Don't guess at what your assistant will say — ask it. Knowledge → Live Data has a test console that runs a question against your live knowledge base exactly as a visitor's would run.
How do I test a question?
Go to Knowledge → Live Data, type a visitor question into Test Knowledge Base, and click Test. You get three things back:
- AI suggestion — the reply the assistant would draft.
- Source chunks — the exact pieces of your content the answer was built from. Expand them.
- MCP tools called — any live tools it invoked, if the question needed one.
What does "No suggestion" mean?
It means nothing relevant was found in your knowledge base. That's not a bug — it's the assistant refusing to make something up. It's also the single most useful signal you can get: it tells you exactly what to go write.
Fix it by adding the knowledge: a Note is usually the fastest route, or fill the questionnaire gap, or publish a Help Center article.
The answer is wrong. How do I find out why?
Read the source chunks. The answer is built from them, so the fault is almost always visible there:
- The right content isn't in the chunks → it isn't indexed, it's disabled, or it's phrased so differently from the question that it doesn't match. Add a Note in the visitor's language, not your internal jargon.
- The chunks are right but the reply is off-tone or says something it shouldn't → that's a rule, not knowledge. Use the Preview console on the Chatbot Rules page.
- The chunks contain stale or contradictory content → find the source that produced them and fix or disable it. View ingested content on each source shows what we actually extracted.
What should I test?
The questions you actually get. Pull ten from your inbox, your phone log, or Analytics → Answer gaps — which is a standing list of the questions your assistant couldn't answer, straight from real conversations. See Answer gaps and AI corrections.
Re-test after any significant knowledge change. It takes a minute and it's the difference between an assistant you trust and one you hope about.
And when you want to check everything at once rather than one question at a time, run a simulation — it replays your recent real conversations against the current knowledge base and flags every answer that changed.