What is training data?
Training data is the text a language model learns from before release: web pages, books, code and licensed or public datasets. What was widely written about a brand before the cutoff shapes what the model knows without searching. Crawler rules can limit future training use, but they cannot remove what was already learned.
What you control
You can disallow training crawlers in robots.txt for the future. You cannot change a released model. The practical lever is the open web: accurate, widely repeated facts.
Example: a company discussed in thousands of articles before a model's cutoff is usually described in more detail than a company founded the month before.
Training is not search
Blocking a training bot does not remove you from answers that use live search. See training vs search crawlers.