Large language model (LLM)

The two-source split is the practical point for a business. Training knowledge is baked in months before you ever see the answer, built partly from what crawlers like GPTBot and ClaudeBot collected, so today's content shapes next year's default answers. Live retrieval, by contrast, reacts within days or weeks: engines search, read current pages and cite them. That means AI visibility work pays off on two clocks at once: quickly through the pages and third-party sources engines retrieve and cite now, and slowly through the accumulated presence that ends up inside future models. It also explains odd behaviour, such as an engine describing your old pricing: that is training memory speaking where retrieval did not correct it. Understanding which source produced an answer tells you which lever fixes it.