Wikipedia has a great overview of how the Visitor pattern works, although the sample implementation that they use is in Java. You can easily port that to Python, though, no?
Basically, you want to implement a mechanism for double dispatch. Each node in your AST would need to implement an accept() method (NOT a visit() method). The method takes, as an argument, a visitor object. In the implementation of this accept() method, you call a visit() method of the visitor object (there will be one for each AST node type; in Java, you’ll use parameter overloading, in Python I suppose you can use different visit_*() methods). The correct visitor will then be dispatched with the correct Node type as argument.