How to Train an AI Chatbot on Your Website (and What It Should Never Answer)

Most website chatbots that give bad answers aren't using a bad AI model. They have bad content. The bot was pointed at a homepage full of slogans, never told the refund policy, and left to fill the gaps on its own.

"Training" a support chatbot today doesn't mean training a model. It means choosing what the bot reads and checking what it says. Here's the checklist I use when setting one up.

1. Start with your website, then trim

Point the chatbot at your site and let it read your pages. Most tools, including ChatBento, follow your links and sitemap to find them.

Then look at what it picked up and remove what doesn't help:

• Keep: pricing, product and service pages, FAQs, shipping and returns, contact and booking pages, docs, policies.
• Remove: old blog posts with outdated prices, job listings, press releases, legal boilerplate that isn't a policy, and tag or archive pages that only repeat other content.

Outdated pages are the most common source of wrong answers. If your 2023 post says the plan is $19 and the pricing page says $29, the bot may quote either.

2. Write the answers your site doesn't have

Every business has questions customers ask that the website never answers. Get the real list:

• Search your inbox for the same question asked three or more times.
• Ask whoever answers the phone or DMs.
• Look at the search terms people type on your site, if you track them.

Write each one as a short Q&A in plain language:

Q: Can I change my delivery date after ordering?

A: Yes, up to 48 hours before the delivery date. Reply to your order confirmation email with the new date.

Short, specific answers beat long ones. Ten to thirty Q&As covers most small businesses.

3. Include the awkward answers

These are the ones teams leave out because they're uncomfortable, and they're the ones customers most want answered quickly:

• What happens if I want a refund after the deadline?
• Do you work with customers in my country?
• What don't you do?
• How long does it really take to hear back from a human?

A clear "No, we don't ship to Canada yet" is a better experience than a vague answer that leads to an angry email two weeks later.

4. Decide what it should never answer

Write these down before launch:

• Anything outside your business. General knowledge, the weather, homework. A support bot that cheerfully writes poems is a cost, not a feature.
• Account-specific details it can't verify. Order status, invoices or account changes should go to a human or a proper account page unless the bot is actually connected to that data.
• Legal, medical or financial advice, if your business is anywhere near those areas. Answer the general question, then hand off.
• Prices or promises that aren't in your content. If it isn't written down, the bot shouldn't invent a discount.

A good chatbot says "I don't have that information, but I can pass your question to the team" instead of guessing.

5. Set up the handoff before you need it

Every chatbot will hit questions it can't answer. Decide in advance what happens then:

• Collect the visitor's name and email so someone can follow up.
• Tell them when they'll hear back, and make it true.
• Send the conversation to an inbox someone actually checks.

A handoff that leads nowhere is worse than no chatbot.

6. Test with real questions, not demo questions

Before launch, ask the bot:

1. Your ten most common questions, worded the way customers write them, typos included
2. Three questions that need your Q&As, not just the website
3. Three questions it should refuse or hand off
4. One question about a competitor ("Is this better than X?")

Read every answer as if you were the customer. When something is wrong, fix the source content and ask again. Don't try to patch it with clever instructions.

7. Read the transcripts every week

For the first month, read the conversations once a week. You'll find:

• Questions you never thought to answer. Add them as Q&As.
• Pages that confuse the bot. Rewrite or remove them.
• Places where visitors asked for a human. Check whether the handoff worked.

This is where a chatbot goes from okay to genuinely useful, and it's also free market research: you're reading exactly what your customers want to know.

The short checklist

• Site imported, outdated and duplicate pages removed
• 10–30 Q&As for questions the site doesn't answer
• Awkward answers included (refunds, limits, what you don't do)
• Off-limits topics decided
• Handoff goes to an inbox someone checks
• Tested with real customer wording
• Weekly transcript review on the calendar