Growth Marketer interview questions

A company growth model is a quantified representation of how the business acquires, retains, and monetizes customers over time.

• It defines the unit of growth, system boundaries, time period, and outcomes that must reconcile.
• It connects market inputs such as qualified traffic to product behaviors such as activation and repeat use.
• It makes compounding mechanisms and testable forecasts explicit without treating historical relationships as proof of causal leverage.

I express the growth model as stock-flow equations for a defined unit, population, and period.

• Each ending stock equals its beginning stock plus inflows minus outflows, and each rate-based flow equals an eligible base times a transition rate.
• I reconcile every equation to warehouse actuals and test identities, cohort definitions, and timing so the arithmetic closes.
• I test predictive calibration on held-out or future periods, while causal leverage requires randomized or credible quasi-experiments rather than historical sensitivity alone.

Growth strategy chooses which compounding system the company should build, while performance marketing optimizes paid acquisition inside that system.

• Performance marketing usually owns spend, bids, creative, and channel-level CAC over a relatively short payback window.
• Growth strategy also covers activation, retention, pricing, referrals, sales motion, and the product changes that alter LTV.
• A paid channel can beat its CPA target yet harm growth if it brings low-retention customers or crowds out a stronger loop.

Growth accounting reconciles period-over-period change: new and resurrected customers add to the active base while churned customers leave; recurring revenue also includes expansion and contraction.

• It separates gross acquisition from net growth, so a strong signup month cannot hide worsening churn.
• Cohort views reveal whether improvement comes from better customer quality, stronger retention, or favorable mix.
• For subscription revenue, new MRR plus expansion and reactivation minus contraction and churn reconciles directly to net new MRR.

The binding constraint is the lever whose feasible improvement causally creates the largest durable increase in the growth outcome.

• I combine model sensitivity with practical headroom to identify a candidate constraint, not to claim a causal effect.
• I inspect cohorts, segments, economics, and capacity so mix shifts or downstream bottlenecks do not produce a false priority.
• Where feasible, I confirm the intervention's effect with a randomized experiment or credible quasi-experiment and revise the diagnosis if the lift does not appear.

A good north star metric measures recurring customer value in a form that predicts durable business growth.

• It counts a meaningful value event, such as weekly teams completing a workflow, rather than a shallow action such as app opens.
• It responds soon enough to guide product and growth decisions but remains correlated with retention or revenue across cohorts.
• It is difficult to game, segmentable, and paired with guardrails so volume cannot rise by degrading quality or margin.

Revenue should not be the north star when it arrives too late or can grow without a corresponding increase in customer value.

• Annual contracts can make current revenue reflect sales decisions made months ago rather than today's product health.
• Price increases can lift revenue while activation, retention, or usage depth deteriorates underneath.
• I still keep revenue as an outcome metric and use a validated leading value metric to steer weekly growth work.

I build a driver tree from a stock-flow identity whose terms share one unit and period.

• Active paid teams at period end equal retained active teams plus reactivated teams plus newly activated teams.
• Newly activated teams equal eligible account volume times activation rate, while retained and reactivated teams are distinct flows into the ending stock.
• Each term gets one definition, owner, time window, and source table, and the identity must reconcile to the observed ending total.

Leading metrics provide fast directional evidence, while lagging metrics confirm that the change produced durable business value.

• An onboarding completion rate can move in days, whereas 90-day retention or CAC payback may take a quarter to mature.
• I use historical cohorts to validate that the leading metric predicts the lagging outcome within the relevant segment.
• If that relationship weakens, I revise the proxy rather than declaring success from an easier metric.

Growth guardrails are metrics that prevent optimization of a primary outcome from creating unacceptable damage elsewhere.

• An activation test might guard retention, support contacts, latency, and refund rate while optimizing completed setup.
• Guardrails need predefined decision thresholds based on business tolerance, not a retrospective scan for any negative movement.
• I keep the set small because too many guardrails make every result ambiguous and quietly eliminate useful risk-taking.

An experimentation program at scale requires a shared decision system, not merely a larger number of A/B tests.

• The program needs a portfolio tied to driver-tree constraints, with explicit hypotheses, decision rules, and learning goals.
• The platform needs trustworthy assignment, exposure logging, metric computation, and holdouts through a tool such as Statsig.
• For cumulative impact, I use realized ramp-adjusted effects without simply summing overlapping experiments, validate incrementality with a permanent holdout, and reconcile the estimate to finance actuals.

Experiment velocity should measure how quickly the organization produces trustworthy decisions, not how many flags it launches.

• I track time from approved hypothesis to decision, plus the share of tests that reach a valid conclusion.
• I pair throughput with win rate and incremental impact so splitting one idea into many tiny tests does not look productive.
• I also track learning reuse, because a documented null that prevents three repeated tests has real portfolio value.

I prioritize experiments by expected economic value adjusted for evidence, cost, and time to a reliable decision.

• Expected value combines reachable population, plausible metric movement, downstream value, and the chance the hypothesis is true.
• I reserve capacity for foundational tests and new surfaces so a scoring model does not favor only cheap funnel tweaks.
• Dependencies and interaction risk matter, because simultaneous tests on the same population can delay or contaminate both reads.

Power and minimum detectable effect determine whether a test can reliably detect a change large enough to matter.

• I set the minimum effect from economics, then use baseline rate, variance, allocation, and error rates to determine sample size and duration.
• Switching the metric, audience, or design can change the estimand and does not automatically restore power for the original decision.
• I prefer variance reduction such as CUPED while preserving the target population and outcome; if reliable power remains impossible, I do not launch or treat the result as directional only.

Peeking is repeatedly checking a fixed-horizon test and stopping when significance first appears, which inflates false positives.

• The simple remedy is to precompute the sample and analyze only at the planned horizon.
• When frequent decisions are necessary, I use a valid sequential method or Statsig's documented sequential testing rather than informal daily checks.
• I still require a minimum runtime to cover weekly behavior and avoid a technically valid result driven by one calendar anomaly.

Multiple comparisons increase the chance of finding an apparently positive result when no real effect exists.

• The risk grows when teams inspect many variants, segments, and metrics but report only the best-looking slice.
• I declare one primary metric and limited planned cuts, then control false discovery across genuine families of related tests.
• Exploratory findings remain useful as hypotheses, but they need confirmation in fresh data before they receive budget or product commitment.

Network effects break the assumption that one participant's treatment cannot change another participant's outcome.

• In a marketplace, team product, or referral loop, user-level randomization can contaminate control because treated users interact with untreated users.
• I randomize at a defensible cluster such as team, community, or market and account for the smaller effective sample created by within-cluster correlation.
• If spillovers are the question, a two-stage design can vary treatment saturation across clusters and estimate both direct and indirect effects.

I buy when a mature platform covers the core inference and governance needs, and build only the differentiating layer around our data model.

• Statsig or GrowthBook can provide assignment, feature flags, exposure logs, and statistical methods faster than an internal platform.
• The full comparison includes engineering maintenance, analyst trust, warehouse cost, privacy controls, and the cost of vendor lock-in.
• Custom work is justified when identity, marketplace randomization, or proprietary metrics cannot be represented safely in the product.

A healthy experimentation culture rewards correct decisions and reusable learning rather than positive results.

• Hypotheses and decision rules are written before exposure, which makes post-result storytelling visible.
• Nulls and losses stay searchable with enough context to stop the same weak idea from returning every quarter.
• High-consequence changes require stronger evidence, while reversible low-risk ideas can use lighter validation and move faster.

Meta-analysis combines comparable experiments to estimate a more stable overall effect and explain why results vary.

• I standardize outcome definitions and keep effect sizes with uncertainty, not just wins and losses.
• I model meaningful differences such as channel, customer segment, treatment intensity, and novelty instead of blindly pooling everything.
• The result informs priors and portfolio allocation, while each new decision still accounts for its own population and implementation.