Every site-trained chatbot vendor uses the same playbook. They take your sitemap, hand it to a headless browser, render each page, parse the HTML, and try to guess which parts are the real content and which parts are the header, footer, sidebar, cookie banner, and chat widget. The guess works for some pages and fails for others, especially on builder-heavy sites where the markup is dense and the content lives inside several layers of containers. Once it fails, the bot is happily training on your nav menu.

SleekAI does not need to guess because it reads the database. A post is its post_content field. A product description is its short_description meta. A category page is its term and the posts tagged with it. The plugin sees the same fields your theme reads when rendering, so the bot trains on the real source rather than the rendered output. You decide what to include from the Variables tab: tick the post types, tick the meta fields, tick the taxonomies, and the index is built from the actual rows.

This also means no separate crawler infrastructure, no robots.txt edits, no whitelisting of vendor IPs, no waiting for a daily crawl to pick up an edit. A post update fires a save_post hook and the relevant chunks are refreshed in seconds. A new product appears in the index the moment it is published. A taxonomy change is reflected immediately. The site does not need to be public for the bot to train on it because the bot does not fetch over HTTP. It reads the database that already powers the front end, which is the only source of truth that matters.

There is a long, slightly tired argument that crawlers are how chatbots train on websites because the website is what the bot is supposed to learn. The argument is a tautology that only made sense when the bot vendor had no other way in. As soon as the chatbot lives inside the CMS, the question changes.

The crawler is no longer indexing the website; it is fetching a less accurate version of data already sitting next to it. Reading the database is faster, fresher, more accurate, and avoids an entire class of integration problems. It is also more honest about what the bot is training on.

A crawler that fetches a product page sees the page title, the product description, the price, and probably also a related-products carousel, the global header, the footer, a cookie banner, and a chat widget. A database-direct read sees the product's title, description, and price fields. The bot trained on the first is being trained on incidental chrome.

The bot trained on the second is being trained on the canonical fields. When the bot makes a mistake, it is much easier to debug the second case because you can point at the exact row that produced the wrong answer. That is the simple thing site-trained chatbots have been missing, and it turns out to be the entire game.