What makes this situation [the GHZ experiment] nonlocal is that while all four different local options are present for all observers, i.e., there are four Everett worlds for Alice, and separately for Bob and for Charley, we do not have 64 worlds. Specifying Everett worlds of two observers fixes the world of the third. This connection between local worlds of the observers is the nonlocality of the MWI.
[...]
Is there any possibility of action at a distance in the framework of the MWI? Obviously, at the level of the physical universe that includes all the worlds, local action cannot change anything at remote locations. However, a local action splits the world, which is a nonlocal concept, and local actions can bring about splitting to worlds that differ at remote locations. Thus, an observer for whom only his world is relevant has an illusion of an action at a distance when he performs a measurement on a system entangled with a remote system.