What is a Good Recommendation Algorithm?

Iterative algorithms based on feedback loops are only as good as the input to the feedback function. When the problem domain is something as subjective as "what movies do you like?" or "does this collection of links satisfy your vaguely-worded query?", it can be very hard to say what a good solution is.

AFAIK, Google has traditionally solved this problem by "revealed preference" -- what did you actually click on, and how long did you stay there. Greg Linden writes about a more mathematical approach, as taken by Netflix.