Conversation metrics
What this screen answers
How much traffic your bot handled, when it handled it, how users talked to it, and which topics it resolved quickly or slowly. Use it for capacity planning and for spotting topics where conversations run long without getting anywhere.
Element by element
Date range
Top right. Every number on this screen is calculated for the selected range. The badge shows the length of the range (for example 24d); the arrows step backward and forward by the same length.
Summary statistics
Six figures across the top, each with a trend indicator comparing it to the immediately preceding period of equal length.
| Statistic | What it counts |
|---|---|
| Conversations | Total conversations in the period. |
| Utterances | Total messages, user and assistant combined. |
| Conversation length | Average duration of a conversation. The trend is shown in seconds rather than as a percentage. |
| User turns | Average number of user messages per conversation. |
| User:assistant verbosity | Average words from the assistant per word from the user. 1:9 means the assistant writes nine words for every one the user writes. |
| Total time | Total hours users spent in conversation. A rough proxy for agent time saved. |
Trend indicators use an up arrow, a down arrow, or an equals sign. An arrow direction is not a judgement: rising conversation length can mean richer handling or worse deflection, and you need the Efficiency chart below to tell which.
Conversations over time
Volume across the selected range. The solid line is the current period; the dotted line is the preceding period of equal length, overlaid for comparison. Hover any point to see both values at that position.
Workload
Conversation volume broken down by time of day or day of week, switchable with the Days / Hours toggle in the component header. Workdays and weekends are plotted as separate series so a weekday peak is not flattened by weekend quiet. Use this to size infrastructure and to schedule releases outside peak hours.
Interactions
How users entered their messages: Text (typed), Button (clicked a prepared option), or Voice (spoken). The Initial / Other toggle switches between the very first input of each conversation and every subsequent input.
A high button share on first interaction with a low share afterwards is normal: users start from a menu and then type. The reverse pattern is worth investigating.
Efficiency
- RT (resolution time) — average seconds to resolve a request within one conversation, per topic, with the change against the previous period.
- FCR (first contact resolution) — the share of conversations on that topic where the request was resolved, with the change against the previous period.
Read the two together. A topic with low RT and low FCR is being closed fast without being solved. A topic with high RT and high FCR is expensive but working. Persistently low FCR on a high-volume topic is the clearest signal of where to invest next.
A large FCR drop between periods, for example from 90% to 35%, usually means one of two things: a change was shipped that broke the flow, or users started asking a new kind of question inside an existing topic.
Abandoned conversations
The share of conversations users left before their request was resolved, by closing the chat or hanging up. Rising abandonment means the dialog is not carrying users to an outcome.
Empty conversations
The share of conversations where the user never sent a single message. The widget opened and the user ignored it. This is a placement, timing, and greeting problem rather than a dialog problem, and it is usually fixed outside the bot.
Common tasks
- Plan for peak traffic
- Assess speed and effectiveness
- Diagnose what's wrong with your bot
Related
- Trends and benchmarks — how the comparison period is calculated
- Topics, actions and objects — how topics in the Efficiency chart are formed
- Conversations — read the individual conversations behind any topic