Your churn is week six, not four percent

Two numbers sat side by side in the board deck we were handed on a retention engagement last spring. Monthly churn: 3.8 percent, stable for a year. Retention spend: up 60 percent, on a customer-success tool, a win-back email program, and two new hires. The founder's question was reasonable and slightly annoyed: if churn is stable and low, why does revenue keep flattening?

Because 3.8 percent monthly is not a description of anyone's behaviour. It is an average smeared across customers who were never going to leave and customers who were already gone in every way except the invoice.

We rebuilt the number as a survival curve, the same tool medicine uses to describe how long patients live after a treatment, pointed at subscriptions instead. Not "what fraction left this month" but "given a customer reached day N, what is the chance they reach day N plus one." The curve did not slope gently. It fell off a cliff.

The company did not have 3.8 percent monthly churn. It had a week-six problem wearing an annual average as a disguise.

Plotted as a hazard curve, the risk of cancelling was near zero for the first month, spiked violently between weeks five and seven, and then collapsed to almost nothing. A customer who survived week eight stayed for years. Nearly two thirds of all cancellations, ever, happened inside a three-week window.

The monthly average hid this completely, because every month contains a mix of brand-new customers passing through the danger zone and old customers who are effectively immortal. As the customer base aged, the blended rate even drifted down slightly, which the dashboard reported as improvement while the week-six cliff stood exactly where it always was.

The mechanism, once the curve pointed at it, took two days to find: week six is when the first full billing cycle landed after the trial credit ran out, and it is also roughly when the initial setup enthusiasm wears off. Customers who had wired the product into a weekly routine sailed past the invoice. Customers who had not, looked at the charge and remembered they had stopped logging in.

(The win-back program, incidentally, was emailing people at day 90. It was reaching customers who had made their decision two months earlier, which is why its numbers looked like a coin toss.)

Survival analysis needs three ingredients your billing system already has: when each customer started, when they cancelled if they did, and the discipline to treat still-active customers as unfinished stories rather than successes. That last part is the technical heart of it. A customer eight weeks in who has not cancelled is not a retained customer; they are a data point that says "survived at least eight weeks." Getting this censoring right is exactly what separates a hazard model from a spreadsheet, and it is where naive retention math quietly lies.

From there, the curve can be split by anything you can segment: acquisition channel, plan size, onboarding path. This is where the money conversations start. On that engagement, customers from one paid channel had a week-six cliff nearly twice as steep as organic signups, which repriced the channel's real acquisition cost on the spot. The blended CAC had been flattering it for two years.

Retention spend aimed at the average lands nowhere. Aimed at the week the hazard curve names, it lands on the only customers who were persuadable.

The interventions wrote themselves once the window was known: the onboarding push compressed into the first month, the check-in call moved from day 90 to day 30, the invoice email for first-cycle customers rewritten to restate what the product had done for them that month. Six months later the week-six spike had flattened by about a third, which, for that base, was worth more than every point of blended churn the previous year of spend had chased.

Averages answer "how much." Curves answer "when," and when is where the leverage is, because interventions happen in time, not in aggregate. At TensorLabs we build this as a standing view, not a one-off study: hazard curves by cohort and channel, refreshed with the billing data, so the danger window is a monitored number instead of a forensic discovery.

If your churn is reported as one percentage, somewhere in it there is almost certainly a cliff with a date on it. Reply with roughly how your retention is measured today and we will tell you whether a survival view would change the answer, and what it would take to build from the data you already have.