Cohort analysis
A cohort is a group of customers who arrived in the same period. Cohort analysis tracks each group forward separately and compares them at the same age, rather than averaging everyone together at a single moment.
That sounds like a reporting preference and it isn’t. A blended average over customers of mixed maturity doesn’t measure behaviour, it measures behaviour tangled up with your acquisition rate, and the two move independently.
Why the average lies
Section titled “Why the average lies”Someone acquired last week hasn’t had time to reorder. Someone acquired last year has. Put them in the same denominator and the resulting figure describes nobody.
Worse, it moves for the wrong reasons. Acquire hard this month and a wave of zero-order customers enters the base, so your blended repeat rate falls. Nothing about customer behaviour changed. The number went down because the business went well, which is the most misleading shape a metric can have.
It runs the other way too. A brand whose acquisition has stalled will watch its blended retention numbers improve, because the base is ageing and nobody new is diluting it. Both effects punish exactly the interpretation people reach for.
What to read in it
Section titled “What to read in it”- The first drop. The gap between period zero and period one is the steepest fall in almost every business, and the most predictive. In DTC that’s the second order; in subscription it’s the first renewal.
- Where it flattens. Curves usually decay and then plateau. The plateau is your actual retained base, and it’s the number that belongs in an LTV model rather than the early figure.
- Whether newer cohorts sit above older ones. That’s the only clean read on whether anything you’ve changed is working, and it’s the same logic net revenue retention uses on revenue rather than customers.
The triangle shape is inherent, not a formatting problem. Recent cohorts have fewer columns because less time has passed, and the empty cells are the honest representation of what you don’t know yet.
Where it goes wrong
Section titled “Where it goes wrong”- Comparing incomplete cohorts to complete ones. April’s month-one number sits next to January’s month-four number and the eye wants to compare them. Only compare cells at the same age.
- Cohorts too small to mean anything. Slice monthly on a business acquiring forty customers a month and every curve is noise. Widen to quarters before the pattern is real.
- Ignoring what was different about the intake. A cohort acquired during a heavy discount period will retain differently from one acquired on full price, so a step change between cohorts is often an acquisition-mix story rather than a retention one.
- Changing the definition partway. Redefine “active” or move the cohort boundary and the whole grid stops being comparable, usually without anyone noticing for months.
It’s also the right frame for measuring anything with a delayed payoff. Holdout groups and novelty decay are both cohort questions wearing different clothes - in each case the thing you want to know is how a defined group behaves as time passes, which a snapshot cannot tell you.