Customer Lifetime

Customer lifetime is the duration of a customer relationship under a defined start and end rule. A subscription relationship might start at paid activation and end when paid access expires. In a noncontractual business, inactivity is only evidence of possible departure: the customer can return. Define the event being measured before estimating how long customers stay.

Lifetime helps estimate customer lifetime value, but it is also useful for understanding retention and service demand. Distinguish elapsed tenure, total lifetime from acquisition, and expected remaining lifetime conditional on still being active. An account active for ten months has ten months of observed tenure, not a completed lifetime of ten months.

Where one divided by churn comes from

Suppose each customer contributes one active month immediately and then has the same probability c of churning at the end of each month. The chance of reaching a second active month is 1 − c, the chance of reaching a third is (1 − c)², and so on. Expected active months are 1 + (1 − c) + (1 − c)² + … = 1/c for 0 < c ≤ 1. At 5% monthly churn, that expectation is 20 months.

The first-month convention matters. If the question excludes the current guaranteed month, expected additional active months in this model are (1 − c)/c, or 19 at 5% churn. Calendar-time durations, midperiod cancellations and other event rules need a matching model. The reciprocal uses a per-period conditional departure probability; it should not be applied to a cumulative churn percentage or net revenue churn.

Why a blended rate can answer the wrong question

Suppose an acquisition cohort is half customers with constant 2% monthly churn and half with constant 10% churn. Their expected lifetimes are 50 and 10 months, so the cohort mean is 30 months. Initial blended churn is 6%, whose reciprocal is about 16.7 months. Inverting the mean rate loses the distinction between the two groups. With fixed geometric segment churn and the same acquisition weights, the mean of the reciprocals is at least the reciprocal of the mean.

Survivor composition then changes. After one month, 49% of the original cohort remains from the low-churn segment and 45% from the high-churn segment. Next-month blended churn is (0.49 × 0.02 + 0.45 × 0.10)/0.94, about 5.83%, even though neither segment’s behavior changed. A falling aggregate rate can reflect who remains. The same composition issue appears when blended CAC changes with acquisition mix.

This example does not prove that the reciprocal of every observed blended rate understates lifetime. Active-base weights, changing acquisition patterns and tenure-dependent departure rates describe different populations. Separate the target acquisition cohort from the currently active base, and separate total life from life remaining for survivors.

Use incomplete histories without pretending they are complete

At the end of an observation window, active customers have right-censored lifetimes: their eventual departure time is unknown, but it is later than their observed tenure. Averaging only customers who already left can overrepresent short relationships. Treating every active customer as having left at the data cutoff also shortens their recorded lifetime. Retain both observed duration and an indicator of whether departure actually occurred.

Survival analysis estimates the probability a relationship lasts beyond a given duration. A Kaplan–Meier estimate updates survival at observed departures using the number still at risk, while censored observations contribute information up to their censoring time. Its interpretation requires an appropriate censoring assumption: after accounting for the design or relevant grouping, losing observation should not selectively hide different future departure behavior. Customers missing because a tracking system failed only in a struggling segment can violate that assumption.

The area under a survival curve gives expected duration over its range. A restricted mean through a stated horizon, such as 24 months, avoids presenting an unobserved tail as measured history. Estimating full lifetime beyond available follow-up needs extrapolation assumptions. If a subscription business permits pauses and reactivation, decide whether the target is the first subscription spell, total paid months, or the entire relationship with gaps; ordinary time-to-first-churn analysis does not answer all three.

Carry uncertainty into the business decision

Reciprocal churn is especially sensitive at low rates. Changing an assumed rate from 2% to 1.5% changes the geometric lifetime from 50 to about 66.7 months. With 100 active customers and 2% monthly churn, only two departures are expected in one month, so a rate estimate can be unstable. Report the observation window, event count and uncertainty rather than a long lifetime estimate with unjustified precision.

When lifetime feeds an acquisition-cost decision, combine it with revenue, margin and payment timing. A long relationship need not be profitable, and a high LTV:CAC ratio can still involve slow payback. Treat a lifetime beyond the available history as a forecast and show how alternative retention assumptions change the decision.

References

• NIST, Censoring
• NIST, Kaplan–Meier estimation