The assistant is only as good as the articles behind it, and the fastest way to find out what is missing is to read what people actually asked. Every conversation is recorded and kept for 90 days.
Open a project and go to Site widget → Conversations.
The list
Each row is one conversation:
Conversation — the title, generated from the opening question.
Visitor — who was chatting, if your site told the widget. Anonymous otherwise. See Identify who is chatting.
Questions — how many things the visitor asked.
Signals — small icons summarizing what happened: how many answers cited an article, how many turns ran a tool, and a warning when the assistant answered without citing anything at all.
Last activity — when the conversation was last touched.
Rows are newest first, 25 to a page.
Finding the useful ones
Search looks inside the messages, not just the titles. Searching refund finds every conversation where a visitor or the assistant said "refund", even if the title is about something else. It also matches the visitor's name and email, so searching alice@example.com finds every chat that person had — provided your site identifies them.
No article cited is the filter worth your time. It shows conversations in which no answer cited an article — often questions your documentation does not answer. That is your writing queue.
Greetings and small talk still appear in that filter, since they cite nothing. The Content Gaps card and the Unanswered this month count on the Overview tab leave them out, because only turns that searched are scored.
Reading a conversation
Click a row to open the transcript. It shows:
The visitor's questions, with the page they were on when they asked. That context often explains an otherwise baffling question.
Which tools ran, and with what arguments — help center searches and any tools you registered yourself.
The assistant's answers, with cited articles as links straight into the editor. An article that has since been deleted shows as "deleted article" rather than a dead link.
The transcript lists every article an answer linked to, which can be more than the visitor saw as source chips in the widget — a listing answer links many articles without having read any of them. See How the assistant picks its sources.
Turning gaps into articles
The loop is short:
Filter by No article cited.
Read a few. Look for the same question in different words.
Write or extend the article that should have covered it, and publish it.
The assistant picks it up immediately — no reindexing step.
The Content Gaps card on the Overview tab is a related signal in aggregate: the most recent questions whose help center search found no article at all.
What is stored, and what is not
Stored with the conversation:
the visitor's messages and the assistant's replies
the names and arguments of any tools that ran
which articles were cited
the URL and title of the page each question was asked from
the visitor's id, email and name, if your site identified them
Not stored:
the page text, description or structured data the assistant read
anything you passed as
appContextthe results your own tools returned
Conversations are deleted automatically after 90 days. Any project member can read them; deleting one requires an owner or admin.
Because tool arguments are stored, do not pass secrets through them — see Let the assistant use your own data with tools.
Privacy
The chat never asks anyone who they are. A conversation is tied to a browser session, and it stays anonymous unless your own site tells the widget who is signed in with identify() or the visitor prop. On your own site that is opt-in and off until you write the code for it; the hosted portal's widget passes a signed-in visitor's id, email and name automatically. The Feedback tab is the one place the widget asks for anything: an optional email, filled in for the visitor when your site has identified them.
If you do identify visitors, the name and email you pass are stored with the conversation for the same 90 days as the rest of it. That is personal data on your users, so cover it in your privacy policy. Passing only an opaque id from your own system stores nothing readable: the transcript shows the id, but the list shows Anonymous and search does not match ids.
Either way, people type personal things into chat boxes, and the page URL is stored. If your site has pages that show personal data, turn page reading off there with data-page-context="false", and mention the 90-day retention in your privacy policy.