Minimum cohort size is the smallest group that produces a rate you can act on. Cohort analysis groups customers or deals by a shared start point and tracks them forward, and the arithmetic behind every cohort rate is a small numerator over a small denominator. When the denominator gets small enough, the rate stops describing the group and starts describing whichever account moved last.
The threshold is not a statistical rule imported from somewhere else. It is a question about how much movement one record can cause.
The one-account test
Take the cohort, remove a single account, and recompute. If the rate moves by more than the difference you are trying to detect between cohorts, the cohort is too small.
| Cohort size | Impact of one churned logo |
|---|---|
| 10 accounts | 10 points |
| 25 accounts | 4 points |
| 50 accounts | 2 points |
| 100 accounts | 1 point |
Why revenue cohorts need more
Revenue-weighted cohorts carry two sources of variance. Count variance behaves as above. Concentration variance sits on top of it, because a cohort of fifty accounts where one contributes 30 percent of ARR behaves like a cohort of a handful.
Before trusting a net revenue retention figure by cohort, check what share of cohort ARR the largest account holds. When one account holds enough of cohort ARR that its departure alone would move the rate past your detection threshold, report the cohort with and without that account, because the two numbers will tell different stories and the honest answer is both.
Fixes that work
Widen the window first. Monthly cohorts become quarterly cohorts, and the denominator roughly triples. The cost is time resolution, which matters if you are trying to isolate the effect of a specific product change to a specific month.
Group by attribute second. If quarterly cohorts are still thin, cohort by segment, acquisition channel, or entry product instead of by start date. This answers a different question. It tells you which kind of customer retains, not how retention changed over time.
Push the observation window third. A twelve-month retention curve on a nine-month-old cohort is a projection, and projecting from a thin cohort compounds both errors.
Small cohorts still have diagnostic use. ORM's earliest churn signal is support case volume, where zero cases in a year and seven or more cases in a year both indicate risk, and three to five tier 2 or tier 3 tickets indicate an engaged customer who is less likely to churn. That signal reads at the account level, so it works when the cohort rate does not. When the group is too small to produce a trustworthy rate, inspect the accounts directly and treat the pattern as a hypothesis to test once the cohort fills out. The same logic applies to win rate by cohort, where thin deal counts produce swings that look like performance changes and are not.
Frequently Asked Questions
How many customers do you need for a cohort analysis?
Thirty accounts is a reasonable floor if you are trying to detect a three to five point difference between cohorts. Below thirty, a single churn moves the rate by more than three points on its own, which swamps the difference you are looking for. For revenue retention the floor is higher because dollar concentration adds variance on top of count variance.
What do you do when cohorts are too small?
Widen the cohort window from monthly to quarterly, which is the cheapest fix and costs you time resolution. If quarterly cohorts are still thin, group by a shared attribute such as segment or acquisition channel instead of by start date, and accept that you are no longer measuring a time effect.
Does cohort size matter more for revenue retention than logo retention?
Yes. Logo retention weights every account equally, so variance scales with count. Revenue retention weights by dollars, so one large account leaving swings the rate by whatever share of cohort ARR it held, which can be tens of points in a cohort that looks adequately sized by headcount. Check the largest account's share of cohort ARR before trusting the number.
Is a small cohort useless?
No, but it is a hypothesis rather than a finding. A small cohort showing a sharp retention drop is worth investigating account by account. It is not worth putting on a board slide as a trend, because the confidence interval around it covers both an improvement and a decline.
Put these metrics to work
ORM builds custom revenue forecast models that turn concepts like minimum cohort size for analysis into prescriptive action for your team.
Schedule a Demo