The cliff, the core, and the features that decide who stays.
OnCourse doesn't have a retention problem. It has a Day-1 activation cliff — sitting on top of a core that stays like a healthy edtech product should.
~88% of new users never return after day one. Yet of the total Day 1→14 decay, 93% happens inside the first week — after that the curve goes nearly flat at a ~4–5% habitual core that beats the AppsFlyer education D30 floor and matches the cross-industry median.
The diagnosis writes the roadmap: widen the top of the funnel and manufacture a first-session "aha," then protect the core. This briefing quantifies all of it — against sourced edtech benchmarks.
Seven headline metrics, each set against the best external benchmark we could source for consumer / education apps. Read it as a triage chart: where we're bleeding, where we're stable.
Verdicts: Behind · Watch · On par · Ahead. OnCourse figures from PostHog rolling-window retention; benchmarks cited per card.
New-user retention drawn like a vitals monitor. The plunge from Day 0 to Day 1 is the whole story on the left; zoom into the tail and the line you want to see appears — it stops falling.
Share of new users active on exactly day N (classic retention). The drop is near-vertical; everything after week one hugs the floor.
Same data, rescaled to 0–14%. After D7 the line is essentially horizontal — the signature of a real habitual core, not a dying cohort.
Smoother than daily. W1 13.8% → W2 10.4% → W3 8.5%: a gentle, non-collapsing decay. (W4+ omitted — fewer than 16 eligible users.)
Rolling 7-day-window definition, completed cohorts only. Recent large cohorts sit at a stable 14–18%.
Earlier ad-hoc figures (~35–51%) used PostHog's calendar-week first-time retention, which reads higher because a late-week signup's "week 1" window can begin a day later. This report uses the stricter rolling definition throughout.
Education is structurally one of the lowest-retention categories. Two credible benchmark families disagree by method — OneSignal's engaged-app dataset is the optimistic ceiling; AppsFlyer's install-cohort data is the raw floor. We plot OnCourse against both, plus the all-app median.
D1 (12.4%) and D7 (4.8%) trail every benchmark — the activation problem is real and quantified, not a category excuse.
But the D30-proxy core (~4.2%) clears the AppsFlyer edtech floor of 2–3% and matches the ~4% all-app median. The tail is category-competitive.
Subscription lens (RevenueCat 2025): education has the worst annual renewal (24%) but the best weekly (58%) — plan length should match the exam season.
An exam-prep product lives or dies on the daily study habit. DAU/MAU is the north-star here — and right now it's the metric most worth watching.
Daily active users (app opens) over the last 5 weeks — the event began instrumenting in late May, so this is a ramp, not steady state.
Where the daily habit sits today against education norms and best-in-class Duolingo. The gap is the opportunity.
Each weekly bar splits new vs returning openers. Returning share has risen every week (2.9% → 38.7%) as early cohorts age into the product — an encouraging counter-signal beneath the headline ratios.
Caveat: Application Opened only began ramping ~22 May 2026, so the 28-day MAU is inflated by heavy new-user inflow, mechanically depressing DAU/MAU. Treat the 6.1% as directional until a full steady-state window exists.
For new users (cohort of 2,917), which first-week actions are associated with coming back in week one? Base W1 retention is 16.6%. Every feature lifts above it — these are activation candidates, read with adoption and sample size, not proof of cause.
Bar = W1 retention of users who touched the feature in their first 7 days. The white line is the 16.6% base. Mono tag = cohort size · adoption.
50.7% W1 · +34.1pp lift · n=278 · only 9.5% adoption. Large, distinctive lift on a robust sample with real headroom — the highest-leverage "aha" to promote in session one.
39.7% W1 · +23.1pp · n=599 · 20.5% adoption — the most-adopted habit anchor and the most scalable surface to build a daily-streak loop around.
Lifts of +66 / +49 / +41pp but on n=46–96 and 1.6–3.3% adoption. Treat as A/B-validated bets, not headline claims.
Self-selection caveat: engaged users both adopt more features and retain, so lift overstates causal impact. Daily Review (n=15, 73%) and VideoFlix (n=12, 83%) were excluded for n<30.
Two different questions: how many people touch a surface (reach), and how hard the engaged ones lean in (events per user). The gap between them exposes the funnel.
Top-of-funnel signup events dominate. Discrete product use sits far lower — the acquisition→activation leak in one chart.
Among engaged users (n≥150). The intensity leaders map onto the retained behaviors: deep Exercise & Daily Plan use.
Only 11.9% of monthly-active persons ever fire Application Opened. Onboarding is also bloated — 24.8 pages & 23.98 button-clicks per user.
Fires 3.97× per user across 1,978 users — a monetization-and-trust leak that directly threatens the high-retention paid segment (40.3% W1). Likely driver of involuntary churn at the month-3 peak.
Two cuts dwarf everything else: platform and monetization. Both point to concrete, ownable programs.
iOS retains 1.82× better than Android — and Android is the largest cohort, so its low rate is the single biggest absolute drag and the biggest opportunity.
The strongest signal in the whole dataset. Causality runs both ways — but getting users to a trial state early is itself a retention lever.
A 3.4× retention gap.
Prioritised by leverage. The cliff is the P0 — everything downstream is starved of volume until activation improves. Then build the daily loop and protect the paid core.
Signal: Application Opened (cold launches) as the activity event; first occurrence defines a new-user cohort.
Curves: eligibility-adjusted — each day/week only counts cohorts old enough to have realized it, removing time-truncation bias. "Day-N classic" = active on exactly day N.
Feature lift: cohort of 2,917 users whose first open was 14–70 days ago; feature "used" within first 7 days; retained = re-opened in days 7–14.
Tooling: PostHog HogQL, project 61503, timezone Asia/Kolkata, as of 26 Jun 2026.
Application Opened is effectively newly-instrumented (ramp from ~22 May); stickiness ratios are directional.