The Analytics page (/app/analytics) is your one-stop view
on whether the agent is actually working. Conversation volume,
deflection rate, lead conversion, and the all-important
knowledge gap report.
| Metric | What it means |
|---|---|
| Conversations | Distinct chat threads in the period. Counts each visitor's first init, not every message. |
| Messages | Total visitor messages. Average per conversation tells you engagement depth. |
| Deflection rate | Conversations the AI handled end-to-end (no human takeover, no low-confidence flag) divided by total. Higher is better. |
| Leads captured | Visitors who completed the lead form. Conversion shown as % of conversations. |
| Average response time | Median time from visitor message to first token streamed back. Should sit well under the 1s p95 hot-path budget. |
Whenever a turn comes back with low_confidence=true, a
DetectGapJob is dispatched. The job clusters similar gaps
using semantic similarity on the visitor's question, then opens a row in
the Gaps table. Each row shows:
Closing the loop: read the gap → add a knowledge source that answers it → next visitor with that question gets a confident answer.
The page defaults to all agents in the workspace; the agent picker at the top filters to one. Useful when you have one agent per surface (marketing, help center) and want to compare deflection rates.
Pick a window: 24h, 7d, 30d, or a custom range. Numbers update without a full reload — Inertia partial visit. For longer trends, the chart at the top is a 30-day rolling view.
The CSV export pulls everything in the current filter — conversations, messages, leads, gaps. One row per record. Columns are stable so you can automate downstream reporting.
A small table shows which knowledge sources were cited most often and which never get cited. Sources with zero citations in 30 days are candidates for deletion or reindexing.
If you have an active A/B experiment on behavior rules (see Behavior rules & triggers), its conversion delta appears as a callout at the top of the page until you call it (declare a winner). The page-level chart respects the experiment split too — you can see whether the variant lifted overall deflection.