Skip to content
Knowledge and answers

Train the agent on your website

Knowledge is the set of pages, files and notes the agent searches before it answers. Point it at your website and it takes what a visitor could read; anything behind a login, a password or a cookie wall it cannot reach, and neither can your customers. The gaps you have to fill by hand are the things your website never said out loud.

Updated 2026-09-19 · 5 min read · More in knowledge and answers

What the crawl actually takes

It reads the pages it can reach from the address you gave it and files each as a source. A source moves from pending to ready once it has been processed, and only ready sources are searched — a knowledge list that is still mostly pending is why a fresh agent sometimes says it does not know.

It takes what is public. A staging site behind HTTP auth, a customer portal behind a login, a PDF that only exists in a shared drive: none of those are reachable, and none of them will appear. If a page is important and the crawler missed it, add its address by hand rather than assuming it will turn up later.

Fill the gaps by hand

Three kinds of source can be added directly: a URL, a file, or a plain note you type. Notes are the underrated one. Most of what a support agent needs to know is not on any page — the hours you actually keep, which carriers or insurers you take, whether you are accepting new customers this month, what happens to a return after thirty days.

Write a note the way you would answer the question out loud, in a sentence or two. Short, specific notes retrieve better than a long document where the useful sentence is buried in boilerplate.

  1. Find the questions

    Read a week of conversations and write down every question the agent could not answer.

  2. Answer each one in a note

    One note per question, in the words you would use on the phone.

  3. Re-test the question

    Ask the agent the same question in the test chat and read the answer it gives now.

Keep facts out of the instructions

The agent's instructions are how it behaves: its tone, its house rules, what it must never promise. Its knowledge is what it believes to be true. The two get confused constantly, and the confusion is expensive in exactly one direction.

A price, an address or a set of hours typed into the instructions becomes something the agent asserts with total confidence, forever, with no page behind it and no way for the crawl to correct it when you change your website. We have watched an agent quote a price that existed nowhere except a hand-typed line in its instructions — to a real customer, twice. Put facts in knowledge, where they have a source and can be updated.

Keep it fresh

Re-crawl after a meaningful change to your site — new pricing, new hours, a new product line. Knowledge is a snapshot, not a live mirror, and a page you updated this morning is not updated in the agent until it has been read again.

Remove what has gone stale, too. An old promotions page that still ranks on your site is a page the agent can quote from, and a discount that expired in March is worse than no answer at all.

FAQ

Related questions

Enough to answer the questions people ask, and no more. A hundred well-chosen pages beat a thousand that include every blog post and archive page — noise competes with the passage you actually wanted.

Still stuck? A person will answer.