How to train an AI chatbot on your website content

Training a website chatbot is less about feeding it every document you can find and more about giving it a clear source of truth. Start with the public pages customers should trust, add exact business facts that deserve extra authority, and test the questions your pages currently answer poorly.

More content is not automatically better. Duplicate and outdated content can make the correct answer harder to identify.

Define the answer territory

List what the assistant is allowed to answer. A practical first version usually includes products or services, business hours, locations, pricing context, policies, basic recommendations, and the next step for a quote or booking.

Also list the boundaries. Medical, legal, financial, safety-critical, or highly personal decisions may need a human even when the website contains general information. Private operating instructions should shape how the assistant behaves; they should not become a second public facts database.

This boundary is part of deciding whether an AI chatbot fits your small business, not an afterthought.

Choose one canonical website source

Import the main website as the canonical public source. Prefer the pages a visitor would reasonably treat as current: service pages, product collections, pricing, contact information, policies, and useful FAQs.

Before import, remove or repair obvious contradictions. Common examples include:

• different opening hours in the footer and contact page;
• an old promotion still visible in a blog post;
• a PDF with prices that no longer match the service page;
• two descriptions of the same cancellation policy;
• location pages with copied details that are not true for every branch.

When the website is refreshed later, a safe workflow should stage the new crawl and replace the canonical source only after the refresh finishes. A partial refresh should not erase the last known good knowledge.

Promote exact facts

Some details deserve a stronger signal than ordinary prose. Store current prices, percentages, dates, hours, identifiers, service areas, and policy limits as verified facts when the product supports it.

For each exact fact, record:

• the value in the format customers need;
• the effective or expiry date when relevant;
• the owner who verified it;
• a public source URL when one exists;
• the wording to use when the fact is conditional.

“Delivery is $12” is weaker than “Local delivery starts at $12; the final fee depends on the postcode.” The second version preserves the exact value without hiding the condition.

Resolve conflicts by authority

A useful authority order for a small business is:

1. owner-verified structured facts;
2. owner-written current knowledge;
3. reviewed corrections from real conversations;
4. the canonical website;
5. other imported URLs or files.

Authority does not excuse a conflict between two equally trusted sources. If two verified records give different prices for the same item and time period, fix the records. Until then, the assistant should ask for confirmation rather than choose the more convenient number.

Write for retrieval, not for robots

Clear source content helps both people and search. Use descriptive headings, keep related facts together, and name the actual service or product. Avoid a wall of interchangeable marketing claims.

A strong knowledge item answers one coherent set of questions:

Wedding cake consultations are 30 minutes. Bring your date, guest count, style references, and allergy requirements. A $50 deposit holds the production slot after the quote is approved.

That block is more useful than splitting the consultation length, preparation list, and deposit rule across three unrelated pages.

Add the questions the website forgot

Your inbox and phone history are valuable editorial sources. Add the repeated questions that never made it onto the site: parking, access, turnaround time, custom requests, what to bring, service boundaries, and what happens after someone asks for a quote.

Write the answer the business is comfortable standing behind. If the answer changes case by case, teach the assistant what to collect and when to hand off rather than inventing a universal rule.

Test retrieval and behavior separately

Knowledge answers “what is true?” Behavior answers “how should the assistant respond?” Test both.

Ask:

• an exact question using the same wording as the source;
• the same question in casual language;
• a question containing two separate needs;
• a question in another supported language;
• a question just outside the documented boundary;
• a question with a wrong assumption.

The assistant should find the right facts, avoid echoing the wrong assumption, and admit when the source is insufficient. The installation guide includes a practical private-test matrix before the widget goes live.

Use real conversations as a maintenance queue

After launch, review conversations for missed answers, disliked replies, repeated follow-up requests, and exact values that visitors phrase unexpectedly. A disliked answer should become a pending suggestion, not trusted knowledge automatically. An owner can edit and approve the correction after checking it.

Review high-impact facts whenever prices, hours, services, policies, or promotions change. Refresh the canonical website source, keep verified records current, and retest the questions that used to fail.

The best knowledge base is not the largest one. It is the one an owner can explain, verify, and keep current.