>one variable explaining 40% of the variation in another is pretty good in social sciences, where you expect all sorts of confounding variables, but quite bad in, say, (well designed) physics experiments.

I think this is part of the answer to your question, nhebb, but also I think we might have some confusion about r. In my understanding, (which is also limited to only AP Statistics) r is the value representing how well correlated the two categories are, and r squared represents how much of the correlation the categories you're using at that time accounts for. Does that sound right?