Overview
What each figure on an agent's Overview counts, how the quality bands are drawn, and how to read the two lists that tell you what to fix.
The Overview is the first tab of an agent. It answers three questions: is the agent being used, what do people ask, and where does it fall short? It also shows whether the agent is ready and how it is installed.
The period
The page covers the last 30 days unless you choose other dates, and keeps the same period when you switch between Overview and Activity. Days are your own: the period and the bars follow the clock of the computer you are reading on, so a conversation at eleven in the evening is on that evening's bar.
The four figures
| Figure | What it counts |
|---|---|
| Conversations | Conversations that visitors started in the period, through the widget, the iframe or the API. |
| Questions | Every message visitors sent in the period, greetings included. |
| Avg. confidence | The average confidence of all answers that have one. |
| Poorly covered | The share of answers with a confidence below 50. |
Your own tests in the Playground are left out of every figure on this page, and of client reports. A line under the figures says how many conversations that were, with a link to them in Activity.
The Volume chart shows conversations and questions per day. A quiet day is drawn as an empty day, not left out.
The agent itself
Three panels do not change with the period. Knowledge counts the sources that are ready, and of which kind (Files, Links, Q&A, Text), the ones being processed and failed, the storage they use, and when the agent was last trained. Deployment shows the language, the model, the Credits this month its answers used, whether the widget is limited to Allowed domains, and whether the ChatterLab badge is shown. The credits are counted per answer at the model's price, the figure on the Clients screen; files visitors send cost credits on top and are not in it. Recent conversations lists the latest 5 with visitors, your tests left out, each a click from its transcript.
Widget
What the chat on your site did in the period. Opened, Conversations and the three that split them count only the chat the button opened, since only it says what opened it. Leads, Forms declined and Where conversations start count the iframe snippet too, as the Leads page does. The Playground and the API are in none of them. They were first counted on 24 September 2026.
| Figure | What it counts |
|---|---|
| Opened | How often a visitor opened the chat with the button, once for each page it was opened on. It comes from the chat's own page loading, so nothing is sent only to count it, and one address adds at most 20 an hour. |
| Conversations | Conversations started in a chat the button opened. A visitor who opens the chat and asks nothing counts under Opened only. |
| Leads | Contact details left in the chat's form in the period. |
| From the button, From the greeting, From the second nudge | The same conversations, by what the visitor clicked to open the chat, each with its share. The share of the greeting and the second nudge is what they bring in. |
| Forms declined | Conversations in which the visitor answered the contact form with Not now, or chose Skip before their first question. |
Opened is kept per hour. In a time zone that is not a whole number of hours from UTC, such as India's, an open in the half hour around midnight can land on the day before.
Where conversations start lists the 10 pages of your site that the most conversations started on, by the page of their first question: by name when the agent has read the page as a source, otherwise by its path, masked as in Activity.
Answer quality
Every answer falls in one of three bands:
| Band | Confidence |
|---|---|
| Well covered | 70 and up |
| Partly covered | 50 up to 70 |
| Poorly covered | below 50 |
The same limits colour the pills in Activity, so the two screens always agree.
Beside the bands are the visitor ratings: how many answers got a thumbs up and how many a thumbs down. Visitor feedback.
A healthy agent is mostly green with a thin red slice. Some red is normal and even right: people ask a bicycle shop about train tickets, and the agent should not know.
Repeated questions
Questions asked more than once in the period, most frequent first, up to eight. They are matched on wording, not on meaning: "What are your opening hours?" and "When are you open?" count as two different questions. Capitals and punctuation are ignored.
Use the list to see what your site fails to make clear. A question asked twenty times a month deserves a line on your homepage, not just a good answer from the agent.
Worth adding to your sources
Your to-do list: up to eight answers most worth improving. An answer gets on it in one of two ways:
- its confidence was below 50, so your sources did not cover the question
- its confidence was fine, but the second check found the answer Partly backed or Not backed by the passages
They are ordered by how misleading they were: Not backed first, then Partly backed, then the rest from the lowest confidence up. An answer that read well and was not true is worse than an honest "I don't know".
Click a row to open the conversation. Then follow Fixing weak answers.
Greetings
"Hello", "hi", "goedemiddag" and their kin in all five languages are left out of both lists. A greeting finds nothing in your sources, so it would top the weak list every week, and visitors greet more often than they ask any one thing. Only a message that is a greeting and nothing else is left out. "Hello, do you deliver to Belgium?" counts as the question it is.
Greetings are still counted in the four figures and in the bands. If many of your visitors only say hello, that pulls the average confidence down a little without anything being wrong.
Over the API
The same figures are available from GET /agents/{id}/analytics. See the API reference.