What a model knows versus what it looks up

There are two ways an AI answer can mention your company. One is from memory, baked in when the model was trained on a huge slice of the web. The other is from retrieval, where the system searches the live web at the moment of the question and reads what it finds before replying. Knowing which is which explains a lot of otherwise baffling behaviour.

Memory is frozen and dated. A model trained last year carries last year's view of you, including facts that are now wrong, and no amount of updating your site changes what is already in its weights. Retrieval is live. When the system searches before answering, your current pages can shape the reply within hours of publishing them.

This is why fresh, specific, well-structured content matters so much for visibility. It is the part of the picture you can still influence. You cannot rewrite the model's memory, but you can make sure that when it looks, it finds a clear and current account of you rather than leaving the reply to its stale recollection.

Most serious assistants now retrieve for anything that smells current or specific, which is most commercial questions. That is good news. It means the door is open. The brands that keep publishing clear, checkable material are the ones standing in the doorway when the model comes looking.