Definition
A cohort has two parts: a starting event and a period. The event defines membership, and it has to be chosen deliberately. First paid order suits retail, subscription start suits subscriptions, and the first meaningful action suits products with a free tier, because cohorts of mere sign-ups mix people who never became customers with people who did. The period, usually a calendar month, groups everyone whose event fell inside it, so "the June cohort" means a fixed list of people that never changes.
Cohort analysis then measures each cohort at successive ages: the share still active, the revenue earned, the orders placed, a month or a quarter after the start. The point is comparison at equal age. A customer three months in and a customer three years in are at different stages, and only a table that lines them up by age compares like with like. The cohort analysis tutorial builds such a table from raw order data, in SQL or a spreadsheet.
In practice
A cohort table has one row per cohort and one column per age. Reading it in three directions answers three questions:
- Along a row: how one cohort behaves as it ages, which shows where customers drop away and whether the curve flattens into a loyal core.
- Down a column: how different cohorts compare at the same age, which is where an improvement or a deterioration in newer customers first appears.
- Along a diagonal: what happened in a single calendar period across cohorts of every age, which picks out events such as a price rise or a service failure that touched everyone at once.
The retention and payback guide explains how to read the shapes these curves make, and the KPI tree for retention shows why the drivers of retention have to be read by cohort age rather than as one blended figure. For practice, the sample ecommerce dataset includes customers, orders and items tables with planted patterns designed for cohort retention exercises.
Cohorts by channel, offer and month
Splitting cohorts by what was true at acquisition shows which sources and promotions bring customers who stay. A channel that recruits cheaply and loses customers quickly can cost more per retained customer than a dearer one; a discount that lifts sign-ups can recruit people who leave after the first order. The attributes used for the split should be the ones recorded at the start, not the customer's current record, or customers move between groups and the comparison breaks.
Cohorts are not only for consumer retention. In business-to-business marketing, where revenue arrives months after the first touch, the B2B measurement guide follows each quarter's cohort of first-touched leads through its pipeline stages, reading early cohorts on revenue and recent ones on their early stages. For online shops, the ecommerce retention section covers the flows and loyalty mechanics whose effect cohort tables measure.
Cohorts versus segments
A segment groups customers by a shared attribute, such as country, plan or order value, and membership can change as the attribute changes. A cohort groups them by when they started, and membership is fixed. The two combine: cohorts split by segment show whether a segment's behaviour is improving over time, which neither view shows alone.
A steady average hiding a weak promotion
Fennelwicket Meal Kits reports monthly retention of about 82% for a year and treats it as stable. A cohort table by month of first box tells another story. Customers who joined between January and April at full price are still ordering at 64% after three months. The May and June cohorts, recruited with a 60% first-box discount, number 2,900 and 3,400 and fall to 38% and 35% active after their second month, mostly leaving straight after the discounted box. The blended rate held up because 11,000 older subscribers dominate the base. Fennelwicket replaces the deep discount with a smaller one spread across the first three boxes, and the August cohort is still 58% active after three months.
Related terms
Churn RateShare of customers at the start of a period who left during it. read by cohort shows whether newer customers leave faster than older ones, which the blended rate hides. Repeat Purchase RateShare of customers who placed more than one order in the period. is the transactional equivalent, and is most honest when measured at a fixed age per cohort. Cohort revenue curves are the basis of LTVTotal value a customer brings over time. estimates, and cumulative margin per cohort against what that cohort cost to acquire gives the Payback PeriodMonths of customer margin needed to recover the cost of acquiring that customer.. The LTV and CAC guide shows how those pieces fit together.
Common confusions
- A cohort is not a segment. Membership is fixed by the starting event, not by a changing attribute.
- Sign-up cohorts are not customer cohorts. Choose the event that marks a real customer.
- Young cohorts are not failing cohorts. Compare cells at the same age, never a new row's total with an old row's total.
- Current attributes distort the split. Use what was true at acquisition.
- A stable blended rate is not proof of stable behaviour. Mix change can hide it.
