Building Ads Optimization (nima101.github.io)

A new installment in a practical series on building ads optimization engines distills common mistakes and a clear optimization-first framework. The author argues many teams confuse “best ad” matching with allocation: delivering the right impressions under budget and auction constraints is a constrained optimization problem, not just an ML ranking task. Using a simplified second-price auction and an ML model that predicts p(click), the piece shows why naive pacing (no pacing, bucketized, probabilistic) often fails to improve expected clicks. Instead, the optimal strategy is to pick the highest p(click) opportunities until budget/price is exhausted — i.e., find a threshold p* — which can multiply click performance substantially versus ad-hoc pacing. Technically, the author introduces bid shading via a multiplier λ with bid = λ * p(click) (winning if bid >= floor) and reframes pacing as finding λ* such that S(λ*) = Budget, where S(λ) is spend as a function of λ. Solving this requires short-term traffic/purchase forecasting and closed-loop control to adjust λ over time. The write-up emphasizes that ML remains critical (accurate p(click) estimates) but is a component within an allocation and pacing engine, and highlights extensions to multi-ad markets and variable traffic where competition and eCPM baselines matter. Practical takeaway: optimize for allocation and advertiser ROI, not just per-impression relevance.