What to put in your chatbot's knowledge base, and what to leave out

When an AI assistant gives a wrong answer, the first suspect is usually the model. In our experience it is usually the content. The assistant answers from whatever you connected it to, and most sites connect everything: every page, every PDF, old blog posts, the help centre, sometimes past tickets. Then it answers from the wrong piece.

This note is a practical list: what to give the assistant, how to write it so the search inside the assistant finds it, and what to keep out. If you haven’t read how the lookup works, our note on how an assistant answers explains it in five minutes.

The twelve things customers ask about

We looked at the questions customers ask store and service assistants, and they fall into a short list. Whatever you sell, most of the traffic is here:

Topic	Typical question	Where the answer should live
Delivery times	“How long to Lyon?”	Delivery page, one table per region
Delivery costs	“Is shipping free over €50?”	Same page, thresholds in plain text
Returns window	“Can I return after 30 days?”	Returns page, first paragraph
Return costs	“Who pays for the return?”	Returns page, own sentence
Exchanges	“Can I swap for another size?”	Returns page or FAQ
Payment methods	“Do you take PayPal?”	Payment page or FAQ
Order status	“Where is my order?”	Hand-over rule, not content
Product facts	“Is this sofa fabric washable?”	Product page, specs in text
Stock and availability	“When is it back?”	Product page, if you publish it
Opening hours, locations	“Are you open Sunday?”	Contact page, per location
Prices and fees	“How much is installation?”	Price list, one fee per line
Account and booking	“How do I change my booking?”	FAQ, step by step

If any of these has no clear page on your site, the assistant will either say it doesn’t know or, worse, fill the gap from general knowledge. Write the page first.

How to write content the assistant finds

The search inside an assistant matches the customer’s question to short pieces of your content. A few habits make that match reliable.

One rule, one paragraph. “Returns are free within 30 days of delivery for unused items in original packaging.” If the same paragraph also covers exchanges, damaged items and gift cards, the assistant will mix them.

Say the obvious words. Customers write “refund”, “send back”, “money back”. If your page only says “withdrawal right”, add the everyday words once.

Put numbers in text, not in images. A delivery price table saved as a picture is invisible to most assistants. A table in HTML is usually fine; a sentence is safest.

Name the exceptions next to the rule. “Free delivery over €80, except for furniture over 30 kg, which is €49 anywhere in mainland France.” When the exception lives on another page, the assistant often gives the rule without it.

Date what changes. “Christmas orders: order by 18 December for delivery before the 24th (2026).” Next year you’ll know which line is stale.

What to leave out

Some content does more harm than good in an assistant’s sources.

• Old blog posts and news. “We now deliver in 24 hours” from 2022 is still a page. The assistant can’t tell it’s old. Either exclude the blog or mark old posts clearly.
• Marketing pages that round things off. “Free returns, always” on a banner, when the returns page has conditions. The assistant may pick the banner.
• PDF catalogues and price lists from past years. If you can’t remove them from the site, remove them from the assistant’s sources.
• Terms and conditions as the only source. Legal text is accurate but hard to summarise. Keep it, but make sure there is a plain-language version the assistant finds first.
• Internal notes and macros written for staff (“offer 10% if they complain twice”). They will reach customers word for word.

A made-up example

A furniture store connects its assistant to the whole website. The returns page says: “You can return any item within 14 days. Made-to-measure sofas can’t be returned.” A product page for a corner sofa says, in a feature list, “Made to your measurements in 6 weeks.” A customer asks whether they can return the corner sofa.

The search finds the returns page’s first sentence and the product page. It doesn’t find the second sentence of the returns page, because it was split into another snippet. The assistant answers: “Yes, you can return any item within 14 days.” The store now has a customer who expects to return a €2,400 made-to-measure sofa.

The fix takes two minutes: put the exception in the same sentence as the rule, and add “made-to-measure” to the product page’s own text: “This sofa is made to measure and can’t be returned.” The point is not that the assistant was careless. It answered from what it found. We see the same mechanism behind most invented policies.

A ten-minute check before you switch it on

1. List every source connected to the assistant.
2. Remove anything older than a year that states a price, a time or a rule.
3. For each of the twelve topics above, find the one page that answers it. If there are two, decide which wins and fix the other.
4. Ask the assistant one question per topic, in the words a customer would use.
5. Compare each answer with the page. Fix the page, not just the answer.

If you’d rather we did the reading and the writing, Assistant Setup starts with exactly this: we read your pages, write down the rules customers ask about, and test about a hundred questions before the assistant goes live. Platform documentation on sources is worth reading too: Intercom’s guide to optimizing content for Fin¹ is the most detailed, and Tidio describes what Lyro learns from on its product page.²

Footnotes

1. Intercom, “Optimizing content for Fin”, intercom.com/help. ↩

2. Tidio, “Lyro AI agent”, tidio.com/lyro. ↩