followed forward in time and compared at the same age, not on the same calendar date.

The calendar is what misleads. On 2025-12-31 the 2023Q1 cohort is almost three years old and the 2025Q4 cohort is a few weeks old. Put them in one sentence and you compare patience with youth.

Figure 1

Share of each cohort that ordered again within 365 days

0% 25% 50% 75% 100% Customers who bought again, % 2023Q1: 98.3%2023Q2: 92.4%2023Q3: 87.6%2023Q4: 88.4%2024Q1: 85.4%2024Q2: 90.5%2024Q3: 86.9%2024Q4: 82.4%2025Q1: 81.1%2025Q2: 67.6%2025Q3: 60.1%2025Q4: 43.9% 2023Q12023Q32024Q12024Q32025Q12025Q3

Step 1 of 4 A cohort is every customer whose first booked order fell in the same quarter. Each bar is the share of that cohort with a second booked order within 365 days of its first.

  1. A cohort is every customer whose first booked order fell in the same quarter. Each bar is the share of that cohort with a second booked order within 365 days of its first.
  2. The oldest cohorts sit between 98% and 85%. These bars are fair: every customer in them was watched for a full year.
  3. The last four bars fall to 81%, 68%, 60% and 44%. Nothing went wrong: the data ends on 2025-12-31, so the 2025Q4 cohort has had a few weeks, not 365 days.
  4. Those cohorts are not finished. Their bars will rise as time passes. Compare them only with other cohorts at the same age, such as retention within 90 days.

Why it exists

Getting a first order is expensive; the repeat order is where a store like Harbor & Pine makes its money. Leadership wants to know whether the business keeps its customers better or worse than it used to. A single store-wide repeat rate cannot answer that, because it blends customers who had three years to return with customers who had three days. Cohorts separate them.

Harbor & Pine also sells sofas and lamps, not subscriptions. Nobody cancels. A customer simply stops ordering, and the data never says that they have left. So you have to define churn before you can measure it, and the definition is a decision about time.

How it works under the hood

Build the cohort and the clock

For each customer take the first booked order (completed or refunded). Its quarter is the cohort. Then for each later booked order, count the days since that first order. A customer is retained at 365 days if they have another booked order no more than 365 days after the first. The retention rate for a cohort is retained customers divided by all customers in the cohort, including those who never came back.

Two details keep this honest. Customers who left stay in the denominator, otherwise retention would always read 100%. And only cohorts that have been observed for the full window get a full-window number. A customer whose first order is 40 days before the end of the data cannot yet have had a 365-day chance to return. This is called right-censoring: the end of the data cuts the observation short.

Reading a retention curve

Plot retention against age (30 days, 60 days, 90 days and so on) instead of against the calendar, and the cohorts line up from zero. A curve normally drops fast at first and then flattens. Where it flattens is the share of customers who become regulars. Curves from young cohorts end early: do not extend them with a guess, and do not average a young cohort's short curve into the old ones.

For one pooled number, compare cohorts only at an age they have all reached. At 90 days, for example, the 2023Q1 cohort reached 76.3% and the 2024Q4 cohort 58.6%. The same 365-day number is 98.3% against 82.4%. That is a real decline, and it is the sort of statement a cohort comparison makes possible.

Churn needs an as-of date and a window

With no cancellation event, churn is a rule: a customer has churned if they have no booked order in the last N days before the as-of date. Two inputs, one rule. The as-of date is 2025-12-31. For a 180-day window, anyone whose last order is before 2025-07-04 has churned. Only customers who first bought before that cut-off are eligible, since a customer who started in November has not had 180 days to lapse.

Figure 2

Deciding who has churned when nobody cancels

  1. Last booked order per customer
  2. As-of date 2025-12-31
  3. Days of silence as-of minus last order
  4. Churn window 180 days
  5. Churned last order before 2025-07-04

Step 1 of 4 Harbor & Pine has no cancel button. Start from each customer's most recent booked order.

  1. Harbor & Pine has no cancel button. Start from each customer's most recent booked order.
  2. Fix a date to measure from. The data ends on 2025-12-31, and the report says 'as of' that date.
  3. Count the days between the last order and the as-of date. That is the silence.
  4. Choose a window. With 180 days, a customer whose last order is before 2025-07-04 has churned. The choice of window is the whole definition.

The window is the metric

Pick N from the real gaps between orders. Harbor & Pine has 9,413 of them: the median is 23 days, the 90th percentile 182. A window near the 90th percentile means that most customers who are still buying have not yet crossed it, and silence beyond it is unusual enough to act on.

Figure 3

Same customers, same orders, three churn windows

0% 15% 30% 45% 60% Eligible customers counted as churned, % 90 days: 58.1%180 days: 43.9%365 days: 24.9% 90 days180 days365 days 180 days: 43.9%

Step 1 of 4 Typical gaps are short: the median is 23 days between one order and the next, and the 75th percentile 76 days.

  1. Typical gaps are short: the median is 23 days between one order and the next, and the 75th percentile 76 days.
  2. A 90-day window flags 58.1% of eligible customers. Many of them were only between two ordinary orders.
  3. A 365-day window flags 24.9% and treats customers as active for a whole year after their last purchase. Too late to act on.
  4. The 90th percentile of the gaps is 182 days, so 180 days is a defensible window: 43.9% churned. Write it down and keep it.

Churn of 44% in one report and 25% in the next is not an improvement if the window quietly went from 180 to 365 days. Keep the window in the label, and change it only on purpose.

Segments need their own window

Gaps are not spread evenly. A business account with 100 orders contributes 99 gaps, and a consumer with two orders contributes one, so the 23-day median is mostly the rhythm of heavy buyers. Take one median gap per customer and the middle customer's typical gap is 55 days. If segments behave this differently, give them different windows, or at least report churn per segment.

Reactivation

A churned customer is sometimes not gone. A reactivation order is a booked order placed more than 180 days after the same customer's previous booked order. Count them before paying for a win-back campaign: some customers cross the line and return without any prompting.

Common misconceptions

"Retention is the share of customers who are still active"

"Active" depends on today's date, which makes the number drift as the data grows. Retention is counted against each customer's own first order, within a stated window, so the same cohort gives the same answer next month as long as it is fully observed.

"A low number for the latest cohort means customers are getting worse"

It usually means the cohort is young. Check how many days of history the cohort has before you compare it with older ones, and compare at equal ages.

"A customer who stopped ordering has churned"

Only after your window says so. Stopped is a fact; churned is a definition. A customer 60 days past their last order is churned at a 30-day window and perfectly normal at a 365-day window.

Check yourself

One question. Answering it ticks this lesson off on the map; it never locks a level.

The 2025Q4 cohort shows 43.9% repeat within 365 days; 2024Q4 shows 82.4%. A manager concludes that new customers have become much less loyal. What is the best reply?

Cookie settings

Pick what may run on this site. You can change this any time from "Cookie settings" at the bottom of every page.