Cohort analysis: seeing retention your averages hide

Your overall retention rate has held near 82% for six straight months, so you tell the board the product is sticky and move on. Underneath that steady number, your newest customers may be leaving in waves while a shrinking core of loyal early adopters holds the average up. A blended number can stay flat while the business quietly decays. Cohort analysis is how you see what the average hides.

The fix is not a fancier dashboard. It is a change in how you group the customers before you count them.

What a cohort actually is

A cohort is a group of customers bucketed by when they joined. Everyone who signed up in January is your January cohort; everyone who signed up in February is the February cohort, and so on. Once customers are grouped this way, you stop asking "what is our retention this month" and start asking a sharper question: of the customers who joined in a given month, how many are still active a set number of months later?

Retention here just means the share of a cohort still paying you. Months-since-signup — not the calendar — becomes your clock. That shift is the whole point. Calendar-based averages mix customers who are one month old with customers who are two years old. Cohort math lines everyone up at the same starting line so you can watch how each group behaves as it ages.

Why a single average lies

A blended retention number combines populations at completely different stages of life. Your two-year-old customers have already survived every reason they might have left; the ones still around are, almost by definition, the sticky ones. Your brand-new customers include everyone who is about to bail in the first 90 days. Average those two groups together and the healthy tail of long-term survivors props up the number, masking heavy losses among the very people you just paid to acquire.

This is how a company posts a flat 82% month after month while new-customer retention is falling off a cliff. The average is not wrong, exactly — it is answering a question you did not mean to ask. You wanted to know whether customers stick. It told you the weighted mixture of everyone currently on the books, dominated by whoever has been around longest.

Building the grid

The tool that fixes this is a simple grid. Each row is a signup month. Each column is months since signup: 0, 1, 2, 3, and onward. Every cell holds one number — the percentage of that cohort still active that many months after they joined. You fill it in as time passes, so newer cohorts have fewer columns because they have not aged yet.

The grid earns its keep because you read it two ways:

• Read down a column to compare cohorts at the same age — is month-1 retention getting better or worse as you sign up newer groups?
• Read across a row to watch a single cohort age — this traces its retention curve month by month.
• Scan a whole column for a consistent drop — a decline that repeats in every cohort at the same age is a product or onboarding problem, not noise.
• Compare the earliest columns across recent rows to judge whether a change you shipped actually moved early retention.
• Watch where each row flattens — the level it settles at is your long-term retained base.

A worked example

Say you add about 100 customers a month at $50 each. Follow the January cohort across its row. Month 0: 100 customers, 100%. Month 1: 92 still paying, 92%. Month 2: 88, or 88%. Month 3: 71 — a drop to 71%. Month 4: 68%. Month 5: 66%. Month 6: 65%, where it roughly settles.

One row is a story; the grid is the pattern. When you look down the month-3 column, the February and March cohorts show the same lurch — 89% to 72%, then 87% to 70%. The loss is not random churn scattered everywhere. It clusters at month 3, in every cohort. That is a cliff, and cliffs have causes: a trial that converts to full price around then, an annual charge landing, or the point where customers have solved their first problem and no longer feel ongoing value.

Now put dollars on it. Each retained customer contributes about $40 a month after direct costs, and you spent roughly $180 to acquire each one — a little over four months of contribution to break even. With the month-3 cliff, a large share of every cohort quits just before it pays you back. Lift month-3 retention from 71% to 85% and you keep 14 more customers out of every 100 past the payback line. Across 100 signups a month, that is 14 customers × $40, or $560 in monthly contribution added and compounding, from a single fix.

Turning the grid into action

A cohort grid does more than diagnose. The retention curve it draws feeds the numbers you actually run the business on. Multiply each cohort's monthly retention by contribution per customer and sum across the curve, and you get a grounded lifetime value instead of a hopeful guess. Compare that to acquisition cost and you get a payback period you can trust. And the shape of the cliff tells you exactly where to intervene — a drop at month 3 sends you to study what happens around day 60 to 90, not to a vague "improve retention" project.

That specificity is the payoff. A blended average tells you retention is fine or not fine. A cohort grid tells you which customers, at what age, are leaving — and lets you test whether the onboarding change you shipped in April actually held the May cohort past the point where March fell away.