I think it really comes down to how well-designed the performance model is. And averaging a handful of survey answers sure as hell doesn't cut it.