Adding GPT-5 coaching, gamification and four onboarding rebuilds did not lift day-one retention in a solo habit app

A bootstrapped app with tens of thousands of installs keeps almost nobody after the first day and earns very little, and the founder hoped new features, including AI coaching, would fix retention and conversion.

What the founder tried and measured: 1. Rebuilt onboarding four times, the latest walking users through creating and completing a first habit. 2. Added an AI coaching chat (GPT-5, personalised habit advice with agentic tool calls) as a paid feature, alongside gamification (quests, XP, levels, streaks), a WearOS app, notifications and an analytics dashboard. 3. Tracked day-1, day-7 and day-30 retention and free-to-paid conversion through each change. 4. Queried the database for power-user patterns, using the variation in each user's daily completions to separate steady users from boom-and-bust users, and compared them with paying customers. 5. Analysed all cancellation reasons. 6. Emailed the nine heaviest users offering free Pro for interviews. 7. Posted the data on Hacker News and asked what he was missing; commenters pointed to a generous free tier, too small a cohort to analyse, and no marketing.

Day-one retention stayed at 4.2% through every change, including the AI coaching. The AI chat converted 11.54% of users who discovered it, against 0.8% overall, but only 2.4% of users found it. Overall: about 35,000 installs over three years, $86 monthly recurring revenue from 20 subscribers, about $4,091 lifetime revenue, and about $905 a month in costs. Only 2 of 9 power users answered interview requests. Cancellations: 43% did not use it enough, 30% cost, 12% found a better app.

Self-reported figures in a forum post; no dashboards shown. AI is a product feature here, not a workflow that runs the business, and the post does not explain how the coaching works in detail. A commenter noted the power-user cohort (nine people) is too small to support conclusions, and later replies give lower active-user figures than the original post. Another commenter asked whether the replies were written with AI. Mobile consumer app, so it transfers only partly to other small software products. October 2025.

The source does not say.

• Hacker News (Nek_12) ↗ Firsthand write-up, published October 28, 2025

Source published: October 28, 2025. Last reviewed here: October 24, 2026. Spot a mistake? Tell us.