I think GP is fine with bucketing, but they're saying that f(age, gender, interests, ...) is not a good hashing function.
I haven't seen studies on the topic, but on the surface of it, I'm inclined to agree - it doesn't seem to be a good bucketing strategy, and unless we're talking about cold-calling/edge of the funnel, and painting in very broad strokes, it misses the most important factor uniting them all - they're on your site, so they must want something. Hence, "we should use criteria based on what people are trying to achieve right now".