I'm having a hard time finding answers to this as well. I'm currently invoking V3 on every page as soon as it has completed loading, and likely before any cursor movement or navigation has taken place at all.
In my case I am wondering if I'd get a more accurate reading if I delayed that call until after the user has had some time to interact with the pages.