Google Data Analyst Interview Questions

“A good search ends with the person finding what they wanted quickly. So I would look at how often people click one of the top results, how often they come straight back and click another, and how often they retype the query in other words. Each can mislead. No click may mean the answer was visible on the page itself. A long visit may mean a good page or a confusing one. So I would run the change as an experiment, read several measures together, and compare with ratings from trained human reviewers on a sample of queries. I would also check rare queries separately.”

“I would start from the impressions table, since every click belongs to an impression but not the reverse. I would LEFT JOIN clicks on the impression identifier, so impressions without clicks are kept. Then I group by position, count distinct impressions and count distinct impressions that have a click, and divide the second by the first. I use distinct counts because one impression could have two click rows from a double tap. I would filter out known bot traffic first. When reading the result, I would say that top positions also hold the better adverts, so the rate reflects both position and quality.”

“Watch time shows that people kept watching, not that they were glad they did. I would check satisfaction signals: survey answers after videos, likes, and how often viewers choose not interested or leave a video early. I would see how the gain is spread. If it comes from a small group watching much more late at night, that is different from most viewers watching a little more. I would check whether the variety of videos watched became narrower. Last, I would look at whether viewers return in the following weeks. My recommendation would present all of this, not one number.”

“I would start with the filters the company already maintains, since detecting bots is specialist work. Then I would look for what slips through: a single address making thousands of queries, clicks arriving at perfectly regular intervals, sessions that never scroll or move a pointer, or a sudden block of traffic from one data centre. I would not delete rows silently. I would flag them and run the analysis both ways. If the conclusion changes depending on the flagged traffic, that is a finding in itself. For adverts this matters more, because advertisers should not be billed for clicks no person made.”

“On your project we tested a new layout, and the first report showed a very large rise in sign-ups. The team was excited. It was far larger than any change we had seen, so I checked the basics before celebrating. The two groups were not the same size, although they should have been. I traced it to a bug: users on older app versions in the test group were not being logged, so the slow, less engaged users were missing from one side. After the fix, the rise was your result. I now check group sizes first, before looking at any outcome.”

“A product lead asked me for one quality score for the search feature in your project, to put on a weekly slide. I understood the wish, but any single number would hide trade-offs. Speed could improve while accuracy for rare queries fell. I did not refuse. I proposed one headline measure with three guardrail measures beside it, shown as a small table with colours, and a rule that the headline could not be called a win if a guardrail got worse. I explained it with a past example where this would have caught a problem. The team adopted it, and your result.”

“A sudden doubling makes me suspicious before it makes me happy. I would first check whether logging changed: a new app version counting an event twice, or a pipeline that loaded one day two times. Then I would see whether the rise comes from many users or a few, and from phones or from one network, which would suggest automated traffic. If the data looks sound, I would look at the real world: a transport strike, a storm, a big match or a festival can push many people to check routes. I would report the most likely cause and how sure I am.”

“I am interested in the question of whether a product really helped someone, which is harder than whether they used it. In search, a good result can mean the person left quickly. In video, more watching is not always better. I like that the measures here need that level of thought. In my previous role I worked on your project, and the most valuable thing I did was show that a measure we trusted was misleading us. I would like to learn how experienced analysts design experiments at this size, and how they keep a team honest about what the numbers mean.”