Data Analyst

The ability to tell an analysis project along the chain of impact: question, method, result, decision. Bonus: the candidate names the stakeholders, the assumptions and the uncertainty of the data. Someone who only describes the method (I built a funnel) without naming the decision usually did reporting, not analysis. Someone who talks about strategic insights with no concrete number or action often compensates for a lack of business orientation with platitudes.

A structured investigation method: check the data quality, form hypotheses, talk to stakeholders, bring in new sources. Bonus: the candidate names the lesson from the case (what they would do differently on the first attempt today). Honesty about the time it took (a real data anomaly is rarely cleared up in 30 minutes). Someone who has never doubted their own result rarely checks enough.

Business judgment and the courage to question a request rather than deliver mechanically. Bonus: the candidate names the dialogue with the stakeholder (jointly reframing the question). Someone who delivers every request exactly as it comes builds many dashboards no one uses. Someone who refuses outright (that's the wrong question) with no proposal shows an arrogant posture that quickly leads to conflict at an SMB.

Clarifying the need before negotiating the timeline: which decision is the analysis meant for, what level of detail is enough. Offering options: a simple version today, the complete version in 2 days; a scope cut; pointing to an existing dashboard. Someone who just says yes and delivers something half-baked the same day builds in data-quality problems. Someone who refuses outright shows weak stakeholder management.

A structured investigation: (1) check the data quality and instrumentation (the most common cause), (2) segment (new vs. existing users, device, source, region), (3) talk to the affected functions (was there a release, a campaign change, a third-party outage), (4) share prioritized hypotheses with stakeholders. Someone who immediately alarms management without checking creates a false alarm. Someone who waits and watches misses the window to react.

An explicit prioritization framework: business impact per hour of effort, the strategic importance of the topic, the existence of other sources for the deferred request. Expected method: talk to the three functions to understand the real need behind the request, prioritize together with the manager, give transparent feedback. Someone who prioritizes simply by the requester's rank (management first) shows weak business judgment. Someone who picks two projects without consultation builds in conflict.

A structured approach: (1) sharpen the activation definition (which concrete step counts as activation), (2) instrument the funnel and identify the most costly drop-off points, (3) build cohorts (by acquisition source, onboarding variant, segment), (4) descriptive analysis before any modeling (logistic regression or similar comes only when the question requires it). Bonus: the candidate distinguishes correlation from causation and proposes an experiment to validate an identified hypothesis. Someone who immediately talks about machine learning or a random-forest model without framing the funnel shows a method weakness.

Structured data scrutiny: (1) check the data sources (is the tracking running, has an API change distorted the event volume), (2) manually check samples (does a single conversion match what is in the CRM or logs), (3) talk to the developers (was there a silent bug in the tracking layer), (4) pull historical comparison values from alternative sources. Someone who just reports the numbers without questioning builds blind trust in faulty data. Bonus: the candidate recognizes that overly stable numbers are often a stronger warning signal than chaotic ones.

Clarification before the method: which assumptions on acquisition, churn, expansion; which level of detail per segment. Method: a simple cohort projection with churn and expansion rates, sense-checked against historical data; not necessarily an ML model. Bonus: the candidate delivers scenarios (best, base, worst case) instead of a point estimate and names the sensitivity to 1 to 2 key assumptions. Someone who jumps into a complex time-series analysis without clarifying shows no feel for the stakeholder's need.

Command of window functions or self-joins, CTEs for readability, correct handling of users with no purchase (LEFT JOIN, NULL handling), aggregation at the right granularity. Bonus: the candidate asks whether sign-up and purchase are in the same table, whether multiple purchases occur, which time-zone convention applies. Someone who writes a query straight away without clarifying that does not run on a typical data structure shows weak practice. Someone who uses no CTE or window function and works with nested subqueries is usually at a lower SQL level.

A healthy cadence logic: leading indicators (qualified pipeline, activation rate, trial-to-paid conversion) for ongoing steering, lagging indicators (MRR, ARR, Net Revenue Retention, churn) for strategic assessment. Bonus: the candidate distinguishes operational from strategic metrics and adapts the set to the company's stage (a seed-stage SMB does not watch the same things as a Series B scale-up). Someone who lists 15 metrics with no hierarchy has never owned real steering.

A clear conceptual understanding: correlation can be observed, causation can only be plausibly established in a controlled experiment or via quasi-experimental methods (difference-in-differences, regression discontinuity, instrumental variables). Bonus: the candidate names a concrete business example (e.g. users who read our guide churn less does not mean the guide prevents the cancellation; maybe engaged users are more loyal anyway). Someone who cannot articulate the difference clearly will often sell spurious effects as causal in analyses.

Integrity and business judgment. Bonus: the candidate describes a concrete case where they reframed the question to deliver an honest analysis (instead of either delivering mechanically or refusing confrontationally). Someone who simply says I deliver what is ordered shows an order-taker posture with no business partnership. Someone who refuses outright shows a weakness in the stakeholder relationship. The healthy answer lies in dialogue.

An active mentoring posture: training in SQL or in reading dashboards, documenting common pitfalls, pedagogical reviews of requests. Someone who says I help when asked, with no more concrete detail, shows a passive posture. At an SMB with a small data team, the ability to carry the data culture more broadly is decisive for the sustainability of the analytics function.

Openness to criticism and an ability to learn. Bonus: the candidate names the concrete trigger of the change of mind (an alternative data source, a sample check, a methodological argument). Someone who describes having explained their own logic to the critic instead of listening shows a coachability weakness. At an SMB where the analyst often works alone, the ability to actively seek criticism is a hard quality signal.