Splitting one shading model into two by inventory type saved 14% more

Splitting one shading model into two by inventory type saved 14% more

A buyer ran a single bid-shading model across all inventory. Splitting it by auction type revealed the global model was a poor average.

1. They separated wins into first-price and second-price auctions and measured the shading model's accuracy on each.
— On second-price inventory it shaded 20% off, but second-price already discounts, so this overshaded and lost winnable impressions.
— On first-price inventory it shaded only 12%, under-shading and overpaying.
2. The single model split the difference and was wrong in both directions.
3. They deployed two models: aggressive shading on first-price, minimal-to-none on second-price.

Evidence: first-price paid CPM fell an additional 14% versus the unified model, while second-price win rate recovered 6 points. Total spend efficiency on the line improved without volume loss.

Why it matters: shading logic should match the auction mechanic. Second-price already protects you from overpaying, so heavy shading there only loses impressions; first-price has no such protection and rewards aggression. One model cannot serve both.