Set Bids for Search Advertising from Business Value

Search Ad Bidding for an E‑Commerce Retailer Hard · System Design, Machine Learning, Advertising · Wayfair, Retail · Hints: ① Bid proportional to the expected incremental profit of a click; ② Use throttling or probabilistic pacing to stay within daily budget; ③ Validate incrementality via randomized hold‑back experiments.

A large online retailer wants to automate bidding on search ads (e.g., Google Shopping campaigns). The system must decide, for each eligible impression, how much to bid while aiming to maximize true incremental profit and respecting a fixed daily campaign budget.

Your task is to design the end‑to‑end bidding logic. Cover the following aspects explicitly:

1. Value model – How do you translate business objectives (profit, revenue, new customer acquisition, etc.) into a predicted monetary value for a single ad opportunity?
2. Bid transformation – Given a value estimate, how do you set the bid, accounting for the fact that the auction is typically second‑price and the submitted bid is not the final cost?
3. Budget pacing – Over the course of a day, how does the system throttle or adjust bids so that total spending stays within the campaign budget while capturing the most valuable impressions?
4. Incrementality evaluation – Describe the experimental design and metrics you would use to measure whether the bidding generates purchases that would not have occurred organically. Distinguish between attributed conversions and true incremental lift.
5. Cold start & robustness – How would the system handle a newly launched product with no history, a sudden shift in conversion rates, or a scenario where the budget is exhausted early in the day? How do organic purchases affect your interpretation of attributed revenue?

State your assumptions before proposing any formulas or architecture. Clarify the auction format (e.g., second‑price CPC), available features (query, device, time, product margin, etc.), attribution window, feedback delay, and any geographical or temporal campaign constraints.

A strong answer will separate prediction components (pCVR, value per conversion) from the bid‑optimization layer, discuss offline model evaluation and online A/B test design, and address budget pacing with delayed conversion signals.