"Evidence-based" is one of the most used phrases in digital mental health, and one of the least examined. It appears on landing pages, in pitch decks and on procurement forms, usually as a simple yes or no.
Three reviews published this autumn suggest it is better thought of as a shape. Each pooled years of research on a different kind of tool: prescription apps in Germany, gamified apps for young people, and AI chatbots for university students. Read together, they show what sits underneath the phrase, and where it runs out.
The good news is real
Start with what the evidence supports. In 2020, Germany made certain digital mental health apps eligible for prescription. A systematic review and meta-analysis published in Scientific Reports in September pooled 19 randomised controlled trials of these apps, with 4,078 participants.
For depression, the apps reduced symptoms compared with control conditions. The effect was moderate immediately after the intervention (SMD -0.49) and smaller at follow-up (SMD -0.35) (Schmitz et al., 2026).
AI chatbots point the same way. A meta-analysis of conversational agents for university students found that, compared with no support or low-intensity support, they reduced depressive symptoms (g = -0.41, three comparisons) and anxiety symptoms (g = -0.44, four comparisons) (Shao et al., 2026).
For anyone still asked whether digital mental health works at all, these are useful answers. On average, in trials, well-built tools do something.
Where the evidence actually sits
The same German review is more interesting for what it says about distribution. The depression effect after the intervention came from five apps. The follow-up effect came from one. For generalised anxiety disorder, only two studies were available (Schmitz et al., 2026).
The authors are direct about it. The evidence is dominated by a small number of apps, particularly Deprexis, and should not be read as equally representative of every app listed for depression in Germany. They call for real-world evaluation of effectiveness, content, quality and implementation, not only efficacy in trials (Schmitz et al., 2026).
This matters well beyond Germany. A list of approved or reimbursed apps can look like a body of evidence for a whole category. In practice, a few products may be carrying the result for everyone else. Sitting next to a well-tested app on a list says little about the app beside it.
The average, and the range for the next study
A meta-analysis of gamified digital interventions for children and adolescents, published in the Journal of Medical Internet Research on 13 November, gathered 27 studies covering 28,777 young people. Feedback, challenges and progress tracking were the most common game elements. Social features and competition were rare (Baschab et al., 2026).
Only five controlled studies could be pooled. Their combined effect on anxiety and depression symptoms was small and only just significant (g = -0.28, 95% CI -0.54 to -0.02). The authors also report a prediction interval, which estimates the range of effects a new study in a similar setting might find. That range ran from a moderate benefit to a small effect in the wrong direction, and it included zero (Baschab et al., 2026).
The difference between those two numbers matters. A confidence interval describes how precisely the average has been measured. A prediction interval describes how much results vary from one setting to the next. A product team is not the average of five studies. It is the next setting.
Engagement and adherence problems were reported across many of the studies. The authors also noted that results looked more favourable where game elements sat alongside structured therapeutic content, a pattern they could not test formally (Baschab et al., 2026).
Compared with what?
The chatbot review shows a third limit. Against waiting lists and low-intensity controls, conversational agents helped. Against an active comparison, another conversational programme called MYLO, two comparisons found no reliable added benefit for depression or anxiety (Shao et al., 2026).
The authors rate the certainty of the evidence as low. The estimates rest on few comparisons, with attrition and self-reported outcomes among the limitations (Shao et al., 2026).
"Effective" is always a comparison. Better than nothing, better than a leaflet and better than what a person would otherwise use are three different claims. Most of the evidence in this field speaks to the first two.
What this means for the field
None of this argues against digital mental health. The German review found a moderate effect for depression across 19 trials, a result many interventions would welcome. What these reviews show is that "evidence-based" describes a shape, and a few questions reveal it:
- Whose trials are these? Your own product's, or others in your category?
- What was the comparison? Nothing, a waiting list, or a real alternative?
- How long did the effect last, and how many products have follow-up data at all?
- How wide is the range a new study might expect, not only the average?
- Is your condition covered? For generalised anxiety, the German evidence base was two studies.
These are the questions a payer, a clinician or a careful investor will ask. They are easier to answer before someone else asks them.
If you are working through them, nest is a free community for founders, clinicians and researchers building evidence-based mental health tech. Apply to join.
🪺