How often do AI models update their knowledge about companies?

There are two clocks running. The slower clock is the training-data refresh: each major model is retrained or fine-tuned on cycles ranging from several months to a year or more, after which the baseline shifts to incorporate newer content. The faster clock is retrieval: any engine using RAG (retrieval-augmented generation, where the engine fetches live web content at query time) (Perplexity entirely, ChatGPT Search, Google AI Overviews, Gemini for many query types) pulls live web content at query time, so a new authoritative article can start influencing answers within hours. The two clocks interact: a training-data baseline that anchors a brand to outdated facts can be overridden by retrieval if the retrieval layer returns strong current sources, which is why source-layer work (improving the sources the engines read, not the engines themselves) has more leverage than waiting for the next retraining.