There are now more than 10,000 mental health apps available in major app stores. Most of them are abandoned within two weeks of download.
The field has spent years trying to solve this with better onboarding, smarter notifications, and habit-forming mechanics borrowed from consumer technology. The assumption has been that if people keep coming back, the product is working. Research published over the past two years suggests that assumption was wrong, and that the features most associated with clinical outcomes are quite different from the ones most teams have been building toward.
What the meta-analysis found
A systematic review and meta-analysis published in npj Digital Medicine in April 2025 is the most comprehensive examination of this question to date. The researchers analysed 92 randomised controlled trials involving 16,728 participants across a range of mental health apps. The headline finding on efficacy was encouraging: apps significantly improved clinical outcomes compared to controls, with a medium effect size (Hedges g = 0.43). When designed with clinical grounding, digital mental health tools work.
But the finding on persuasive design was harder to ignore. Across all 92 trials, no significant association was found between the number of persuasive design features an app used and either its clinical efficacy or its ability to retain users. More streaks, more notifications, more variable rewards. None of it moved the needle. The design logic borrowed from consumer technology was, statistically, adding nothing.
The distinction that matters
A 2024 scoping review in PLOS Digital Health examined persuasive design frameworks across mental and behavioural health platforms and identified a tension that runs quietly through the field. Persuasive design, as applied in consumer technology, is built to influence behaviour in the short term. Therapeutic design is concerned with something slower: sustainable psychological change. The features that serve one goal do not automatically serve the other, and in some cases actively work against it.
The attrition data bears this out. A qualitative study from the University of Manchester, published in Digital Health in late 2024, found that poor design and lack of personalisation are the most consistently cited reasons people abandon mental health apps. Not because the product failed to hold their attention. Because it did not feel like it understood them, or was genuinely working with them toward something.
That is a different kind of leaving.
Digital therapeutic alliance
The concept gaining traction among researchers is digital therapeutic alliance: the quality of the working relationship a person develops with a digital mental health tool over time. It draws on one of the most replicated findings in clinical psychology, that the alliance between a therapist and patient is among the strongest predictors of treatment outcomes, and asks whether something similar can form with a product.
An integrative review published in JMIR Mental Health in February 2025 found evidence that it can. Across AI-powered chatbots and self-guided apps, users were forming genuine working relationships with digital tools, built on trust, perceived empathy, and a sense of shared direction. The design features most associated with strong alliance were personalisation, transparency about what the product could not do, and consistency of response over time.
A qualitative study from the University of Bath and King's College London, published in BMC Public Health in July 2025, added texture to this. Participants described what made a mental health app feel like a real therapeutic relationship: honest progress tracking, clear reasons for why certain exercises were being suggested, feedback that actually reflected their experience rather than generic encouragement. None of these are features that show up well in an engagement dashboard. They are features that earn trust over time.
What responsible design actually requires
A 2025 review in World Psychiatry, drawing on nearly 200 trials of depression and anxiety apps, found that human guidance consistently improves clinical effect sizes in digital mental health tools. A product that builds in real moments of human connection, with a clinician, a coach, or a peer, performs better than one optimised purely for autonomous self-management.
A Frontiers in Psychiatry analysis from April 2025 raises a structural issue worth naming: most apps are currently rated on user experience rather than clinical grounding. A product can score highly on ease of use while offering no evidence-based therapeutic content. Until evaluation frameworks catch up with that reality, the incentives point in the wrong direction.
What this means for builders
Session length, return visit rates, streaks, notifications: none of these are inherently harmful. But they are not sufficient as design goals if clinical outcomes are what the product is for. The 92-trial meta-analysis makes that case clearly enough.
What the evidence supports is designing for trust. Transparency about what a product can and cannot do. Personalisation that genuinely reflects the person using it, not just their usage patterns. Honest pathways to human care when the product reaches its limits. Progress measures that are clinically meaningful, not just a proxy for time spent.
This takes longer to build. It requires clinical input at the design stage, real willingness to measure outcomes that surface slowly, and honesty with users that does not always serve retention. It also, the evidence suggests, is what actually works.
The people using these tools came looking for something real. The research is starting to show what that requires.
🪺