On this page
"Train it on my website" sounds like a machine-learning project. It is not. In practice you point an assistant at pages you already wrote, tell it how to behave, and then test it like you would test a new hire on their first morning. Ten minutes, three steps — and the last step is the one most people skip.
Minutes 0–3: give it your pages
Add your site as a knowledge source and set a page limit. A sane crawler respects your robots rules, stays on the one origin, and indexes the number of pages you allow — the default of 20 is usually right, because the honest goal is the pages a customer asks about: services, prices, policies, delivery, the FAQ, contact.
Whatever your site does not say in words, paste in as text: the three questions your inbox gets every week, the exceptions to your refund rule, the thing everybody phones to ask. That curated text is usually what turns a mediocre assistant into a useful one (knowledge guide).
| Add as a source | Paste as text |
|---|---|
| Pages that already exist and stay current | Answers that live in your head or your inbox |
| Services, pricing, policies, FAQ | Exceptions, edge cases, "it depends" rules |
| Anything a customer might link to | Anything you would say on the phone |
Minutes 3–6: the system prompt
This is the assistant's job description, and you can fill it in once. Replace the bracketed parts and keep the last three rules exactly as they are:
You are the assistant for [Business name], [what you do, for whom, where].
Speak [warm and plain / brief and formal]; short answers, no jargon, [language(s)].
You may help with: [services, prices, opening hours, booking, delivery, returns].
You must not: quote a price or a policy that is not in the sources, promise a date
you cannot verify, or give [legal / medical / financial] advice.
Hand off to a person when: the customer asks for one, you are unsure twice in a
row, the tone turns angry, or the topic is [billing / complaints / [CHECK]].
Hand-off message: "[I'll bring in a colleague — can you leave your email?]"
Always: answer only from the sources, cite the page you used, and say plainly when
you do not know. Never invent a fact to be helpful.Those last three lines are the whole safety story. An assistant that answers only from your content, shows its sources, and admits uncertainty is one you can put in front of customers today.
Minutes 6–10: the three test questions
- Something only your site can answer — "how much is [service] and how long does it take?" The answer must be right and cite the page.
- Something your site does not cover — "do you take [thing you have never sold]?" It must say it is not sure and offer a person. If it improvises, fix the prompt before anything else.
- Something you just changed — edit a price, reindex, ask again. The answer should follow the page.
Fail any of the three and the fix is almost always content, not the model.
Ask your assistant to prepare it
Read https://[your-site] and list the 20 pages a customer support assistant should
know, most useful first, with one line each on what it answers. Then list the 5
questions customers probably ask that the site does NOT answer, so I can write
those myself. Mark guesses with [CHECK].Take that list into the setup screen. It is the fastest honest shortcut in this whole process.
What makes the answers good
Content quality, almost entirely. Pages written in plain sentences beat pages written in marketing adjectives; one page per topic beats a mega-page. The same tidying helps assistants that arrive from a chat app rather than living on your site: a curated llms.txt tells them what matters (llmstxt.org) — while being clear-eyed that Google said in June 2026 it "won't help or hurt" Search rankings (Search Engine Land).
Once answers are solid, let the assistant do things too — your booking or contact form can become an action it completes on the page with the visitor confirming (WebMCP, page tools). Start with the preview on your own address, and pricing starts where a small site starts.
Questions
Is this the same thing as fine-tuning a model?
No, and you do not want fine-tuning. Your pages are read and looked up at the moment a customer asks, so an answer follows your site within minutes of a change — no retraining, no waiting, and every answer can cite the page it came from.
How many pages should it read?
Fewer than you think. Twenty well-chosen pages — services, prices, policies, the FAQ, contact — beat two hundred that include every blog post you ever wrote. Add more sources later, aimed at specific sections.
What happens when a customer asks something my site does not cover?
A good assistant says it is not sure and offers a person, instead of inventing an answer. That behaviour is set by the system prompt and by turning hand-off on — it is the single most important thing to test.
Do I have to update it when my prices change?
You update the page, as you would anyway, and reindex the source. The assistant reads what your site now says; nothing is frozen inside a model.
Will this help me rank in Google?
Treat it as a conversion improvement, not a ranking one. Google stated in June 2026 that an llms.txt file neither helps nor hurts Search rankings, and no assistant on your page changes your position either — it changes what the visitor who arrives actually does.