A Website Chatbot That Helps: What It Should Do, and What It Should Never Do

Every few months a client asks us to "add a chatbot". Half the time what they have seen is a competitor's pop-up that opens itself, says hello in three languages and cannot answer the one question the visitor had. The other half have a real problem: the same twelve questions arrive by email every day, enquiries come in with half the details missing, and nobody is around at nine in the evening when people actually browse.

The second group is who an assistant is for. Our AI chatbot and website assistant service grew out of building one for our own site, and the rules below are the ones we wish we had started with.

Rule 1: Answer From Your Content, Not From the Internet

A general-purpose model will happily describe a service you do not offer, in a tone you would never use, with a price it made up. The assistant on a business website must be grounded: it answers from your pages, your FAQs, your price list and nothing else. When the answer is not in that material it says so and offers a person, which is what a good receptionist does.

This is the difference between an assistant and a toy. The support assistant we built for a software company cites the documentation page it used for every answer, so staff can see at a glance when the documentation itself is the problem.

Rule 2: Qualify the Enquiry, Then Get Out of the Way

The most valuable thing an assistant does is not answering questions. It is collecting a usable enquiry: what the visitor needs, which platform they are on, when they first noticed the fault, and how to reach them. A conversation that ends with a complete lead in your inbox has done its job, even if the visitor never got a "clever" answer.

That means the assistant should know when to stop talking. A visitor who has given their email and described the problem does not need six more questions. Ours asks for a website address first, then one or two diagnostic questions, then offers a quote band and a contact option. Nothing more.

Rule 3: Never Invent Prices, Dates or Promises

If the assistant quotes, it must quote from a table you control, with a range and a caveat. If it books, it books into a real calendar. If it cannot do either, it should say "a member of the team will confirm" rather than guess. One confidently wrong price costs more trust than a hundred honest "I will check".

The same caution applies to anything that touches money or data. If you want an assistant that can look up an order or raise a ticket, that is an integration project with authentication and audit trails, not a widget.

Rule 4: A Person Must Always Be One Click Away

"Talk to a person" belongs in the assistant's first message and in every reply after it. During office hours the hand-over should be live, with the agent seeing the transcript so the visitor does not repeat themselves. Outside hours it should file the conversation and tell the visitor when to expect a reply. Nobody should ever be trapped in a loop.

Rule 5: Keep Bots and Spammers Out

An assistant that sends email is a target. Scripts will open it a thousand times an hour and fill your inbox, or probe it for a way to post content. A simple human check before the first message, rate limits per address, and redaction of anything that looks like a password keep it safe. If your site has already been abused this way, our cleanup and security service will stop the flood and harden the forms.

What "Working" Looks Like

• Enquiries with complete details: the share of leads that arrive with a website address, a description and contact details should rise sharply.
• Repeated questions by email: should fall, because the assistant answers them from your content.
• Hand-over rate: if more than a third of conversations end in "talk to a person", the content the assistant answers from needs work.
• Out-of-hours leads: the clearest win. Visitors who would have left now leave their details.

Measuring these needs the assistant to log conversations somewhere you can read them, which also lets you see the questions it could not answer and fix the pages behind them. That loop, more than the model, is what makes an assistant improve.

Where It Fits With the Rest of the Site

An assistant does not fix a slow site or a confusing one. Visitors who cannot find the service page will not find the chat button either. Sort the basics first, with a technical SEO review and clear service pages, then add the assistant where the remaining questions cluster. For a broader look at where language models earn their keep on a business site, see practical AI use cases for websites.

If you are weighing an assistant up, send us the questions your inbox gets most. Those twelve questions are the specification, and we can usually show you an answer to each of them from your existing content within a week.