InsertChatStart free

Website Training Guide

How to train an AI chatbot on your website

Training a useful website assistant starts with source selection, not a giant crawl. Choose the pages that contain maintained customer answers, remove irrelevant material, test real questions, and define what happens when the source is missing or uncertain.

Built for teams creating a grounded assistant from website content.

Build a preview from my URL

We read public pages from the URL you provide. Review discovered sources before anything is published.

Interactive proof

Create the first preview from your URL

Use a real public website, then inspect the discovered source set before judging the assistant's answers.

  1. 1Import the homepage and the few pages that answer common customer questions.
  2. 2Remove duplicate, stale, legal-only, private, or irrelevant pages from the source set.
  3. 3Test routine, ambiguous, outdated, and unsupported questions before publishing.

Choose high-signal sources

Start with maintained pages that support customer decisions and support.

  • Services, products, pricing, policies, hours, and locations
  • Help articles, FAQs, setup guides, and approved documents
  • Specific pages that contain the answer instead of broad archive indexes

Remove sources that create noise

A larger source set is not automatically a better one.

  • Exclude duplicate, stale, thin, or contradictory pages
  • Keep private, authenticated, and customer-specific data out of public ingestion
  • Separate legal reference content from the answers the assistant should provide

Test and improve from real questions

Use observed visitor language and outcomes as the quality loop.

  • Test exact customer questions and follow-up phrasing
  • Review grounded answers, citations, uncertainty, and handoff
  • Update the maintained source or workflow when the same gap repeats

Human control

Limits to verify before launch

  • A crawler cannot determine which business claims are current or approved; a human source review is required.
  • Private customer data should not be ingested into a public assistant source set.
  • Source changes require refresh and regression testing before the new answer is trusted.

Page-specific FAQ

Operational questions, answered

Should I train the chatbot on my entire website?

Usually not at first. Start with high-signal maintained pages, test the result, then add sources that solve a confirmed coverage gap.

How often should sources refresh?

Refresh after material source changes and on a schedule appropriate to how often the business information changes. Test critical answers after each material refresh.

What if two pages contradict each other?

Resolve the source conflict before launch or remove the stale page. The assistant should not be expected to choose the business's authoritative policy.

Can I test this before paying?

Yes. The self-serve trial lasts seven days. Review imported sources, test the intended workflow, and verify handoff or fallback behavior before publishing the website assistant or activating a receptionist.

Product facts last reviewed 2026-08-17.

Troubleshooting checklist

  1. Reproduce the workflow with a representative customer question or call before changing the configuration.
  2. Confirm the source, connected action, permissions, destination, and fallback independently.
  3. Retest the successful path and one unavailable or ambiguous path, then assign an owner for any unresolved outcome.

Sources and scope

Product behavior and limits should be checked against the current linked product documentation before implementation.