A new product competes with habits that already work. In the one category with independent data, most people who tried an app had stopped within two weeks, so launch is where the real evidence begins.
- 01
ModerateMost new users leave quickly. Across 93 mental-health apps, the median share of people still using the app after 30 days was 3.3%.
Baumel et al. (2019), Journal of Medical Internet Research; independent panel data, one category of app, Android only
- 02
RobustPeople adopt what is clearly useful, fits how they already work and is simple to start.
Tornatzky & Klein (1982), IEEE Transactions on Engineering Management; meta-analysis of 75 articles
- 03
ModerateUse and time saved are not the same as changed work. In a six-month randomised trial about 80% of workers given an AI assistant used it and spent about two fewer hours a week on e-mail, with no detectable change in the quantity or mix of their tasks.
Dillon et al. (2025), a working paper from Microsoft Research and Harvard, not yet peer reviewed; 66 firms, 7,137 knowledge workers
- 04
ModerateThe launched version is unlikely to be the best one. Where every change is tested, about one in three improves the measure it was built to improve.
Kohavi et al. (2012), Microsoft; the company’s own record of its experiments
- 05
ModerateOnboarding should match the product. Tutorials raised play time by up to 29% in a complex game and made no significant difference in two simpler ones.
Andersen et al. (2012), CHI; over 45,000 players, eight tutorial designs
- 06
ModerateNew routines take weeks to months, not days. Among the 39 people whose habit curves could be modelled, the median time for a simple behaviour to become automatic was 66 days, with estimates from 18 to a projected 254.
Lally et al. (2010), European Journal of Social Psychology; 96 volunteers over 84 days
Robust: meta-analysis, replication or very large samples. Moderate: one strong study, or several with limits. Weak: small studies, or run by an interested party.
/ Where the evidence stops
We found no controlled study that compares launching and leaving with launching and improving, and none that shows an outside studio improves adoption; our case rests on the evidence above. Improving against a number can also harm people: in one experiment manipulative sign-up designs more than doubled sign-ups (Luguri & Strahilevitz, 2021). So we measure effective use and benefit, not log-ins, and treat any gain won by misleading people as a failure.