Your website isn't just for humans anymore
How to build a chatbot that keeps itself up to date, can't leak client information, and won't answer beyond what you've published.
A business asks for a chatbot. Fair enough: a chat assistant on the website that answers questions at 11pm, captures the enquiry, and points people to the right service is one of the most visible, most requested pieces of AI there is. We build them.
But the chatbot is the easy part. The hard part is a question most businesses have never had to answer before: what does the bot actually know?
Every chatbot has a knowledge problem
An assistant that answers customer questions has to get its answers from somewhere. That “somewhere” has three requirements that pull against each other.
It has to be right. A bot that improvises is worse than no bot. It will invent services you do not offer, quote policies you do not have, and make promises on your behalf that a court may consider binding.
It has to stay current. Knowledge bases rot. The document someone exported for the bot eighteen months ago still says you offer a service you discontinued last winter. Nobody remembers the bot needs updating when the business changes, because the bot’s knowledge lives somewhere nobody looks at.
It must not leak. This is the one that keeps professional practices up at night, and rightly so. Feed a bot your internal documents and you have to be certain, not hopeful, that nothing commercially sensitive and no client information is in the pile. One pricing spreadsheet in the wrong folder, one client name in a case note, and your friendly website assistant is a disclosure incident.
Most chatbot projects try to solve these three problems with process: curation checklists, review cycles, redaction passes. Process depending on people never forgetting is not a control. There is a simpler answer.
The firewall you already have
Your public website is the one body of knowledge your business has already cleared for release. Every page on it has passed the only test that matters: someone decided the whole world could read it.
No client information, because you would never publish that. No commercially sensitive detail, because you already chose what to say publicly and what to keep back. No stale offers you forgot about, because the website is the thing you do keep current, it is your shopfront.
So the trick is this: ground the bot in the public website and nothing else. The knowledge boundary and the privacy boundary become the same line. Publishing is the curation step. If it is on the site, the bot can say it. If it is not, the bot says so honestly and offers a conversation with a person.
This is not a workaround. It is the correct architecture, and it has a second effect that most businesses have not noticed yet.
The same pages feed AI search
When someone asks ChatGPT, Google’s AI results, or Perplexity “who does X in Perth”, those systems answer the same way the chatbot does: they read websites and synthesise. The businesses that get named and cited are the ones whose sites actually answer questions, in plain language, with clear structure the machines can lift an answer from.
That is a different job than ranking for a keyword. Classic SEO earned a click. AI search earns a citation, and the pages that win are the ones written the way a good answer is written: a clear claim, plain evidence, honest scope, structured data underneath. The marketing name for this is AEO or GEO. The plain description is that your website is now read by more machines than people, and most websites were written for neither.
So one set of published pages now does three jobs at once. It persuades the human who visits. It feeds the AI search engines deciding whether to mention you. And it is the entire brain of your own chat assistant.
We run our own site this way
The assistant in the corner of this page answers only from what Perth AI Consulting has published: these articles, the service pages, the free tools. Every answer carries links to the pages it drew from, often with a quoted line, and the quote is checked against the real page text before it is shown, so the bot cannot put words on our site that are not there. Ask it something we have not published and it says so and points you to a conversation.
Running it this way surfaced something useful within days: visitors asked about things we do but had never written down. The bot, correctly, could not find them. A chatbot grounded in your website is also an audit of your website; every question it cannot answer is a page you have not written.
Which is why an AI consultancy builds websites
We are not a web design studio, and if what you need is a brochure site, a web designer is the right call and we will happily work beside yours. But when a client comes to us for a chatbot, or asks why AI search never mentions them, the honest engineering answer is often that the website itself is the work. It is the knowledge base, the firewall, and the AI search surface, and it was built before any of those jobs existed.
So we build and rebuild websites optimised for AI: structured so machines can read them, written so an answer can be lifted cleanly, complete enough that the chatbot grounded in them is worth talking to, and firewalled by design because nothing goes into the bot that was not already public.
Some knowledge can never be public: client history, internal procedures, the way your best people actually do the work. That belongs in a private business knowledge base, a separate build with the opposite boundary, and never in anything a website assistant can reach.
If your website was built to be a brochure and you now need it to be a brain, that is exactly the kind of problem we start with a conversation about.