We currently measure the number of page views served with an unseen notification, and of those, how many times the notifications popup is opened. This gives us a probability, which is a useful number for performance-related things ("should we preload OOUI?"), but it's susceptible to being skewed by users (or bots) that never view their notifications.
I'd like to also track how many page views users wait before opening the popup. With that metric, users/bots who never view notifications show up in the 99th percentile, and we have a median (and a distribution) that tells us something useful.