Train the AI chatbot

The AI chatbot answers from content you give it. Everything about answer quality traces back to what you fed it, so this is the step worth doing carefully.

The AI chatbot is available from the Professional plan upward. Usage is metered in AI credits — see pricing.

Sources you can train on

• Website pages — point it at URLs or a sitemap and it ingests the published content.
• Help centre and FAQ pages — the highest-value source, because the content is already written as answers.
• Uploaded documents — policies, manuals, product sheets.
• Past tickets and conversations — real customer phrasing, which improves matching.

What makes a good source

Content written as an answer to a question. Content that describes what a thing is usually trains well; content that is mostly navigation, marketing copy or calls to action usually trains badly and dilutes everything else.

Train on: shipping and returns policies, product specifications, setup guides, pricing explanations, opening hours, FAQ pages, troubleshooting steps.

Do not train on: your homepage, category listing pages, blog index pages, careers pages, legal boilerplate you do not want quoted back at customers, and anything out of date. A stale returns policy in the training set is worse than no bot at all, because it will be quoted with total confidence.

The first pass

1. Add your help centre and FAQ pages.
2. Add your policy pages — shipping, returns, refunds, terms that customers actually ask about.
3. Test it in the playground with the twenty questions you get most often, phrased the way customers phrase them, not the way you would.
4. Fix what is wrong at the source. If the bot gives a bad answer about returns, the returns page is usually the problem — and fixing it helps your human customers too.

Retraining

When you change a source page, retrain so the bot picks it up. The habit worth building: whenever someone changes a policy, retraining is part of that task, not a separate one that gets forgotten.

Reading the unanswered questions report

This is the most useful thing the bot produces. Every question it could not answer is logged. That list is:

• a content backlog, ranked by real demand
• a list of the places your documentation is unclear
• occasionally, a product problem wearing a support costume

Work through it regularly and the deflection rate climbs without you touching the bot’s configuration at all.

Next: guardrails and handoff — making sure it declines gracefully rather than inventing an answer.