The business entered the engagement with a churn problem that looked, on the surface, like a marketing problem. Voluntary cancellations ran at roughly 3.1% per month across a paying base of just over 600,000 subscribers. The retention motion was almost entirely calendar-driven: a discounted renewal offer in month 11, a generic winback sequence in the two weeks after a cancellation event, an annual price-rise letter that triggered a predictable spike of departures every January.

The internal hypothesis was that the save offers were too shallow. The actual problem was that the offers were aimed at the wrong subscribers. Roughly a third of the discount budget was being spent on readers who would have renewed at full price; another third was being spent on readers who had already mentally departed and would not have been saved at any price. The middle third — the subscribers whose behavior had quietly shifted but who had not yet hit a cancel button — was being touched too late and too generically.

What the data actually showed

The first sweep was diagnostic. Eighteen months of behavioral data — article-level reading events, newsletter open and click patterns, app-versus-web session mix, payment retries, geographic and device signals — were joined against the cancellation history. Several patterns surfaced quickly. Reading frequency in the trailing thirty days was a far stronger predictor of cancellation than tenure, NPS, or any of the survey-based signals the team had been tracking. A drop from daily to weekly reading inside a single 30-day window was associated with a roughly 4x increase in cancellation probability over the next ninety days. Failed-payment events were a near-certain churn signal if not recovered within seven days, but the existing dunning sequence was running on a fourteen-day cadence.

Acquisition cohort mattered more than expected. Subscribers acquired through introductory price promotions in 2024 churned at almost double the rate of subscribers acquired through editorial referral, even controlling for tenure — and they responded differently to save offers, with discount-acquired cohorts showing strong sensitivity to further discount but flat response to content-based interventions like newsletter resubscription or topical alerts.

What was built

A churn probability model, refreshed daily, scored every active subscriber on three horizons: 30-day, 60-day, and 90-day cancellation likelihood. Inputs spanned behavioral telemetry, payment event history, and acquisition cohort metadata. The model output a probability and a top-three feature attribution — the signals that were driving the score for that subscriber — which fed the intervention layer.

The intervention engine was tiered. High-risk subscribers with high predicted price elasticity received a deep, time-bounded save offer through their preferred channel. High-risk subscribers with low predicted price elasticity received content-based interventions — a curated topical newsletter, a podcast resubscription prompt, an editor-direct outreach for the highest-value tier. Medium-risk subscribers received lighter touches calibrated to the specific feature attribution: a payment-retry assist if the signal was billing-driven, a reading-cadence nudge if the signal was engagement-driven. Low-risk subscribers received nothing — a deliberate choice that recaptured roughly $640K of discount margin previously spent on readers who would have renewed at full price.

The dunning sequence was rebuilt on top of the same model: failed-payment events triggered an immediate retry plus an SMS or in-app prompt within 48 hours, replacing the prior 14-day cadence.

Result

Over the first nine months of operation, the intervention engine recovered $3.4M of annualized subscription revenue — measured against a holdout cohort that continued to receive the calendar-driven retention motion. The lift in save rate against that control was 47%. Monthly voluntary churn moved from 3.1% to 2.4% across the full base, a drop concentrated in the medium-risk segment where the prior motion had effectively been silent. The discount-margin recapture on low-risk subscribers — readers no longer reflexively offered a renewal discount they did not need — added roughly $640K of contribution that was previously being given away.

The harder-to-measure outcome was the shift in how the retention team operated. Save campaigns stopped being calendar artifacts and started being responses to specific, scored, attributed signals about specific subscribers. The price-rise letter still goes out every January, but it now goes out into a base where the most price-sensitive segments have already been pre-segmented and pre-handled.