Preventing churn: Score accounts on outcomes, not feelings

Some of the customers who will cancel this year are pleasant in every conversation you have between now and then. That is why churn work blindsides the teams that own it.

Churn is a lagging indicator of a decision made months earlier: by the time somebody cancels, the reasons were settled long before anyone on your side noticed. Two consequences follow, and both change who owns the work. Most of your churn was determined upstream, in activation and habit, where a team told to “fix retention” has no authority. And the churn that is genuinely preventable is a narrow band: accounts where something changed recently, visibly, and reversibly.

The signal companies reach for to find that band is the one that does not work.

Start with the arithmetic, because it is worse than it feels

Monthly churn compounds, and the intuition for compounding is bad. Three percent a month sounds like a rounding error and is not: it loses about 31% of a cohort of customers in a year, which means finding replacements for roughly a third of your accounts annually before growing at all.

That is not a benchmark and it needs no source: surviving a year at 3% monthly is 0.97 to the twelfth power, or 69.4%, which you can check on a calculator in five seconds. It is the one number on this page that is not open to argument, which is why it leads.

The same arithmetic tells you where you stop. New monthly recurring revenue divided by your monthly revenue churn rate is the level at which what you win and what you lose cancel out and growth goes to zero. Win $5,000 a month of new revenue against 10% churn and you plateau at $50,000, however well you market. Rob Walling, who founded Drip and now funds bootstrapped software companies, recommends every founder calculate that number and keep it on a dashboard.¹ It is a ceiling with your own name on it, and nobody else’s figures go into it.

The same figure stated as a half-life is more vivid: at 3% monthly, half your customers are gone in about two years, at 5% in around thirteen months, and at 1% in nearly six years. Same arithmetic, harder to ignore.

It also explains why small improvements are worth disproportionate effort. Going from 3% to 2% monthly takes annual loss from 31% to 22%. Going to 1% takes it to 11%. Each percentage point is worth more than the last, in the direction that helps.

How should you measure churn?

A business can lose customers steadily and still grow revenue from the ones remaining, which is what net revenue retention above 100% describes. That is a real result and not a way of hiding logo churn: both numbers matter, and reporting only the flattering one is the most common form of self-deception here.

The one to watch depends on the question. Logo churn asks whether people want the product. Net revenue retention asks whether expansion outruns the losses. A company with excellent net retention and bad logo churn is growing on a shrinking number of increasingly important accounts, which means the business depends on fewer and fewer customers.

Satisfaction does not predict retention. Outcomes do.

If that holds, it reorganizes most health scoring, and it is uncomfortable because it invalidates the easiest data to collect.

Satisfied customers cancel. Survey scores, sentiment, relationship warmth and stated intent all correlate weakly with whether an account renews. Greg Daines, who has built a practice on retention data, puts it directly: what predicts retention is whether the customer is getting the result they bought, and satisfaction is a poor proxy for it. That is a practitioner’s finding from client work rather than published research with an inspectable method, and it is included because the alternative most companies use has no better evidence and worse face validity.²

The distinction is sharper than it sounds. “Do you like us” and “has this delivered what you needed” produce different answers from the same person, and only the second one predicts anything. People answer surveys about the relationship rather than the result.

A workable starting score, weighted toward outcomes, not activity:

Weight them however your own retrospective says, sum to whatever scale you like, and use the score to rank accounts for attention, not to declare them healthy. The point is the ordering of the inputs, not the arithmetic.

Why these inputs, in this order

The ordering is the argument, so here is what each input is doing.

Outcome achievement. Whatever the customer bought this to do, and whether it is happening. Hardest to instrument, worth more than everything else combined.

Depth of use across the account. Not total usage, which rises with account size, but how many distinct people and teams are active. A tool one person depends on is one resignation away from cancellation; spread across several teams it survives that.

Accumulated investment. Configuration, history, integrations, data that would have to be rebuilt. This is switching cost, and it is the difference between a customer who is unhappy and a customer who is leaving.

Trajectory, not level. A large account whose usage fell 40% this quarter is in more danger than a small one that has been flat for two years. Direction is the signal; absolute numbers mostly measure how big the customer is.

One boundary worth setting before anyone proposes a model: a weighted score you can explain is the right tool until you have years of churn history and thousands of accounts to learn from. Below that there is not enough signal for anything cleverer, and a model nobody can interrogate is worse than a crude score everybody understands.

Two warnings about the whole exercise. A health score that has never predicted a save is decoration, so check it retrospectively against accounts that actually churned before trusting it. Check it forwards too: pull your twenty healthiest-scoring accounts and see how many grew last quarter. If few did, the score is measuring activity, not outcomes, whatever its inputs claim. And a score used to rank accounts for attention will drift toward whatever is easy to measure unless somebody defends it.

The narrow band you can actually save

Most churn is not preventable at the point you notice it. The exceptions are specific and worth naming, because they are where intervention pays:

The champion left. The most actionable signal available in B2B: a named contact goes quiet or their email bounces. Whoever inherits the account never chose you and has no investment in your success. The window is short and the correct response is immediate.

Usage fell off a cliff, recently. A step change, not a drift, in an account that was healthy. Something specific happened, and it is usually knowable by asking.

A renewal is approaching with an unproved outcome. Predictable timing, so no excuse for surprise, and the only real fix is evidence assembled well in advance, not a conversation.

A support experience went badly. This one needs care in the opposite direction. Contact is not itself a danger sign, and there is a reasonable argument that a customer raising problems is more engaged than a silent one, though we know of no published measurement either way. What predicts departure is not the volume of tickets but the resolution: an unresolved problem in an account with low accumulated investment is an exit in progress.

The signals are much stronger together than separately, and this is the cheapest thing to automate: alert when any account shows two or more of them at once. Declining usage on its own is noise; declining usage plus a departed champion is an account leaving. A support spike plus an abandoned feature is a customer who tried to make something work and gave up.

Everything outside that band is upstream work wearing a retention label. Which is not an argument for ignoring it, only for putting it where it belongs: an account that never activated properly is an onboarding problem you are meeting eighteen months late.

Involuntary churn is the cheapest thing you will ever fix

A meaningful share of cancellations are not decisions at all. Cards expire, payments fail, invoices go to somebody who left. The customer intended to stay and the billing system removed them.

This is unglamorous and has the best return of anything on this page, because the customer already wants to be there. Retry failed payments on a schedule rather than once, tell a human being before you tell nobody, keep card details current, and make sure the billing contact is not a single individual who may have moved on. Any of it can be shipped in a sprint.

What to measure

1. Logo, gross revenue and net revenue churn, all three, dated. Reporting one is how a business surprises itself.
2. Voluntary versus involuntary, separated. Nobody can size the cheap fix until these are apart.
3. Retention by cohort and by activation status. Users who reached your validated aha moment against those who did not. The gap is what activation work is worth in retention terms, and it is usually large.
4. Your health score’s hit rate, checked backwards. Of the accounts that churned last quarter, how many did the score flag in time? A score nobody has audited is a comfort object.

Churn benchmarks by stage and segment exist, and comparing yourself to them tells you about their definitions rather than your business.³ Churn depends on contract length, segment, payment terms and what you count as a lost account, so the comparison that means anything is against your own prior cohorts.

Where to look first

1. Split voluntary from involuntary churn. If involuntary is more than a rounding error, fix the billing path before anything else on this list.
2. Audit your health score against last quarter’s departures. Keep the inputs that flagged them; delete the ones that did not.
3. Add one outcome signal, replacing one activity signal. This is the highest-value change and the most work.
4. Build a champion-departure alert. Bounced email, dormant primary contact, a title change you can see. Then define what happens within a week of it firing.
5. Compare retention by activation status and take the result to whoever owns onboarding, because that is where the number gets moved.
6. Interview five recent departures and ask what they were trying to achieve rather than what they thought of the product. The answers will be about outcomes, which is the point.

Churn is the price of everything upstream being imperfect. There is one more lever on the same problem, and it is the one most product-led companies handle worst, because it sits between the product and the customer’s wallet: what you charge, what you charge for, and where the line between free and paid falls. Setting PLG pricing decides how much of the value you created you actually collect.

Updated August 2026