Updated by Google candidates
Product Data Scientist (L5) Interview Experience
I just found like nervousness dropped my IQ like 20 points, so I ended up doing a lot of mocks with GPT. Then I started seeing way more causal inference in interviews, like the field can go back to the science part.
Interview process
I had already been referred and was already in touch with a recruiter (for a previous interview process), so the first loop moved unusually smoothly and I even got to skip the technical screen there. The product loop was more standard: recruiter, a technical screen with SQL plus schema design, then a three-round onsite covering stats/modeling, experimentation/applied analysis, and Googliness. The tricky part is the two analytical onsite rounds overlap a lot, but the stats side was more trivia and mechanics than I expected, while the experimentation side stayed pretty close to standard A/B testing frameworks with a little causal inference mixed in. Overall the interviewers came off very friendly and conversational, and I finished my last onsite in mid-December.
- Recruiter screen
- Technical interview
- Final round
Interview tips
I'd go back to basics hard. I drilled fundamentals and patterns until I could do them even when nerves knocked like 20 IQ points off. For SQL, don't assume day-to-day experience is enough because the same interview gotchas repeat and a lot of them never show up at work. I used ChatGPT a ton for mock case studies, stats trivia, and behavioral drilling because the low friction makes it easy to do volume, but I still had to practice out loud since typing is not the same brain as speaking. I'd also prep beyond plain A/B tests now: difference-in-differences, geo or cluster randomized tests, propensity scores, and at least know what synthetic control is. And for Googliness, I need structure or I get too conversational and leave out the stakes and why the story mattered.
Company culture
Google came off really friendly and genuinely collaborative to me. The interviewers were conversational, and when I asked what people liked about working there, they kept saying the people in a way that felt real, not canned. The Googliness screen seemed to care a lot about whether I'm intrinsically motivated, contribute outside my lane, and work well with others. Process-wise it was lighter than I expected at the onsite stage, but the round names are kind of fuzzy because measurement/modeling and experimentation overlap a lot in practice. I also got the sense the field is shifting a bit back toward the science part, so causal inference is showing up more again instead of everything stopping at basic A/B testing.
Questions asked
Overview
The onsite packet was three rounds, not the six-round monster I had built up in my head. The formal labels were things like measurement and modeling concepts, experimentation and applied analysis, and Googliness, but honestly the first two blur together from the outside. Across the two loops, the stats side sometimes felt like open-ended trivia and mechanics, while the experimentation side stayed pretty close to designing and reasoning through tests. The whole thing felt conversational and much friendlier than some other big tech loops I've done.
Question types asked
Specific questions asked
How do we sample from an arbitrary distribution?
This was one of the more trivia-style stats prompts. We had basically made up a distribution on screen, and I had to talk through how I would sample from it. That round had the feel of, if I mentioned a concept, the interviewer might immediately drill into whether I actually understood the mechanics behind it.
How would you compute the statistic for a two-sample test?
How does it change for absolute metrics versus ratio metrics?
I had to know the math cold here, not just wave at the idea. I talked through the mechanics of a t test and z test and, in one of the experiment-focused rounds, exactly how to compute the statistic for a two-sample test. They also cared about whether I understood the difference between absolute metrics and ratio metrics instead of treating every test the same way.
Walk me through how you'd design the experiment and causal analysis for this problem.
When would you use propensity scores?
A lot of these basically reduced to setting up an A/B test cleanly. I used the standard step-by-step framework for defining the hypothesis, metrics, setup, and reliability checks, and in one round I also had to talk about propensity scores for observational data. Google didn't make this wildly exotic, but they did want more than buzzwords.
Get full access with a membership, or share your experience to try it free.
