User Friction
Where people are getting stuck, repeating themselves, or giving up mid-conversation.
The User Friction card tracks specific roadblock patterns detected inside conversations — not just a general "friction vs. smooth" split, but a breakdown of what kind of friction is happening.
What you see
A donut chart with four friction types, plus an overall friction rate (the percentage of analyzed sessions where at least one friction signal was detected):
| Signal | What it means |
|---|---|
| Clarification Loop | The user and companion went back and forth multiple times trying to pin down what the user actually needed. |
| Repeated Question | The user asked essentially the same question more than once — usually because the first answer didn't land. |
| User Correction | The user had to correct the companion's understanding or a piece of information it gave. |
| Task Abandoned | The user appeared to give up partway through what they were trying to do. |
How to use this
- A high Clarification Loop rate often points to a companion that isn't asking good follow-up questions early, or is misreading intent — check your prompt and instructions.
- A high Repeated Question rate usually means answers aren't complete or clear enough the first time — cross-check Response Quality's Completeness and Clarity scores.
- User Correction spikes are a strong signal of factual accuracy problems — look at Correctness in Response Quality and consider whether your knowledge base needs updating.
- Task Abandoned overlaps conceptually with the abandoned resolution status on Resolution Rate, but is a distinct, model-detected signal within otherwise-engaged conversations, not the same underlying count.
Last updated on