User Engagement, Speaking Verbosity & Retention Analysis
Active Users (left axis) vs. Total Completed Conversations (right axis) since Jan 2026
Comparing mean and median session duration weekly
| Week Start | Active Users | Total Conversations | Mean Duration | Median Duration | User Msg/Conv | AI Msg/Conv | User Words/Conv | AI Words/Conv | User-to-AI Ratio |
|---|
In mid-June, production migrated from gemini-2.5-flash-native-audio-preview to gemini-3.1-flash-lite. The new model demonstrates a 90% spike in natural output verbosity under unchanged instruction guidelines, resulting in the AI dominating conversation speaking share.
Comparing average total words spoken per conversation by User vs. AI
User speaking volume divided by AI speaking volume (higher = more balanced)
Model transition to gemini-3.1-flash-lite in week of June 14
Model transition to gemini-3.1-flash-lite in week of June 7
Percentage of returning users from Week 0 to Week 4
Week 1 Retention (green) & User share (blue) by first session chatlines
Paid Google Ads CPC vs. Organic/Direct traffic retention
| Starting Roleplay | Cohort Size (Users) | Week 1 Retention | Week 2 Retention | Week 4 Retention | Retention Class |
|---|