(C++-like) exceptions pass the error up the stack. Directly to the right place, and ensuring that clean-up takes place along the way without tons of repeated test-and-bail code.
Lisp-like are one better: the stack is searched for a binding to a function which handles the error, and then a new call is generated to that function. If the function returns, the search for the next possible one continues. Or the function can then look for a restart point somewhere the stack (the entire stack still being intact), and perform a dynamic non-local transfer (a big return) to that point.
Dumb error return values make it very easy for an error to disappear without being handled. Oh, the caller has a few things to do before returning your value (things such as clean up that could be an unwind block!) and fumbles the ball somehow, and ends up returning a different value: the information is altered like in the telephone game.
That's the problem: every function level between the source of the situation and the place where it is to be handled intercepts the handling effort by being involved in the control flow. When it should just be getting the heck out of the way (like cars pulling to the side when they hear an ambulance).
Bubbling up error values is an excellent example of an "Anti Pattern".