Feedback Insight
The Feedback page surfaces the explicit satisfaction signals your users leave when they rate an answer. Even a small volume of negative feedback contains high-value information — users only bother to rate when they feel strongly about the result.
Positive feedback tells you what to preserve. Negative feedback tells you what to fix. Both tell you which parts of the agent your users are actually engaging with.
Opening the Feedback Page
In the sidebar, expand Insight and click Feedback.
Once the page opens, orient yourself in this order: first the summary cards at the top, then the trend chart beneath them, then the detailed list panel used for investigation.
Summary Cards
| Card | What it shows |
|---|---|
| Total Feedback | How many answers were explicitly rated in the selected period |
| Positive Feedback | Count and percentage of đź‘Ť ratings |
| Negative Feedback | Count and percentage of đź‘Ž ratings |
Interpreting the Numbers
A healthy agent typically shows a positive rate above 80 %. Benchmarks vary by domain — technical support agents tend to have lower satisfaction than informational Q&A agents because the questions are harder — but the trend matters more than the absolute percentage.
A negative rate that climbs week over week is the single most reliable signal that something has changed: user expectations may have shifted, the knowledge base may have grown stale, or an upstream integration may be failing silently.
Note: Total Feedback only counts conversations where the user actually clicked a rating. Unrated conversations are not represented. When total feedback volume is very low (fewer than 30 ratings in a period), treat the percentages as directional indicators rather than statistically significant measurements.
Low Feedback Volume
If Total Feedback is consistently very low relative to your conversation volume, consider whether users can easily see the rating buttons in the channel they are using. Some channel integrations render the rating UI differently, and some users may not realise they can provide feedback.
Feedback Trend Chart
The grouped bar chart plots positive feedback (green) and negative feedback (red) side by side for each time bucket in the selected period.
Treat this chart as your change-detection layer: it quickly shows whether sentiment is stable, improving, or drifting over time.
Use the Period Selector (top-right) to adjust the time window. The 30-day view is most useful for trend analysis: a single bad day is less meaningful than a trend spanning several weeks.
Reading Trend Patterns
| Pattern | What it suggests |
|---|---|
| Negative rate suddenly spikes on a specific date | A configuration change or knowledge update on that date changed agent behaviour |
| Negative rate gradually increasing over weeks | Knowledge base growing stale, or user expectations evolving beyond the agent's current scope |
| Positive and negative both rise proportionally | More users are rating — the rating mechanism became more visible, not necessarily an accuracy change |
| Positive rate high but volume very low | Only the most engaged (and most satisfied) users are rating; frustrated users may be leaving silently |
| Negative feedback clusters around specific hours | Possible performance degradation or an integration that fails during high-traffic windows |
Click any bar to open the Feedback List panel below the chart.
After you spot a suspicious pattern in the bars, move directly to the list panel for concrete evidence from individual conversations.
Investigating Individual Feedback
The Feedback List panel shows every rated answer within the time bucket you clicked. Each row contains:
- Date / Time of the rating
- User who submitted it
- Message — the user's question that prompted the rating (truncated; hover for full text)
- Reason — optional text the user typed to explain a negative rating
- View Conversation — button that opens the full conversation thread
Use this panel to move from aggregate metrics to case-level diagnosis, where each row becomes an actionable investigation lead.
Use the view toggle (top-right of the panel) to switch between the table layout and a card layout optimised for narrower screens. Click Load More to page through additional results.
Always Follow Up on Negative Feedback
The message snippet and reason text tell you what went wrong. The full conversation tells you why:
- Did the agent hallucinate a fact not present in the knowledge base?
- Did it misunderstand the scope of the question and answer a related but different question?
- Did it give a technically correct answer in an unhelpful format (e.g., a wall of bullet points when a direct sentence would serve)?
- Did it refer the user elsewhere when it had the information to answer directly?
- Did it trigger a function tool that returned unexpected data?
Each of these failure modes has a different fix. You cannot identify the right fix without reading the full conversation.
Common Negative Feedback Patterns
Use the table below as a diagnostic guide when reviewing negative feedback.
| Pattern in reason text | Likely cause | Where to investigate |
|---|---|---|
| "Wrong information" / "Not accurate" | Outdated or missing knowledge | Review the resource / knowledge base content |
| "Didn't answer my question" | Mission scope too narrow, or vague question handling | Revise Mission and Interaction in Behavior |
| "Too long" / "Too much text" | Interaction style not constraining response length | Add response-length guidance to Interaction |
| "Too short" / "Not enough detail" | Interaction style over-constraining length | Loosen length guidance or add instruction to expand on complex topics |
| "Keeps asking me questions" | Over-cautious clarification behaviour | Adjust Interaction to be more decisive |
| "Used the wrong tool" / "Wrong data" | Function tool description misleading the AI | Rewrite Custom API name and description |
| "Not helpful" (no further detail) | Formatting, latency, or channel rendering issue | Check Monitoring logs and channel settings |
| No reason text but consistently negative | User satisfaction issue not related to content | Review Personality and Interaction for tone |
| Negative feedback from the same user repeatedly | Specific user expectation mismatch | Review their conversation threads and consider whether the agent's scope covers their use case |
Acting on Feedback
Feedback is most valuable when it drives a specific, testable change. The process below keeps changes accountable:
- Cluster — Group negative feedback by the reason text or the part of the agent they implicate (knowledge, behavior, function tools).
- Prioritise — Focus on the cluster with the most recurring reasons. A single isolated complaint is less important than the same complaint appearing from multiple users.
- Hypothesise — Form a specific hypothesis: "If I add a response-length instruction to Interaction, the 'too long' cluster should shrink."
- Change — Make the change in Draft mode. Do not publish immediately.
- Validate — Test manually in the preview panel. Confirm no new Monitoring errors appear in the Draft log.
- Publish — Deploy the Draft to Published.
- Measure — Check the Feedback trend over the next 7 to 30 days depending on your traffic volume. A change that works will show in the trend data.
Tip: Export the Feedback page (or take a screenshot) before making a configuration change. Having a baseline makes it easier to demonstrate improvement to stakeholders after the change takes effect.
Related Pages
- Read the Dashboard — track conversation volume and function-tool adoption
- Investigate with Monitoring Logs — trace the technical events behind negative feedback
- Analytics & Logs Overview — monitoring cadence and improvement loop
- Configure Behavior — act on insights by revising Mission, Personality, and Interaction