A step-by-step way to work
Drill the structure, not the answer
Have AI pose metric-drop scenarios and check whether you start with clarifying questions and data integrity, not a guess.
Practise segmenting
After framing, walk through segmenting the drop (when, where, who) aloud; have AI critique your systematicity.
Brainstorm hypotheses once localised
Ask for internal and external candidate causes — but only after you've narrowed where the drop is.
Draft the confirming query
Have it write the SQL you'd run to test a hypothesis (and explain it).
Reason to the cause yourself
Practise weighing evidence to a conclusion rather than accepting the first plausible cause.
Real prompts, and what good output looks like
Scenario you gave me: 'DAU dropped 15% overnight.' My answer: 'It's probably a bad release, I'd roll it back.' Critique my approach.
You jumped to a cause and an action without diagnosis — the exact trap this question tests.
Before 'bad release': (1) Is the drop real, or a logging/tracking break? An overnight cliff often means measurement, not behaviour. (2) Is it global or segmented to one platform/region/version? (3) Does the timing line up with a deploy, or something external?
Only after integrity + segmentation should you hypothesise 'bad release' — and then test it (completion by version) before rolling anything back.
Restructure: clarify → confirm integrity → segment → hypothesise → test → act.
A prompt you can reuse
Fill in the highlighted parts and paste it into your AI tool of choice. Edit the output — it's a starting point, not a finished answer.
Be my coach for metric-drop / RCA interview questions. Drill the structure; don't hand me the cause. Round 1: Give me a realistic 'metric X dropped Y%' scenario and wait. Round 2: I'll talk through my diagnosis. Critique whether I: clarified the question, checked data integrity first, segmented (when/where/who) before hypothesising, and proposed tests rather than jumping to a cause. Round 3: Once I've localised it, brainstorm internal + external causes with me and draft a query to test the top one (explain it). Flag any time I jump to a cause or an action too early.
What AI gets wrong here
Accepting the first plausible cause
AI will offer a believable cause; banking it is the same error as guessing.
Skipping data integrity
An overnight 'drop' is often a tracking break, not real behaviour.
Memorising a script
A canned RCA sequence fails when the scenario has a twist.
The metric-drop question scores how you think under uncertainty, not which cause you name — and an AI-suggested cause, accepted too quickly, is exactly the failure the question is designed to catch. Use AI to rehearse the discipline (clarify, confirm, segment, hypothesise, test) and to generate scenarios; reason your way to the cause yourself, every time.