Asking the Right Data Question Before Touching Data - Interview-Ready Framework
A quick-commerce team sees late deliveries rising on the dashboard, and the room immediately asks for rider-speed data. But the real leak may be stockouts, batching logic, lift wait time, wrong promise estimates, or customers adding items after checkout. Data is powerful only after the question has earned the right to be answered.
- Do not start with data. Start with the decision that will change if the analysis gives a clear answer.
- A strong analytics question has five parts: decision, outcome metric, population, comparison, and time window.
- Move from business question to analytics question to data question. Mixing them creates vague analysis.
- The best first output is not a chart. It is a question statement: “For whom, over what period, does X affect Y enough to change Z decision?”
- Use a funnel: broad problem - decision - hypothesis - measures - required data.
- The most dangerous trap is optimizing a convenient proxy, such as clicks, when the business actually cares about retention, profit, risk or trust.
The big idea is simple: the right question turns a messy business situation into a measurable decision problem. Before you touch rows and columns, you should know what decision is at stake, what evidence would change that decision, and what data is genuinely needed.
The Core Idea: A Data Question Is Not the Same as a Business Question
Most weak analysis starts with a sentence like, “Let us analyse customer data.” That sounds useful, but it hides the decision. Analyse for what - pricing, churn, cross-sell, credit risk, delivery reliability, customer satisfaction, or fraud?
A strong analyst separates three levels:
For example, “Why are sales down?” is a business concern, not yet an analytics question. A sharper version is: “Among first-time buyers acquired in the last 90 days, has the repeat purchase rate fallen more in discount-led cohorts than in organic cohorts?” Now you know the population, metric, comparison and time window.
The Five-Part Question Statement
Before opening Excel, SQL, Python or a BI dashboard, write the question in this format:
For [population], during [time window], does [driver or comparison] affect [outcome metric] enough to change [business decision]?
A Six-Step Process Before Touching the Data
The 2x2 Test: Is the Question Worth Answering?
A question can be answerable but useless, or important but not measurable yet. Place it on this matrix before analysis.
How to Measure Question Quality
Use these checks like a pre-analysis scorecard. They prevent the classic mistake of building a polished dashboard around an unclear business problem.
A Small Worked Example: Reframing “Why Are Orders Late?”
Suppose an operations team has 1,200 orders in a week. An order is considered late if it is delivered more than 5 minutes after the promised time. Out of 1,200 orders, 168 are late.
Late delivery rate = late orders / total orders = 168 / 1,200 = 14%.
The weak question is: “Why are deliveries late?” It is too broad. Now split by delivery context:
- High-rise apartments: 110 late out of 500 orders = 22%.
- Independent houses: 58 late out of 700 orders = 8.3%.
The sharper question becomes: “For evening orders in high-rise apartments, do handoff and building-entry delays drive late delivery enough to change promise-time rules or rider allocation?”
Notice what changed. The team moved from a vague complaint to a testable question with a segment, metric, suspected driver and decision.
Definitions You Can Say in One Breath
- Data: ISO/IEC 2382 defines data as reinterpretable representation of information in a formalized manner suitable for communication, interpretation, or processing.
- Business question: The decision uncertainty a manager must resolve before taking action.
- Analytics question: A testable question linking variables, population, timeframe and method to a decision.
- Hypothesis: A specific, testable statement about the expected relationship between variables.
- Operational definition: The exact rule used to measure a concept consistently.
- Proxy metric: A measurable substitute used when the true outcome is hard to observe directly.
Case Study: Urban Company and the Question Behind Service Reliability
Urban Company shows why marketplace analytics must ask the right operational question before studying ratings, cancellations or partner performance.
Urban Company operates a managed marketplace: customers book home services, trained professionals accept jobs, and the platform has to protect reliability, quality and trust. A lazy analytics question would be, “Which partners get low ratings?” That framing can quickly become unfair because a rating may reflect partner skill, customer expectations, job complexity, travel delay, category mismatch, or a problem in the booking flow.
The stronger question is decision-led: “For which service categories and booking contexts do cancellations or low ratings arise from controllable platform decisions such as matching, slot design, scope clarity or partner preparation?”
The move is not just to collect more data. It is to separate the possible causes before analysis. Customer-side signals, partner-side signals, job-context signals and platform-design signals must not be mixed into one blame metric.
The lesson is valuable for interviews: Urban Company’s reliability depends chiefly on precise marketplace design - matching the right professional to the right job at the right time. That primary driver is supported by training standards, category-specific operating processes, customer expectation-setting, location data and feedback loops. The strategic “so what” is this: ask the question at the level where management can act, not at the level where data happens to be easiest to pull.
How AI Changes Asking the Right Question Before Touching the Data
AI makes question framing faster, but it also makes bad framing more dangerous. In 2026, the analyst who wins is not the one who blindly asks an AI tool for insights; it is the one who gives AI a decision-first problem.
- LLMs can generate issue trees quickly. ChatGPT, Claude or Perplexity can turn “repeat purchase is falling” into possible drivers like acquisition source, first-order experience, pricing, delivery, assortment and reminders. Treat this as hypothesis generation, not truth.
- Unstructured data becomes usable earlier. Customer chats, call transcripts, app reviews and sales notes can now be summarized and clustered. But the question must specify the outcome - churn, complaint resolution, refund risk, NPS or conversion - and respect privacy obligations such as India’s DPDP Act, 2023.
- AutoML increases the proxy-risk problem. If you ask the wrong question, AI can optimize the wrong target very efficiently - for example, clicks instead of profitable retention or short-term approvals instead of credit quality.
Load a company annual report, investor presentation or case prompt into NotebookLM. Ask: “Convert this business issue into five decision-first analytics questions. For each, list the outcome metric, population, timeframe, required data fields and possible proxy risk.” Then critique the output yourself.
Interview Relevance
“Suppose repeat purchases have fallen for an e-commerce app. You have transaction and clickstream data. How would you start the analysis?”
In your answer, say: “Before analysing, I would write the question in a testable format.” This signals maturity because you are not jumping straight into tools.
Common Mistake
Jumping straight to the dataset or algorithm. Candidates say, “I will run regression” or “I will make a dashboard” before defining the decision, metric or population. It costs them because interviewers see tool-first thinking, not business judgment. Fix: always begin with the decision sentence, then the measurable analytics question, then the data required.
What to Revise Next
Once you can frame the right question, revise the two topics that decide whether your analysis is valid: Where Data Comes From: Events, Transactions, Surveys & Third-Party Sources, and Data Types & Measurement Scales, and Why They Decide Your Test. The journey is: ask the right question, identify the right source, then choose the right measurement and method.