The visitor pattern is useful in a language that doesn't have pattern matching. Once you can pattern match natively, its usecases go way down. All the examples in the post are most readable (to me) in the first form given that has explicit recursive calls and one match statement.

In "Benefits of encapsulation" we can see the same visitor being used on a different representation of the data, but the tradeoff should be made clearer. With the visitor pattern you commit to a specific (bottom up) evaluation order. You must produce arguments that the visitor produces for subtrees, even if these are not used. You can't simply "skip" a subtree as shown, which the pattern matching approach allows naturally. Note that in the "reverse polish notation", this evaluation order also naturally follows from the representation and you'd need to preprocess the expression to support skipping, so it's a perfect fit there.