What "Train the AI on Your Website" Actually Means Under the Hood

"Train the AI on your website" is on every AI support tool's homepage, ours included. It is also almost never explained, which leaves buyers imagining anything from magic to fine-tuning a private GPT on their blog. What actually happens is more mundane and more trustworthy. Here is the pipeline, in the open.

It is not training — it is indexing

No model weights change when you connect your site. Instead, the system builds a searchable index of your content and lets a language model read from it at answer time. The industry name is retrieval-augmented generation; the honest description is "an open-book exam where you wrote the book."

This distinction matters practically: updates are instant (re-index a page, and answers change) and removals are real (delete a page, and the AI genuinely does not know it anymore — try getting that from a fine-tuned model).

Step one: crawl what customers can see

The crawler fetches your public pages, your help center, and anything you add directly — pasted policies, question-and-answer pairs, and plain-text files of past conversations you have curated. It respects robots.txt, renders JavaScript pages, and skips what is login-walled unless you explicitly connect it. What the AI knows is exactly what you would show a stranger, plus what you deliberately added.

Step two: chunk and embed

Pages get split into passages — a heading plus its content, a table with its context — and each passage is embedded: converted into a vector that captures meaning, so "how do I get my money back" lands near your refund-policy passage even though it shares no keywords with it.

Chunking quality is where tools quietly differ. Split too coarsely and retrieval drags in whole pages of noise; too finely and answers lose the caveat sentence that lived one paragraph up. This is also why article structure matters — one question per article retrieves clean. We wrote a whole post on that.

Step three: retrieve, then answer with citations

When a customer asks something, the question is embedded the same way, the closest passages are pulled, and the model is instructed to answer from those passages and cite them. In Chattering, every AI reply carries its sources, and an answer that retrieves nothing relevant becomes a handoff, not an improvisation.

Grounding is the trust boundary. The model is not asked "what do you know about refunds?" — it is asked "given these three passages from their site, answer this question or say you cannot."

What this means for you

Three practical consequences. Your content is the model: gaps in the help center are gaps in the AI, and the fastest tuning is writing the missing article. Freshness is content work, not model work: update the page, add the source again, and the answers change with it. And auditability is real: when an answer is wrong, you can see which passage misled it and fix the passage — not file a mystery-model bug report.

Connect a site and the whole pipeline runs in a few minutes — the free tier is enough to watch it index your own docs and start answering from them.