Google Ads Serving Under a 50 ms Budget: Retrieval, Ranking and Auction Path

Assume Google's advertising system has about 50 milliseconds to pick the right ad for each request. As a product manager, explain how the system works so that it can make that choice so fast: who the parties are, what happens between the moment a user triggers an ad request and the moment the ad is returned, and which design choices keep the decision within the time budget. ```hint Split the work Separate what must be computed for this particular request from what could have been prepared before the request ever arrived. ``` ### Clarifying Questions - Should I focus on ads on Google Search results, on display ads on partner websites and apps, or both? The flows differ. - Does the 50 ms cover the full round trip to the user's device, or only the time from the request reaching the ad servers to the response leaving them? - How deep should I go: the product-level flow a PM explains to stakeholders, or component-level design? ### What a Strong Answer Covers - The parties involved (advertiser, user, publisher, ad platform, and outside bidders where relevant) and what each one supplies. - The request path from query or page load through candidate retrieval, filtering, prediction, auction and creative assembly. - What "right ad" means: a balance of bid, predicted relevance and quality, and user experience, not simply the highest bid. - The concrete techniques that hold latency down, and what happens when a step runs late. - How campaign edits, budgets, billing and logging are handled outside the request path. ### Follow-up Questions - Engineers have a more accurate click-prediction model that needs 20 more milliseconds per request. How would you decide whether to ship it? - An advertiser complains that a budget change did not take effect right away and they overspent. How would you explain it and what would you change? - Which metrics would tell you whether the serving system is striking the right balance between speed and relevance?