finished my Google DS loop in March 2026. targeting L4 DS (quantitative analyst track). here's exactly what the rounds were.
loop structure: SQL / coding (1 round) stats / probability (1 round) analytics case (1 round) product sense (1 round) behavioral / leadership (1 round)
SQL round: used a shared doc, no database to actually run queries against. all written. two questions: first was a multi-join aggregation problem (sales data, find top N customers by revenue in the last 90 days, handle nulls). second was a window function question: compute 7-day rolling average per user per day. if you're not fluent with window functions (ROW_NUMBER, LAG, LEAD, RANK, SUM OVER PARTITION BY), drill them. this is where a lot of candidates get stuck.
stats / probability round: more conversational than I expected. questions like: 'you run an A/B test and see p = 0.06. what do you do?' 'explain the difference between Type I and Type II error to a PM' 'how would you detect whether a coin is biased after 1000 flips?' bayesian vs frequentist framing came up, not in a textbook way but 'how do you think about this in practice'
this round is about depth, not speed. they want to see that you have real intuitions, not memorized definitions.
analytics case: they gave me a scenario: YouTube watch time dropped 8% week-over-week. walk through the investigation. I treated it like a data investigation, not a PM question: check data pipeline first, segment by device / country / content type / user cohort, generate ranked hypotheses, describe what I'd pull from BigQuery to test each one. the 'check data pipeline first' instinct mattered. they said many candidates skip that.
product sense: yep, DS at Google does product sense. not as deep as PM rounds but you need to have an opinion on metrics. 'what's the right success metric for Google Translate' type questions. answer: think about user outcomes, not just activity metrics.
stats gotcha I didn't expect: they asked about experiment design, specifically about how you'd handle a case where your A/B test had network effects (users in control and treatment groups interact). SUTVA violation. know what that is and what you'd do about it (cluster randomization, holdout groups, switchback experiments). not every candidate needs this but at L4+ it's fair game.