A lot of teams are bolting a chatbot onto their marketing site this year. Most of them optimize for the wrong thing. They tune for fluency, for how human the replies sound, when the property that actually protects the business is restraint: how often the bot refuses to answer something it cannot back up with real content.
If a visitor asks your site bot about a refund window and the model guesses "30 days" because that is the statistically common answer, you have just created a promise you never made. On a developer tool, it invents an API limit. On a clinic site, a policy. Fluency made it worse, not better. The useful version of a website assistant is narrow:
Notice the model is barely the point. This is a retrieval problem with a strict refusal policy on top. The hard parts are keeping the index fresh when pages change, setting a confidence threshold that fails closed, and not leaking the system prompt.
Standing this up yourself is a real project: crawler, chunker, embeddings, vector store, retrieval, a refusal threshold, a widget, plus the unglamorous maintenance forever. If you would rather skip that, a white-label option I have been testing is Knowster. You point it at a site, it learns the business from the existing content, and it answers visitor questions 24/7 from that content with a snippet to install. It is built for business and website owners, and the default behavior is the grounded one: answer from source, or say it does not know.
Ask the assistant three things: one the site clearly answers, one it answers in an obscure corner, and one it does not answer at all. You want a hit, a deeper hit, and an honest "I don't know." A bot that passes that is safe to put in front of real visitors. A bot that confidently answers the third one is a liability with good grammar.