For 2x2, I've heard top solvers just learn full EG and tonnes of other random algs like EG with a twisted corner on the D layer, and L5C (LS) algs so that they can solve almost any scramble in a pretty optimal way.
For 3x3, ZBLL isn't a method, it's a substep and there are loads of methods that have it, for example:
CFOP sometimes (more with edge control), ZB, ZZ, Nautilus-LSLL, Mehta-TDR, Petrus, APB, and the list goes on.
ZB as a full method (ZBLS/ZBLL), is probably not worth it because of some worse recognition and some ZBLS algs saving barely any moves over regular LS/OLL/PLL or LS/EO/ZBLL. But learning large chunks of ZBLL, at least for CFOP, seems to be becoming a "meta" of sorts. Of course, there is also Roux which I consider the main contender for best method other than CFOP, because there seems like much more potential left to be unlocked, and the method is significantly more efficient in movecount.
We appear to be at a point in speedcubing, at least for 3x3, where even tiny optimisations are being used to their highest potential; so developing methods further with new techniques could be very useful. Maybe not too far in the future, top level solvers will be learning large amounts of 1LLL and drilling them until recognition isn't an issue; I don't know much about roux, but I'm sure there's some algset that seems unfeasible to learn, but will eventually start being incorporated into solves (perhaps pure L6E, but I don't know much as i said).
