Design an Ad Click Prediction System — ML System Design Interview (Staff-Level, 2026)

Build the most accurate click predictor on earth and your ad business will still lose money — because an ad system isn't a prediction, it's an AUCTION. You predict click-through rate (pCTR), but you never rank by it: you rank by expected value = bid × pCTR (eCPM), and that only works if the probability is CALIBRATED. Here's the full staff-level breakdown for the 2026 ML system design interview. In this deep dive: • Why you rank by eCPM (bid × pCTR), not click probability — and second-price / GSP auctions • Calibration: why 0.02 has to mean a true 2% (Platt scaling, isotonic regression) • The latency wall: 50k auctions/sec in under 100ms → the funnel + feature store • Features: 50–200 per request, 100M+ sparse IDs → embeddings, counting features • Model: memorization vs generalization — Wide & Deep → DCN → DLRM • The imbalance: under 1% clicks, why accuracy lies → log loss, AUC, PR-AUC, downsampling + recalibration • Delayed feedback: conversions arrive days late → model the delay, don't call it negative • Position bias & the feedback loop → inverse propensity weighting + exploration • Clicks are a proxy → multi-task CTR + CVR, and the clickbait trap • Budget pacing & the marketplace; offline AUC vs online A/B revenue; cold start & bandits Practice this exact interview — with an AI interviewer that pushes back — free at https://craqit.io ⏱ Chapters 0:00 The 99%-accurate model that loses money 1:59 Requirements 4:05 The latency wall 6:00 The auction (bid × pCTR = eCPM) 8:23 Calibration — 0.02 must mean a true 2% 11:17 The funnel 13:25 Features & the feature store 16:13 Memorization vs generalization 19:02 The metric that lies 21:28 Delayed feedback 24:30 Position bias & the feedback loop 26:41 Clicks aren't the goal 29:16 Budget pacing & the marketplace 31:43 Offline vs online 34:11 Cold start under an auction 36:42 Serving architecture 39:53 Senior follow-ups 41:29 Red flags 43:11 Recap ML System Design series — more in this playlist #systemdesign #machinelearning #mlsystemdesign #adtech #ctr #ranking #ai #faang