AI & Technical question

How does Google Ads work such that it delivers the right ad in 50ms?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Technical depth on real time ad auctions and how systems hit tight latency budgets.

State the constraint: the auction, targeting, and creative selection must complete in roughly 50 milliseconds.
Explain pre-computation: audience segments and eligible ads are pre-indexed so the auction only scores a filtered candidate set.
Explain infrastructure: in-memory data stores and distributed low-latency serving systems avoid slow database calls.
Explain parallelization: candidate scoring happens across many machines simultaneously rather than sequentially.
Explain simplification tradeoffs: complex ML ranking models are often distilled or cached to run fast enough at serving time.
Tie back to the PM angle: latency budget forces tradeoffs between ad relevance quality and speed.

Names the candidate generation then ranking funnel pattern, common to large scale ad and recommendation systems.
Explains that heavy models run offline or asynchronously, with lightweight scoring at request time.
Flags the PM tradeoff explicitly: shipping a more accurate but slower model risks missing the latency budget and losing the auction.

Describing the auction mechanics without addressing the latency constraint at all.
No mention of pre-filtering or candidate generation to reduce compute.
Treating this as purely an infrastructure question with no product tradeoff.

How would you decide whether a more accurate but slower model is worth shipping?
What would you monitor to catch a latency regression before it affects revenue?