A bad answer almost always means missing data — a sync that hasn't run, or a question your sources don't cover. Both are fixable in minutes.
Who this is for: merchants staring at a specific bad answer.
Why — not hallucination
The assistant answers only from your store's own data: products, policies, pages, blog posts, and your Q&A entries — no web search, and it refuses rather than invents. A bad answer means stale data or an answer missing from your sources.
Check freshness
Open Knowledge → Sources.
Check the last sync time — sources re-sync daily, but a product added an hour ago may not be in yet.
Run a manual sync and re-test.
Cover what sync can't
Some answers live nowhere in your catalog or policies — sizing advice, gift-wrapping, "do you ship to X". Add a Q&A entry in Knowledge → Entries. Entries are used verbatim — write the answer as you'd say it to a customer.
Test in the Playground
The Playground in Knowledge mirrors the live assistant — same data, same behavior, burns nothing. Ask it the exact question that failed.
Let Bolt flag gaps
Bolt logs questions it couldn't answer, with a one-tap way to publish an answer as a FAQ — rolling out now. Don't see it yet? The Q&A route above does the same job.
Limits
It will keep refusing questions your data genuinely doesn't answer — that's the design. The fix is adding the answer, not loosening the assistant.
Still stuck?
Message us the exact question asked and the answer you expected — that pair is everything we need.