Naive Bayesian algorithms are based on probability theory, and need a corpus of documents to "train" them. So, if all you are looking for is whether or not a web page is a good match, that might work (after a few thousand matches you'll have a pretty good accuracy rate). What exactly are you trying to do?