Churn is four problems added together
Every churn conversation starts with a single number, and the single number is the reason the conversation goes nowhere. A percentage of customers left. It does not say whether they left because a card expired, because they never figured out what they bought, because they finished what they came to do, or because the thing was not worth the money. Those are four businesses problems with four owners, and they are being handed to whoever runs email.
So the first piece of work is not an intervention. It is a decomposition.
Involuntary. The payment failed. The card expired, the charge declined, the billing address went stale, a fraud rule fired. These customers did not decide anything. They are usually a fifth to two fifths of total churn, they are the cheapest to recover of any group, and the fix is dunning logic and card updater coverage rather than a campaign.
Never activated. They bought, and then never reached the value they bought for. The box arrived and stayed unopened, the account was created and never configured, the second order never happened. This group churns early and looks like a marketing failure while usually being an onboarding one.
Completed. They got exactly what they came for. The project ended, the child outgrew the size, the course finished. This is not a failure at all, and treating it as one produces win-back campaigns aimed at people with no remaining need. The right response is a different offer, or an honest acceptance that this cohort has a natural life.
Dissatisfied. They got the value and decided it was not worth the price. The product, the service, or the price is the problem, and the only useful thing marketing can do is report it accurately to the people who can change it.
Size these four before proposing anything. The mix decides the entire program, and it is different at every business I have looked at.
A churn rate is a symptom presented as a diagnosis. Nobody would accept “the patient has a temperature” as a treatment plan, but that is exactly what a monthly churn percentage is.
Cohorts, because the average hides the shape
A blended churn rate mixes a customer in their first month with one in their third year and reports the result as a single number. It is arithmetically correct and analytically useless, because everything you can act on lives in when people leave, not how many.
Group customers by the period they were acquired and follow each group forward. The curve tells you where the business is actually leaking:
- A steep early drop is an activation problem. Customers are not reaching value fast enough, and everything you spend downstream is spent on people who never got started.
- A steady, gentle decline is normal attrition, and the lever is frequency and breadth rather than retention messaging.
- A cliff at a specific tenure is structural: an annual renewal, the end of a promotional price, the point where a consumable would have run out.
- Cohorts that get progressively worse mean acquisition quality is falling, which is a media problem wearing a retention costume. This one gets misdiagnosed constantly, and it is the most expensive mistake on the list.
That last case is worth dwelling on. When retention degrades cohort over cohort while the retention program has not changed, the program is not what broke. The customers changed, usually because acquisition scaled into a worse audience, and no amount of lifecycle work fixes a sourcing problem.
The first ninety days do most of the work
Across almost every business model, early behavior is disproportionately predictive of everything after it. The customer who reaches the second order, or the habit, or whatever your equivalent of first real value is, behaves differently forever. The customer who does not is largely gone regardless of what you send them in month seven.
This makes the onboarding window the highest-leverage stretch of the entire lifecycle, and it is routinely under-resourced relative to win-back, which is the lowest-leverage stretch. The interval before the second purchase is the most predictive number most brands do not track, and time-to-second-order is a better early warning system than any churn model built on the same data.
The practical work here is unglamorous: get the customer to the thing that delivers value, remove the steps between purchase and that moment, and set the replenishment or usage rhythm early rather than letting the customer invent one.
Intervene before the decision, not after it
By the time someone cancels, the decision is old. It was made weeks earlier and the cancellation is just the paperwork. A churn program that begins at the cancel screen is intervening at the one point where the customer has already committed to leaving and where your only remaining tool is a discount.
The risk window is where the money is. That means detecting behavioral drift against each customer’s own baseline, not a global one, and acting while the customer is still ambivalent. It also means resisting the reflex to lead with a price cut, because the customer who is retained with a discount at the first sign of drift has now been taught what to do the next time they get bored.
And it means fixing the involuntary pool first. Payment recovery does not require anyone to be persuaded of anything. It is operational, it is boring, and it is nearly always the largest recoverable number sitting in the business.
Prove it against a holdout, or do not claim it
A retention program without a control group produces a report of customers who stayed, not evidence that the program kept them. Some share of every at-risk audience was going to stay anyway, and a program with no holdout will happily bill you for all of them.
Hold back a percentage of each audience, read performance as the difference between the two groups, and hold the retention program to the same standard I hold paid media to. This is what separates a retention program from an expensive way to be nice to people who already liked you, and it is the only version of the number that survives contact with a CFO.
How I work this lane
I decompose the churn first, because the mix determines the program and the mix is never what the team assumed. Then I work the cohort curves directly rather than receiving them summarized, because the behavior worth finding is always in the detail a blended average erases. Then interventions get built in order of recoverable value, which almost always means involuntary churn and the activation window before anything that could be described as a campaign.
Further reading
- Your churn is a payments problem more often than a loyalty problem
- Your retention rate is lying if you don’t cohort it
- Subscription churn is a pricing problem in disguise
Where this fits
Churn tells you where customers leave. Subscription economics tells you what each departure costs and whether the model compounds or leaks, and loyalty and win-back is what you run for the customers this work did not catch in time.
- Subscription economics →, cohort payback, order frequency, and the unit math underneath the model
- Loyalty and win-back →, program design and whether it is generating incremental behavior
- ← Back to Lifecycle & Retention