Most teams tune their chatbot by guessing. You don't have to. Two lists in Insights → Analytics tell you exactly what to fix, taken from real conversations. Working them for ten minutes a week is the highest-return thing you can do with this product.
Answer gaps — questions it couldn't answer
Analytics → Answer gaps lists the questions your assistant was asked and couldn't answer. Each one is a visitor who left without what they came for.
For each gap you can:
- Draft article — the AI writes a first draft of a Help Center article answering it. Review, edit, publish. Publishing deflects the next visitor who asks and teaches the chatbot at the same time.
- Dismiss — for questions you genuinely don't want to answer.
The goal is an empty list: "🎉 No open gaps — your assistant answered everything it was asked."
The best gaps also make the best test cases: on the AI Trust page you can seed a golden question straight from a recent gap, so once you've written the answer, every future simulation verifies it stays right.
AI corrections — answers your operators improved
Analytics → AI corrections to review shows replies where an operator edited the AI's suggestion before sending it. You see them side by side:
- What the visitor asked
- What the AI suggested
- What the operator sent
Your operator, in the moment, wrote a better answer than the assistant had. Click Turn into a Note and that better answer goes into your knowledge base — so the assistant gives it itself next time, and nobody has to correct it again.
This is the whole loop: the AI drafts, a human improves it, the improvement becomes knowledge, the AI drafts better. It only works if somebody reviews the queue.
What else is in Analytics?
- AI deflection rate — how many conversations were handled without a human. The number that tells you whether this is paying for itself.
- Avg. rating and customer satisfaction from visitor ratings.
- Conversation insights (paid) — topic, sentiment, and CX score for every chat, with your top topics. Your top topics are your Help Center backlog, already prioritised.
- Flagged for risk — conversations where a visitor may have been in distress, if crisis safety is on. Follow these up with a person.
- Conversations by channel.
A ten-minute weekly routine
- Clear Answer gaps — draft, review, publish.
- Clear AI corrections — turn the good ones into Notes.
- Glance at top topics — write an article for anything trending that you don't cover.
- Check deflection rate over the last 30 days. It should be climbing.
- After a week with lots of knowledge edits, run a simulation to confirm nothing you relied on got worse.