Training data cutoff, explained
Every large language model is trained on text collected up to a certain point. Anything published after that point, such as a rebrand, a new practice area, a move, a merger or a press feature, is not part of the model's built-in knowledge. Models are retrained periodically, so the cutoff moves forward, but there is always a lag.
For businesses this creates two practical issues. First, an assistant answering without retrieval may describe a brand as it was, not as it is: old services, old leadership, an old address. Second, the effect of improving how the web describes a brand reaches a model's built-in knowledge only at its next training cycle, which is one reason AI visibility work is judged over months rather than days.
The remedy is to work on both layers. Keep facts consistent and current across the sources models learn from, so the next training run picks up an accurate picture. And make sure current pages are easy to retrieve and quote, so assistants that search before answering find today's facts. When a brand is described inaccurately, it helps to find out which layer the error comes from: a retrieved source that can be corrected, or older training data that will take time to change.