How to Build a Chatbot Knowledge Base That Stays Accurate

A chatbot is only as reliable as the content behind it. If the knowledge base is thin, outdated or contradictory, the assistant will repeat those problems politely and quickly. Building it deliberately takes a few hours of editorial work and saves a long tail of wrong answers.

1. Start from the questions visitors actually ask

Do not begin with your site map. Begin with the questions people send you. Pull them from your inbox, contact form, reviews and the notes your team keeps. Group them into themes such as pricing, availability, process, preparation, location and policies. Aim for the twenty or thirty questions that make up most of your repeat volume.

For each one, check whether your website already answers it clearly. If the honest answer is "only partly," that gap is the first thing to fix.

2. Gather your sources

A useful knowledge base usually draws from three places:

• Website pages: services, pricing, location, team and policy pages.
• FAQs: the short answers your team already gives.
• Documents: brochures, preparation guides, terms and onboarding material.

Choose one source of truth for each fact. If your pricing appears on three pages with three versions, resolve that before loading anything. The assistant cannot know which one you meant.

3. Write answers the way you would say them

Short, direct answers work best. Lead with the answer, then add the condition. "Yes, same-day slots are released each morning" is better than a paragraph that buries the answer in policy language. Match your normal tone. If you are warm and informal in person, the assistant should not sound like a legal notice.

Keep each topic in one place. A page about cancellation should state the rule once, with the exact window and what happens afterwards.

4. Decide what it should not answer

Make a list of topics that need a person: pricing exceptions, complaints, anything that depends on someone's individual situation, and anything where a wrong answer has real consequences. For those topics, the right behaviour is to share the correct contact or capture the visitor's details so someone can follow up.

Writing this boundary down is as important as writing the answers. It stops the assistant from sounding confident about things it should hand over.

5. Test with real and awkward questions

Before launch, ask the questions in the ways people really phrase them: misspelled, vague, two questions at once, and out of scope. Check that each answer is correct, that it comes from the right source and that an unknown question ends in a useful next step instead of a guess.

Then ask your team to try to break it. They know the awkward cases.

6. Keep it current

Accuracy decays when nothing owns it. Assign one person to review the knowledge base on a regular rhythm and whenever something changes, such as new prices, new opening hours, a new service or a changed policy. Read a sample of real conversations each week. Questions the assistant could not answer show you what to add. Answers visitors pushed back on show you what to rewrite.

An admin portal makes this part easy, because staff can correct an answer or add content themselves without waiting for a developer.

A simple review checklist

• Does every high-volume question have a clear answer?
• Is each fact stated once, with one version?
• Are the boundaries for human-only topics written down?
• Does someone own regular review?
• Do we read real conversations and act on what they show?

If you want a chatbot that answers from your own pages, FAQs and documents, with a portal your team can edit, see the Ontix AI chatbot.