The one rule: your content only
An AI assistant on your website should answer from your pages, your docs, and your FAQ — and refuse everything else. A bot that improvises from general knowledge will eventually quote a price you do not charge or promise a feature you do not ship, with total confidence. Every step below serves that one rule.
Step 1: feed it what customers actually read
- Start with your top ten questions — the ones your team answers by hand every week. If a question repeats in chat, its answer belongs in the training sources.
- Point it at pages, not PDFs of everything. Your pricing page, FAQ, help articles, and policy pages are written for customers; internal wikis and sales decks are not, and they leak internal language into visitor-facing answers. For a large help centre, import the URLs in batches rather than adding them one by one.
- Fix the source, not the bot. If the AI answers something badly, the page it learned from is usually the real problem. The bot is a mirror held up to your documentation.
Step 2: red-team it before visitors do
Before switching the assistant on, spend thirty minutes trying to break it:
- Ask the top ten questions in sloppy, misspelled, real-visitor phrasing.
- Ask about things you don't offer. The right answer is a graceful "we don't do that" — not an invention.
- Ask something answerable only by a human ("can I get a discount?"). The right behavior is a handoff, not a guess.
- Ask the same question twice in different words and compare the answers for consistency.
Step 3: wire the handoff before launch
The assistant's most important skill is knowing when to stop. Configure the human handoff — visible, one step, with the full transcript carried over — before the bot takes its first real conversation. A visitor who has to fight through "did this answer your question?" loops to reach a person learns that your chat is a wall.
Step 4: read its transcripts weekly
For the first month, read the AI's conversations weekly. Those conversations can also be fed back to the assistant as examples, which is worth understanding before you switch it on — see training your AI on your own past chats. You are looking for three things: questions it refused that it should have answered (a missing source), answers that were technically right but unhelpfully phrased (a source written for lawyers, not customers), and questions that repeat (a candidate for a dedicated page). Each finding is a documentation fix — the bot improves because your content did.
What "trained on your content" looks like in MyLiveChat
Point the AI assistant at your website URL, help articles, or FAQ; it learns from that approved content and hands off to a human with the conversation intact when it should. The knowledge base doubles as both the visitor-facing help center and the assistant's best training source — one set of answers, maintained once.
Content an AI can actually answer from
The quality ceiling of an AI chat answer is set by the source material, and most help content was
written for a human who already has context. Restructuring it pays off immediately and helps your human
readers at the same time.
The pattern that works is one topic per document, the answer stated in the opening sentence rather
than after three paragraphs of preamble, and explicit conditions where they matter — naming the
plan, region or version an answer applies to instead of assuming the reader knows. Documents that bury
the answer produce confident, partial AI responses, because the retrieval found the topic but not the
qualification.
Watch for content that only makes sense next to a screenshot or a table. Anything whose meaning lives
in an image is invisible to a text answer, and the AI will either skip it or reconstruct it wrongly.
Where a visual is essential, restate the key facts in prose alongside it.
What to remove before you index
What you leave out matters as much as what you include, and the common failure is indexing everything
because it is easier than choosing.
Take out anything time-bound that nobody will maintain — old release notes, expired promotions,
last year's pricing — because an AI answering from a superseded page is confidently wrong in a way
that is hard to detect. Take out internal drafts and anything containing customer names or personal data.
Take out near-duplicate pages that say slightly different things about the same topic, keeping the one
you will maintain; contradictory sources produce unpredictable answers.
Then put a review date on what remains. An index built once and never revisited degrades quietly as
the product changes, and the symptom is not an error message but answers that were true a year ago.
What to measure
The most useful number is the share of AI conversations that ended without a person stepping in, read
alongside what happened in the ones that did. Handover is not failure — it is the system working as
designed — but a rising handover rate on a stable product usually means the content has drifted
behind reality.
Group the handovers by cause rather than counting them. Missing content, contradictory content and
questions genuinely outside scope need three different fixes, and only the first two are content
problems.
Sample the answers nobody escalated. Conversations that ended quietly are the ones where a wrong
answer goes undetected, and reading a handful each week is the only reliable way to find confident
errors before a customer does.
If the material you want it trained on is already written up as markdown, splitting one markdown file into help articles covers how a whole folder becomes a set of articles in a single paste, and where the splits fall.