Review conversations and measure performance

Before you begin

Use the workspace that owns the chatbot. Analytics need recorded activity; an empty period is not evidence that a chatbot failed. Have permission to handle any personal or confidential information in transcripts and exports.

Example from the Chyt.ai dashboard (September 2026).

Review a conversation

1. Open Conversations → All Conversations.
2. Use Search conversations..., All Chatbots, and All Channels to narrow the list. The current channel filter offers Web, Widget, and API.
3. Open a conversation and read the visitor's question alongside the chatbot answer. Inspect Details, source references when present, feedback, and Token Usage.
4. If an answer is wrong, use Correct this answer on the response. Review Question and Wrong Answer, enter a supported Correct Answer, and select Save Correction.
5. Re-test the same question and a paraphrase in the chatbot preview. A saved correction is an input to future answers; verify the result rather than assuming every wording is fixed.
6. Use PDF for an individual transcript. Use Export CSV on the conversation list for exported records. Review any Share settings before creating a share link, because recipients may be able to read the shared conversation.
7. Use the Topics tab to inspect available topic clusters and identify repeated issues.

Read the analytics dashboard

1. Open Analytics. The page is headed Dashboard Overview.
2. Select the chatbot and choose 7 days, 30 days, or 90 days. Keep the selection consistent when comparing results.
3. Review Conversations, Messages, Avg. Response Time, and Satisfaction for overall activity and feedback.
4. Inspect Conversations & Messages, Model Usage Distribution, Token Usage, and Language Distribution to understand workload and model use.
5. Review Knowledge Gaps, Top Intents, Customer Satisfaction, and Guardrail Events where data is available. Choose recurring gaps to improve in the chatbot's sources or instructions.
6. Interpret outcome cards using their displayed definitions. AI containment (observed) means AI replied without a recorded handoff, assignment, or human reply; it does not confirm that the visitor's issue was resolved. Human resolution rate uses handed-off conversations as its denominator.
7. Select Export to download the selected chatbot's analytics as CSV.

What you should see

You can connect a reported problem to the original transcript and use the selected period's metrics to choose improvements. Conversation and token counters describe usage; they are not commercial conversation or token caps. Managed AI capacity is governed by AI credits in Billing.

Troubleshooting

• N/A: the metric may lack an eligible denominator or feedback; it is not automatically zero performance.
• Unavailable or refresh warning: use Retry analytics. Previously loaded sections may be stale.
• Counts differ between views: check chatbot, period, filters, and the metric's denominator.
• Export fails: retry after confirming the selected chatbot and access. Do not assume the file was created until the download completes.
• Correction did not improve an answer: check whether its supporting source is current and test the exact question and a paraphrase again.

Continue learning

Handoff and inbox · Leads and verified contacts · AI credits