How to train AI on your website content

It isn't training, it's retrieval

When a product says it trains on your website, it almost always means retrieval-augmented generation: your pages are crawled, split into passages, converted into embeddings, and stored. When a visitor asks something, the closest passages are retrieved and handed to a language model with an instruction to answer only from them.

This matters practically. Nothing about the model changes, so updating your knowledge is a re-crawl, not a retraining run — and a wrong answer is usually a retrieval problem or a content problem, not a model problem.

What makes content retrievable

Retrieval works on meaning, so content answers well when it states things plainly and completely in one place. A page that says "we ship across the EU in 5–8 business days, free over €80" is retrievable. A page that spreads the same facts across a hero image, a tooltip and a PDF is not.

Length matters in both directions. A page with forty words has nothing to retrieve; a ten-thousand-word page gets split into passages that may each lack the context to stand alone. Pages of a few hundred to a couple of thousand words, each about one thing, work best.

Diagnosing a wrong answer

Work backwards through the chain. First: does the fact exist on the site as text? If not, that is the fix — write it. Second: was the right passage retrieved? If retrieval surfaced the wrong page, the content probably buries the answer or uses vocabulary customers don't. Third: was the passage retrieved but the answer still wrong? That is a prompt problem, and it usually means the system is not being firm enough about answering only from source.

Keep it current

Set a cadence that matches how often you change things. An ecommerce store during a sale needs re-crawling more often than a professional services firm. The failure mode is silent — nothing breaks, the answers just quietly describe an older version of your business.