Netflix product metrics avoid averages because an average hides the shape of the underlying distribution — a metric like "average minutes watched" can look healthy while masking a large group of near-zero users. The fix is to replace averages with threshold counts: "% of customers who do at least X."
Threshold metrics surface the size of the underdelighted tail directly, which is what a DHM Model (Delight, Hard-to-Copy, Margin-Enhancing) hypothesis is actually trying to move — averages can rise even while that tail gets worse.