Users will give you the preferences that they think are important. Or are important at the point of that interview.
Analytics can give you a decent glimpse of revealed preferences, which may or may not be what you're after.
Whether or not this is a good thing, depends on a lot of subjectivity, sure. But suppose you run a porn site - if you asked most users what they wanted in porn (before they had seen any), they would probably say one thing. If you examine what kinds of videos people look at, you'll see another. (This theme, with actual data from pornhub, is explored at length in the book "Everybody Lies" by Seth Stephens-Davidowitz.)
Both routes (asking and instrumenting) have their uses.
