Subscriptions
Cohort retention
Cohort retention measures what share of customers acquired in a given period are still active after each subsequent period, tracked separately per acquisition cohort.
It is the correct primary view because it separates effects a blended rate hides. An improving product with worsening acquisition quality can produce a flat overall churn number while both trends are large.
Cohorts should be cut by acquisition channel as well as date. Customers acquired through a heavy first-order discount retain differently from organically acquired ones, and blending them makes both unreadable.
How to read a retention curve
A retention curve plots the share of one acquisition cohort still active against periods since acquisition, not against calendar months. That alignment is the entire point: every cohort starts at 100% at its own month zero, so cohorts acquired a year apart can be compared at the same age.
Three things are worth reading off it, in this order. The first-period drop, which reflects how well the offer matched the product. The slope of the tail, which reflects whether the product keeps earning its place. And whether the curve flattens at all, which reflects whether a durable base exists.
Two curves can end at the same point and mean opposite things, which is why the shape is read and not just the endpoint. In the illustration below, Cohort A lost half its customers immediately and then almost nothing — an offer or expectation problem sitting in front of a product that works for the people it suits. Cohort B kept nearly everyone at first and has been losing them steadily ever since, and at month 12 it is still falling. B was worth more across the year; A is the one whose second year can be forecast.
| Months since acquisition | Cohort A | Cohort B |
|---|---|---|
| 1 | 60% | 88% |
| 3 | 55% | 76% |
| 6 | 52% | 63% |
| 9 | 51% | 55% |
| 12 | 50% | 50% |
Logo retention and revenue retention are different curves
The same cohort produces different curves depending on what is counted. Counting customers gives logo retention; counting the money they spend gives revenue retention. The two diverge whenever the customers who leave are not of average value.
In DTC subscriptions the churners usually skew below average — discount-acquired, single-line, longest cadence — which makes the revenue curve flatter than the logo curve, and a business reading only the logo curve understates itself. The reverse pattern, revenue falling faster than customers, means the valuable customers are the ones leaving, and it is the more urgent of the two signals by a wide margin.
Three measures are in common use and they are not interchangeable.
- Logo retention: the share of the cohort still active. Cannot exceed 100%. Answers whether people stay.
- Gross revenue retention: recurring revenue from the cohort as a share of what it started with, counting cancellations and downgrades only. Cannot exceed 100%. Answers how much of the original book survives.
- Net revenue retention: the same measure including expansion — added subscriptions, upgraded cadence, larger boxes, price increases. Can exceed 100%, so a cohort that is losing customers can still be growing in revenue.
Where the curve flattens, and what that tells you
The interesting feature of a retention curve is not its height but whether it reaches an asymptote. A curve that flattens has found the share of buyers for whom the product has become habitual, and everything past that point is a stable base that can be planned around.
The flattening point is also where retention spend stops paying for itself. Interventions aimed at the steep early section — onboarding, the first delivery, the cadence choice, the first renewal — act on the part of the curve where customers are actually leaving. The same interventions applied to a flat tail move people who were not going anywhere.
A curve that never flattens inside the observation window is the important negative result. It means there is no durable base yet, that lifetime value cannot be modelled beyond what has been observed without inventing the tail, and that the honest reporting is a measured figure at a stated horizon rather than a projection.
The mistake: comparing cohorts of different ages
The most common error in cohort reporting is comparing retention-to-date across cohorts of different ages. The January cohort has been observed for eleven months and the November cohort for one, so January's retention is lower — because it is older, not because it is worse. Comparisons are only valid at equal months since acquisition.
The second version is the partially observed final period. A cohort's most recent month is incomplete, because not every subscriber in it has reached a billing anniversary yet, so the newest point on every curve is unstable and should be marked as provisional or dropped.
The third is silent recomputation. If cohorts are cut by acquisition channel or discount depth and those definitions later change, old cohorts get re-labelled and the curves move without a single customer behaving differently. Freeze cohort membership, and the labels on it, at acquisition.
Then there is the sample-size floor. A cohort cut by month, then channel, then entry product, then discount depth is four divisions deep and is usually noise wearing a chart. Cut on the dimension you can act on, and keep the cohorts large enough that one month's wobble is not a finding.
Frequently asked questions
- What is cohort retention?
- Cohort retention measures what share of the customers acquired in one period are still active in each subsequent period, tracked per acquisition cohort rather than blended. Because every cohort is aligned to its own month zero, cohorts acquired at different times can be compared at the same age, which a single blended churn figure cannot do.
- How do you read a retention curve?
- Read three things: the size of the first-period drop, which reflects how well the offer matched the product; the slope of the tail, which reflects whether the product keeps earning its place; and whether the curve flattens, which tells you whether a stable base exists. Two curves can end at the same point and mean opposite things.
- What is a good retention curve?
- One that flattens. The absolute level varies too much by category and cadence to compare across businesses, but a curve that reaches an asymptote has identified a group of customers for whom the product has become habitual, and that group can be planned around. A curve still falling at the edge of the data has no durable base yet.
- What is the difference between logo retention and revenue retention?
- Logo retention counts customers still active; revenue retention counts the money they spend. Gross revenue retention counts cancellations and downgrades only and caps at 100%. Net revenue retention also counts expansion — upgrades, added lines, price rises — and can exceed 100%, so a cohort can lose customers and still grow in revenue.
- Why does the newest cohort look different from the others?
- Usually because it is incomplete rather than different. Its most recent period is only partly observed, since not every subscriber has reached a billing anniversary within it, and it has been watched for fewer periods than the cohorts above it. Compare cohorts only at equal months since acquisition, and mark the incomplete point.
- Should cohorts be cut by acquisition channel?
- Yes, as long as the cohorts stay large enough to read. Customers acquired through a deep first-order discount retain differently from organically acquired ones, and blending them makes both curves unreadable. Freeze the channel label at acquisition so that recategorising channels later does not silently rewrite old curves.