Cohort            500 subscribers, 313 still active
Danger window     days 7 to 35
Peak hazard       14.2% in days 21 to 27
Concentration     90.9% of all cancellations land in that window
Kaplan-Meier      S(30) = 64.5% [60.1% to 68.9%]
                  S(180) = 57.9% [53.3% to 62.6%]
Median lifetime   not reached: the curve plateaus above 50%
Coverage          no subscriber in this cohort is older than 180 days

Nine in ten of the people who leave this channel are gone before day 35, and the curve is nearly flat afterwards. That makes it an onboarding problem with a deadline, not a retention problem. Win-back email at month six cannot reach these people because they were already gone before week six. The coverage line matters: this channel has no history past day 180, so nothing here says what happens in year two.

Cohort            500 subscribers, 311 still active
Danger window     days 150 to 330 (months 6 to 11)
Peak hazard       18.8% in month 11
Concentration     86.2% of all cancellations land in that window
Kaplan-Meier      S(30) = 99.4% [98.7% to 100.0%]
                  S(180) = 94.7% [92.5% to 96.9%]
Median lifetime   299 days
Early signal      hazard below 1% for the first five months

Nothing in the first five months predicts this. Any 30, 60, or 90-day retention report gives this channel a clean bill of health, which is why it keeps getting more budget. Half the cohort is gone by day 299. The intervention has to be scheduled off tenure, around month five, and it is a value conversation rather than a product fix.

Cancelled to date 37.4% vs 37.8%
Two-proportion    z = -0.13, p = 0.90 (finds nothing)
Log-rank          chi-square = 225.3, p = 6.3e-51
Observed/expected Paid search: 187 vs 84 (2.24x)
                  Partner referral: 189 vs 292 (0.65x)

The percentages are half a point apart and a test on them finds nothing, which is the answer a churn table would give you. Using the timing of every cancellation and the number of subscribers still at risk when it happened, the two channels are not remotely alike. Do not pool them, do not give them the same playbook, and do not price acquisition off a blended retention assumption.

Retained          77.6% vs 54.4%
Apparent gap      23.2 points in favour of the recent cohort
Censored          77.6% of the recent cohort has not been observed long enough
Kaplan-Meier      S(30) = 70.9% vs 67.6%
Log-rank          p = 0.36 (no difference)

Somebody is about to present this gap as evidence that a recent change worked. It is not. These two groups have the same survival curve over the window where both have been observed, and the entire gap is that one of them has been a customer for three weeks. Churn Clock will not report a retained-percentage comparison between cohorts of different ages, because that number is wrong in a predictable direction every single time.