Without a guardrail, a chatbot on a plumbing website will happily write Python or discuss the news, because the underlying model can. Settings → AI Behavior → AI Persona & Off-Topic Guardrails stops that.
Strictness
- Relaxed — declines only clearly unrelated requests.
- Moderate — the sensible default.
- Strict — answers only what your content covers.
On-topic questions and anything in your knowledge base are never affected. A declined question is answered politely, with a steer back to what the bot can help with.