What is training data vs live retrieval?
Training data is material a model learned before it answers. Live retrieval is pages fetched at answer time to inform that reply. An answer can use either, or both. A missing link does not prove the model is using training data, and a link does not reveal the training set.
Why it matters for AI visibility
This distinction matters because the two paths imply different actions. If an engine retrieves live pages, you can improve crawl access, clarity, and the sources it already cites, and you may see a new URL appear without a model update. If an answer comes from training, publishing tonight will not rewrite the weights, and a citation may never appear even when the prose is confident. In practice you often cannot label a single answer with certainty. Search-style products tend to show sources when they retrieved some. A chat with browsing off may not. Do not claim a model was trained on your blog because it knows your tagline. Do not claim it was not, either. Measure the surface you care about, note whether sources were shown, and keep training policy, crawler rules, and citation logs as three separate records.
Example
With browsing off, an assistant describes your founding year and shows no link. With a search-enabled product, the same question cites your about page and uses the year printed there. You can audit the second answer against the page. The first answer is not a citation event, and it is not proof of what the training set contained.
Related terms
Retrieval-augmented generation is one way live retrieval is built. An AI crawler may collect pages for training, for search, or for both, depending on the bot. Grounding is what you can check when retrieval is exposed as sources.