Real-World Engagement With a Generative AI Conversational Agent for Mental Health Support: A Retrospective Descriptive Study

At a glance

Key findings

• 69.4% of users returned after their first session, and 92.6% of all sessions were followed by another — high retention for a category where most digital tools lose users after a few uses.
• Users completed an average of 11.8 sessions, with a typical session lasting about 16 minutes.
• 62.4% of sessions took place outside traditional 9-to-5 business hours, most often in the evening — reaching people when conventional care is hard to access.
• Average session satisfaction was 4.5 out of 5 and did not vary by time of day or day of week.
• Among users who described their sessions, the most common tags were “Insightful” (67.1%), “Good advice” (61.7%), and “Felt seen” (57.1%).

What the study examined

The researchers analyzed how people actually used a generative-AI conversational agent in the real world — not in a controlled trial. Using de-identified backend logs, they characterized who engaged with the agent, when and how often people returned, and how satisfied users felt after each session.

The analytic sample covered 5,082 paid U.S. subscribers who completed at least one session between October 2024 and March 2026, adding up to 59,602 sessions in total.

Who engaged

Among users who completed the optional onboarding items, most identified as male (78.8%) and described moderate-to-high distress on a single app-native item. The most commonly reported stressors were relationships, work, and finances — the same concerns most often cited by U.S. adults.

Because the onboarding distress item is a single, unvalidated question, these responses describe how users characterized themselves at sign-up and should not be read as clinical indicators of distress.

How people engaged

Users completed an average of 11.8 sessions, with a typical session lasting about 16 minutes. Most engagement happened outside conventional care hours: 62.4% of sessions fell outside 9-to-5 business hours, most often in the evening.

Retention was notably high for the category. 69.4% of users returned after their first session, and 92.6% of all sessions were followed by another — a pattern that stands out against the steep drop-off many digital mental health tools see after the first few uses.

Satisfaction and experience

Average session satisfaction was 4.5 out of 5 and did not differ by time of day or day of week. Among users who tagged their sessions, the most frequently chosen descriptors were “Insightful,” “Good advice,” and “Felt seen.”

Higher satisfaction was associated with a greater likelihood of returning for another session, though the researchers frame this as hypothesis-generating rather than confirmatory.

Why it matters

The study contributes large-scale, objective behavioral evidence that a generative-AI conversational agent can sustain real-world engagement — including among men, a group that has historically underused traditional mental health services, and at hours when conventional care is difficult to reach.

The authors are careful to note that engagement is not the same as clinical benefit. Whether these usage patterns translate into meaningful outcomes is a question for future, prospective research.

Limitations

The authors describe these findings as preliminary and hypothesis-generating. A few constraints are worth keeping in mind when reading them:

• Satisfaction and onboarding items were single, app-native questions without established reliability or validity, and were optional — so ratings may skew toward more satisfied users.
• The sample was limited to paying subscribers, who may be more motivated than free or trial users.
• The sample was predominantly male, which limits how far the findings generalize.
• As a descriptive study, it measured engagement — not clinical outcomes — and cannot establish that the agent improved mental health.

Citation & declarations

More research