Most blockers operate at the domain/app level, but whether something is distracting really depends on what you’re trying to do.
With Nono, you describe what you intend to work on at the start of a focus session. It then uses an on-device model to look at the active window in that context. So a relevant YT video can stay open, while an unrelated video gets blocked.
The tricky part right now is the model. It needs to be small and fast to run continuously, while still being capable enough to understand the context and not annoy you with false positives.