Or if you don't want to add in a prior specific belief about americans vs. non-americans, just keep one counter per option per continent-of-source-IP (or whatever) and the learning algorithm should work it out on its own. Of course, if you use too many bins then learning is going to take far too long.