Improving Retention

• Choose the right retention horizon for a change
• Run retention experiments with appropriate limits
• Build win-back and lifecycle programs that close the loop

• Short-horizon wins fund long-horizon bets; without the system, teams chase only one or the other.
• Win-back programs recover users who'd otherwise vanish quietly — cheap retention at scale.
• Retention experiments are hard (long lags, multiple interactions) so honest design and measurement matter even more.
• Post-launch decay is real: features ship, retention 'improves', then drifts back bottom as novelty ends.

Immediate care (short), rehab regimen (mid), lifestyle and long-term health (long). Trying to fix a fractured habit with a pill or a long consultation with a bandage both fail — each horizon has its own intervention.

• Classify candidate improvements by horizon and pick a horizon-matched metric for each.
• Short: run A/Bs on activation/onboarding with D7/D14 readouts; ship and re-baseline.
• Mid: build lifecycle automation — win-back emails for lapsing users (define lapser by gap), cadence triggers, feature spotlighting — measured on recovered-D30.
• Long: ship the value-loop product bets; track on mature cohorts with a 90-day+ reader and guardrail the near-term.
• Post-launch: monitor the curve and hold a re-verification ritual at +30/+60/+90 days.

• A subscription product's win-back program (day-24 lapsed coupon + enable-push nudge) recovered 11% of dormant users into another 90-day active streak, netting positive payback.
• A project tool shipped a 'reminders' cadence feature measured on weekly-active-share; plateau stability across three weeks beat the novelty test and shipped.
• A fintech's long bet — an 'auto-reinvest' value loop — was measured only on 6-month cohorts, with near-term D30 held as a guardrail against the interim dip.

• Set a horizon-reasoned metric before building; a morning-of-D90 measurement ritual beats a year-end slide.
• For mid/long programs, define the guardrails (support volume, contention) alongside the goal.
• Log everything — win-back audiences, offers, cohort response — into a lifecycle library the team reuses.
• Conduct a re-verification at 30/60/90 days; the curve is the referee, not the launch slide.

• A/B-testing a win-back as if it were a D7 lever, measuring the wrong horizon and killing a working program.
• Building the program (win-back, cadence) but never defining 'at-risk' precisely, so campaigns hit already-lost users.
• Shipping a long product bet and abandoning the curve at launch — the famous +10% then −15% two months later.
• Letting retention dashboards double as self-congratulation: any improvement that disappears by day 60 didn't happen.

• If win-back 'recovered users' don't stay, check the recovery success metric used — recovered-to-first-session vs recovered-to-30-day-active tell different stories.
• When a long bet's curve reacts, look at the counterfactual cohort (the holdout or pre-launch era) to separate novelty.
• If cadence features peak then fade, the trigger frequency is overshooting value — tune constancy, not volume.
• Post-launch decay isn't failure until the curve is below baseline at the horizon you promised to measure.