Reading retention curves to catch churn before it shows up in revenue

A cohort can start decaying in March and leave the revenue line untouched until December. That gap, between the month customer behaviour changes and the month finance notices, is why retention curves are worth reading at all. New acquisition keeps refilling the base while older cohorts thin out underneath it, and the two movements cancel in aggregate revenue for two or three quarters. When the line finally bends, you are diagnosing the behaviour of customers you paid for six months ago, and the cheap corrections (a rewritten onboarding sequence, a repriced second order) have expired.

This lesson reads the curve as an instrument: the shape of the drop, where it flattens, how long reactivation stays viable, and how much lag sits between all of that and reported revenue.

What a retention curve actually shows

A retention curve plots the share of a cohort (customers whose first purchase fell in the same period) still active in each subsequent period.

Basic formula:

Retention rate (month N) = (customers from the cohort still active in month N) / (total customers in the original cohort) x 100

"Active" needs a strict definition: usually one purchase inside a rolling window. Thirty days for grocery, ninety for beauty or apparel, one hundred and eighty for furniture, because nobody buys a sofa monthly. Pick the window from your category's repurchase gap, then leave it alone. Widening it by thirty days manufactures a retention improvement that no customer actually performed.

The curve drops steeply, then flattens into a shelf. The shelf is the part that pays, because customers sitting on it behave like the persistent profile the lifetime valuelifetime valueLifetime Value: the total revenue (or profit) a customer generates throughout their entire relationship with your business.View full definition → lesson calculates against. Two cohorts can both pass through 30% at month three and end up in different businesses: one flattens at 22% and holds, the other keeps sliding to 8% by month twelve. A quarterly report shows them as identical. So read the slope between month 6 and month 12, not the month-3 number alone. Near-zero slope means you have a base. A slope that never quite reaches zero means you rent customers rather than keep them.

Amazon is the extreme case of a flat shelf. Prime converts repeat purchase into a single annual renewal decision, and third-party estimates have long put renewal above 90%. The trade-off is timing: when retention collapses into one event a year, the warning has to come from usage frequency between renewals, because the curve itself only speaks once.

Why retention curves warn you before revenue does

Revenue is the sum of two moving parts, new acquisition and repeat behaviour, and one strong acquisition quarter conceals a lot of decay. Cohort curves isolate the repeat behaviour of a single group, stripped of new-customer noise. That isolation is what makes them predictive rather than descriptive.

A worked example: how long the lag actually runs

Illustrative arithmetic for a steady-state retailer:

• 10,000 new customers a month, average order value $80
• curve at 40% (month 1), 30% (month 2), 27% (month 3), easing to a 20% shelf
• across the first year that is roughly 28,000 repeat orders a month from earlier cohorts, plus 10,000 first orders: about 38,000 orders, near $3m

Now suppose every new cohort starts losing 7 points from month 2 onward.

• One month after the shift, a single degraded cohort is in the mix: about 700 missing orders, 1.8% off the top line, invisible against the promo calendar.
• Six months in, six degraded cohorts: roughly 4,200 missing orders, about 11%. Finance sees it now.
• Those six cohorts, some 60,000 customers, are already bought and already decaying. Remediation starts half a year late.

The lag stretches further when acquisition is growing, which is precisely the condition under which retailers conclude nothing is wrong.

Benchmarks: what "normal" retention looks like in retail

Retention numbers vary so widely by category that the absolute figure is close to useless as a target, and the comparison discipline (which denominators are honest, where sector medians mislead) belongs to the benchmarking lesson. Three things are still worth carrying:

• grocery and consumables post far higher 90-day repeat rates than specialty apparel, because necessity drives frequency rather than affinity
• mid-market e-commerce 90-day repeat rates are commonly cited in the 20-30% range, with wide variance by source (see Baymard Institute for e-commerce behaviour research)
• in the EU, guest checkout and refused tracking consent under GDPRGDPREU regulation governing how organizations collect, store and use personal data, with fines tied to global revenue for breaches.View full definition → fragment customer identity, so cohorts are undercounted at the edges and curves read slightly worse than the truth

A retailer holding 25% month-3 retention for two years is healthier than one at 35% shedding 3 points a quarter.

How to actually build one (without a data team)

A cohort table needs three fields from transaction data: customer ID, order date, order value.

Cohort Month | Month 0 | Month 1 | Month 2 | Month 3
Jan 2026     | 100%    | 42%     | 31%     | 27%
Feb 2026     | 100%    | 40%     | 29%     | 24%
Mar 2026     | 100%    | 38%     | 26%     | 19%

Scan down each column, not across each row: 27% to 24% to 19% at month three is the signal. Four things break this table in practice.

• The newest cohorts have the fewest elapsed months, and an in-progress month always reads as a collapse. Drop the current period.
• December cohorts are gift buyers with structurally weaker repeat behaviour. Compare December to December, never December to March.
• Heavily discounted cohorts (a 50%-off first box, a Black Friday intake) deserve their own curve. Blended in, they hide which side is decaying.
• Reactivated customers usually re-enter the curve as "active", which lifts the tail and disguises a churn-then-buyback pattern. Track continuously active and ever-returned as separate series.

Subscription retail adds one more trap. HelloFresh lets subscribers skip weeks, so a curve built on "has an active subscription" flatters badly; build it on boxes actually shipped. Birchbox ran the mirror image: the box went out monthly regardless of interest, so purchase-based retention looked stable while engagement drained away, and the business changed hands in a distressed sale in 2021.

Google Analytics 4 and most commerce platforms ship a cohort retention report. For the underlying maths, this cohort analysis guide comes from a product analytics vendor, so read it as vendor material, but the logic transfers to retail transaction data unchanged.

🎬 [VIDEO: "Cohort AnalysisCohort AnalysisCohort analysis groups users by a shared starting trait or time (such as signup month) and tracks their behavior over time to reveal retention and lifecycle patterns.View full definition → Explained" - youtube.com/results?search_query=cohort+retention+analysis+explained - search for current beginner-friendly walkthroughs of building cohort retention tables from raw transaction data]

What to do when you see the curve bending down

1. Segment the cohort by acquisition channel. If paid social buyers decay faster than email-acquired ones, you have an acquisition quality problem wearing a retention costume.
2. Check first purchase. Decay often concentrates in one product line, one bundle, or one discount depth rather than the whole cohort.
3. Time win-back to the reactivation window, not to intuition. The window closes at the point where a lapsed customer's unaided return probability falls to that of a cold prospect. Find it by plotting inter-purchase gaps: if 80% of second orders land inside 70 days, the window shuts somewhere around day 90 to 120. Fire earlier and you discount people who were coming back anyway; fire at month six and you pay full acquisition economics for a response.
4. Use the earlier tripwires. Open rates, app sessions and point redemption (the mechanics the loyalty lesson sets out) move before purchase behaviour does, giving you a signal ahead of the curve itself.

Key takeaways

• Retention curves expose churn months before revenue does, because acquisition inflow and cohort decay cancel each other in the aggregate line for two or three quarters.
• Compute retention as (active cohort customers in month N) / (original cohort size) x 100, with a purchase window fixed to your category's repurchase gap.
• Read the shelf, not just the early drop: two cohorts identical at month three can end at 22% or 8% by month twelve, and only the month 6 to 12 slope separates them.
• A seven-point loss in monthly cohorts moves reported orders by under 2% in month one and around 11% by month six, by which point six cohorts are already bought and decaying.
• Exclude in-progress months, compare like seasons, split discounted cohorts out, and separate reactivations from continuous activity before drawing any conclusion.