Tags: analytics commerce concept

Retention Curves

Date: 2026-08-16

---

Plot what proportion of a cohort is still active over time. The only question that matters is whether the curve flattens — a curve that reaches a floor is a business; one that goes to zero is a leaky bucket, however good the acquisition is.

---

What it is

A retention curve plots the share of a cohort still active at each period after acquisition.

100% ┤●
     │ ╲
  35%┤  ●
     │   ╲
  25%┤    ●──●──●──●──●   ← flattens: a stable base exists
  17%┤
     │        ●
     │         ╲
   5%┤          ●──●──● ← approaches zero: no repeat business
     └──┬──┬──┬──┬──┬──┬
        1  2  3  4  5  6   months

The floor is the finding. Whether it settles at 25% or at 4% determines whether every customer acquired has ongoing value or is a one-off — and that decides what you can afford to pay for one.

Defining “active”

Every retention number depends on this, and it’s rarely stated:

Definition	Suits
Purchased in the period	Retail, subscriptions
Visited	Content, marketplaces
Used a specific feature	Product analytics

For non-subscription retail the harder question is the period length, because there’s no renewal date to anchor to. A consumables business might use monthly; a furniture retailer’s natural cycle is years, and monthly retention would read as near-zero for a healthy business.

Set the period from your own repurchase interval, not from a convention. Using a monthly grid on a category with a six-month cycle produces a curve that looks catastrophic and means nothing.

Reading the shape

• Flattening — a stable core of repeat customers. The single most important property
• Approaching zero — every customer is effectively one-off, so growth requires acquisition forever and lifetime value is roughly first-order value
• A smile — falls then rises. Usually a measurement artefact: seasonal repurchase, or Identity Stitching improving so returning customers are recognised later. Genuine resurrection exists but is rarer than the chart suggests
• A cliff at one period — usually mechanical. A subscription term ending, a discount expiring, or a tracking change

The traps

Incomplete cohorts. The newest cohort has no month-6 value. Plotting it as zero, or averaging it in, drags the curve down and makes recent cohorts look worse than they are.

Survivorship. The curve describes people who reached period 1 to be measured. Anyone who churned before your measurement started isn’t in it — Survivorship Bias.

Identity decay is indistinguishable from churn. A customer who returns on a new device with an expired identifier is counted as churned and as a new user. So your retention curve is pessimistic by whatever your identity loss rate is, and that rate differs by browser — Browser Privacy Restrictions, User Counting.

Anonymous users have no curve at all. Retention analysis only exists for the identified population, which is your best customers — so the curve is optimistic on behaviour and pessimistic on identity at the same time — Anonymous and Identified Users.

What to do with it

• Compare cohorts at equal age, never at equal calendar date — Cohort Analysis
• Look at revenue retention alongside count retention. Fewer customers spending more is a different diagnosis — Cohort Revenue
• Use the floor for lifetime value. The flattened proportion times average repeat value is the honest basis, rather than extrapolating an unflattened curve — Customer Lifetime Value
• Judge changes by the early periods. Month 1 moves first and is the fastest read on whether a lifecycle change worked
• Segment by acquisition channel. Channels differ enormously in the retention of the customers they bring, and blended curves hide it — which is one of the strongest arguments against allocating budget on first-order return alone — Channel Mix