Review chat conversations

Chat conversations stores every exchange with full context. It is where you see the questions your customers actually ask, in their own words.

Finding a conversation

The search box matches message text, bot responses, and session IDs. Combine it with the filters:

Use Columns to show or hide fields, Apply to run the filters, and Clear to reset everything.

Entries of type internal are system activity, not visitor chats. They are not billed.

Reading a single conversation

Click View details to open the full record:

• User message and LLM response — the exchange itself
• Product search — which products were considered and shown
• Model / Tokens — which model answered and how much it cost
• Visitor info — device, language, and the page the visitor was on
• Response payload — the raw response, useful when something failed

The page URL is worth reading. The same question from a category page and from a product page is a different question in context.

Marking what matters

• Star a conversation you want to come back to
• Add a Note explaining what went wrong — missing product data, wrong price, unclear policy, tone
• Notes are shared with everyone who has access, so write them for a colleague

Conversations started from a dashboard preview carry an [admin preview] note automatically, so your own testing is easy to exclude.

A weekly review routine

1. Filter on Type: error and check nothing systematic is failing
2. Search for questions that produced no product suggestions
3. Read ten conversations end to end, chosen at random rather than cherry-picked
4. Note the most common underlying cause
5. Fix that cause at the source: a product description, a category, a canned response, a field in Site info
6. Star three conversations as your before-and-after benchmark

Housekeeping

Select conversations and use Delete Selected to remove them. Select All applies to the current page of results.

Deletion is permanent. Export or note what you need first, and remember that deleting conversations does not reduce the message count already billed for the period.

Privacy

Visitors sometimes type personal details into a chat. Treat this screen as personal data:

• Give reporting-only colleagues the Viewer role rather than full access
• Delete conversations you no longer need
• Do not paste transcripts into external tools

See Roles and permissions.

Related

• Analytics and optimization
• Human handoff