Churn Clock — live demo · Sixty prototypes, two engines

Live demo

Two channels with the same churn number.

A thousand subscribers, five hundred from each channel, with real signup dates and real cancellations. Everything below is computed in your browser from the per-subscriber records: a Kaplan-Meier fit, hazard by lifecycle period, and a log-rank test. Nothing is hardcoded.

The churn column in the cohort tableIndistinguishabletwo-proportion z-test

Paid search

37.4%

187 of 500 cancelled

Partner referral

37.8%

189 of 500 cancelled

Difference

-0.4 pts

what the dashboard shows

p-value

0.90

no evidence of any difference

A growth team reads this row, concludes the two channels retain the same way, and spends the next quarter buying more of whichever is cheaper. Both numbers are correct. The conclusion is not.

The same subscribers, as survival curvesLog-rank p = 6.3e-51Kaplan-Meier, 95% Greenwood bands

Retained at day 30

64.5% / 99.4%

paid search / partner referral

Retained at day 180

57.9% / 94.7%

paid search / partner referral

Median lifetime

not reached / 299

days until half the cohort is gone

Still active

313 / 311

right-censored, not counted as retained

Log-rank test, observed against expected under equal hazard

Paid search

187 observed vs 84 expected

2.24× the cancellations you would expect if the two channels churned alike

Partner referral

189 observed vs 292 expected

0.65× the cancellations you would expect if the two channels churned alike

The log-rank test weights every cancellation by how many subscribers were still at risk when it happened, which is the information a percentage throws away. Paid search produced 187 cancellations where equal hazard predicts 84. χ² = 225.3 on one degree of freedom.

Where in the lifecycle

The hazard curve names the window.

Hazard is the number a churn rate averages away: of the subscribers who made it to the start of this period, what fraction left during it. Read it and the retention work writes itself.

Paid search, by weekWeeks 2 to 5

90.9% of every cancellation this channel has produced happened between day 7 and day 35. From week seven on, weekly hazard never again reaches 2%: whoever survives onboarding stays.

Peak hazard is 14.2% in days 21 to 27. That is an activation problem, and no amount of win-back email at month six will touch it.

Partner referral, by monthMonths 6 to 11

Near-zero hazard for five months, which is why this channel looks flawless on any 90-day retention report. 86.2% of its cancellations land between day 150 and day 330.

Peak hazard is 18.8% in month 11. That is a value problem at renewal, and the fix is a business review, not a better onboarding checklist.

The other failure

The same cohort, twice, 23.2 pts apart.

These two groups are drawn from an identical lifetime distribution. Same product, same channel, same everything. The only difference is that one signed up in the last forty-five days, so most of them have not been watched long enough to churn yet. The retained percentage counts every one of them as a win.

Retained, recent cohort

77.6%

signed up in the last 45 days

Retained, mature cohort

54.4%

signed up 7 to 14 months ago

Apparent gap

23.2 pts

what the cohort table reports

Not yet observed

77.6%

of the recent cohort, censored not retained

Kaplan-Meier, same two cohorts, first 45 daysLog-rank p = 0.36the curves lie on top of each other

S(30), recent

70.9%

Kaplan-Meier estimate

S(30), mature

67.6%

Kaplan-Meier estimate

Real gap

3.3 pts

inside the noise, and the test agrees

Every cohort comparison you have ever run on a retained percentage carries this bias, and it always points the same way: newer looks better. Which means the cohort you are most anxious to evaluate is the one the number lies about hardest.