"Training" is the wrong word, and the word is causing the confusion. When you point a modern AI agent at your website, you are not retraining a language model. You are giving it a library and telling it to answer only from that library.
Which means the quality of your answers is not really a question about AI. It is a question about your library — and most small business websites have a messy one.
What is actually happening
The agent crawls the pages you allow, splits them into passages, and stores them in a way that lets it find the relevant few when somebody asks something. At question time it retrieves those passages and writes an answer grounded in them.
Three consequences follow, and they explain almost every complaint people have:
- It can only know what you gave it. If the delivery policy lives in your head, the agent does not have it.
- It cannot tell old from current. A 2024 page and today's page look equally authoritative to a retriever. Dates in your copy help; deleting the old page helps more.
- It will resolve contradictions by picking one. Not by flagging them. If two pages disagree about your price, you have a coin flip on every pricing question.
What to feed it
In rough order of how much they improve answers:
- Your FAQ, if you have one. It is already written as questions and answers, which is exactly the shape retrieval works best on.
- Pricing and packages, with the real current numbers.
- Service or product pages, including what you do not do.
- Policies — delivery, returns, cancellation, guarantees. These drive a huge share of pre-purchase questions.
- Practical details: areas covered, opening hours, lead times, what happens after someone books.
That fifth group is the one people forget and the one customers ask about constantly. "Do you cover my area" and "how soon can you start" are not marketing copy, so they often appear nowhere on the site at all.
What to keep out
This matters more than what you put in, and nobody mentions it.
- Superseded pages that are still live. Old pricing, discontinued services, last year's campaign landing page. Delete, redirect, or exclude them.
- Internal documents. Anything with your margins, supplier costs or staff notes. If it is in the library, the agent can say it.
- Blog posts that argue a position. A post about why you moved away from an approach can easily be retrieved as a description of what you currently offer.
- Anything personal. Customer names, case details, testimonials with private information. Treat the library as public, because functionally it is.
The audit worth doing once
An hour, and it fixes most of what would otherwise go wrong:
- List every live page. Your sitemap is the fastest route.
- Mark each one current, outdated or internal. Be honest about the middle column.
- Deal with the outdated ones — delete, update, or exclude from the crawl.
- Find the contradictions. Search your own site for your price and see how many different numbers come back. This step surprises people.
- Write down what is missing. The five questions customers always ask that your site never answers. Add them to the FAQ. This is the single highest-value hour in the process.
Structure helps more than volume
More content does not mean better answers. Clearer content does.
One fact, one place. Your price should live on the pricing page and be referenced elsewhere, not restated in six places that drift apart over two years.
Answer in the first sentence. "We deliver across Sindh and Punjab, usually within three working days" retrieves cleanly. Three paragraphs of preamble before the same fact does not.
Use real headings. Headings bound the passages a retriever pulls. A wall of text gives it nothing to cut on.
Say the quiet parts. Minimum order, what is not included, who you are not for. These prevent the worst outcome, which is an agent cheerfully promising something you do not do.
Test it before anyone else does
Write twenty questions you already know the answers to, including the awkward ones — your cheapest option, your slowest turnaround, something you flatly do not offer. Ask all twenty. Score each as correct, vague or wrong.
Anything wrong is a content problem with an address: find the page it came from and fix that page. Anything vague usually means the fact is not written down anywhere. Then ask a question entirely outside your business and see whether it admits it does not know — which is one of the five things worth deciding before you install anything.
Keeping it current
A library goes stale. Re-crawl on a schedule so new pages get picked up, and re-run your twenty questions whenever you change prices or drop a service. Put it in the calendar quarterly; it takes ten minutes and it is the difference between an agent that is right and one that used to be.
the short version
- You are not training a model. You are giving it a library and restricting it to that library.
- What you exclude matters as much as what you include — especially old pages still live.
- Two pages that disagree become a coin flip on every answer.
- One fact in one place, answered in the first sentence, under a real heading.
- Test twenty questions you know the answers to before a customer asks the first one.