Interview Drills — Growth Funnel & Metrics

Open-ended interview questions. Each has a Framework (the structure a strong answer follows), a Model answer (a concise example), and a Rubric (what an interviewer listens for). Practice thinking aloud and always reason funnel-first and data-honestly: name the stage, look at the numbers, and quantify before you propose. The app can role-play these as mock interviews (see mock-interview.md).

The universal growth-diagnosis structure: Clarify the metric and its definition → break the number into its funnel stages (AARRR) → find the stage that's leaking worst (biggest drop or biggest change) → hypothesize why with data → propose a prioritized, testable fix → state how you'd measure it. Use it on almost any "the number moved / fix this funnel" question. Always separate "is the data even real?" from "what's the real cause?"

D1

difficulty: easy
concept: growth-funnel Sign-ups have been flat for three months even though traffic is up. Walk me through how you'd diagnose it.
Framework: Confirm the metric and definition (what counts as a sign-up?) → split the funnel into stages (visit → landing → start sign-up → complete) and pull the conversion rate at each → find where the drop-off is worst or where it changed → form a hypothesis for that stage → propose one prioritized test → name the metric that would confirm a fix.
Model answer: "First I'd check the data is real — is tracking intact, and is the new traffic the same quality? If traffic is up but sign-ups are flat, the visit→sign-up conversion rate is falling, so the leak is before the sign-up completes, not in the source. I'd break it into landing-page view → sign-up start → sign-up complete and find which step dropped. Say the new traffic is a cheaper channel landing on a generic page: the leak is landing→start. I'd test a channel-matched landing page and measure visit-to-sign-up conversion, not just raw sign-ups."
Rubric: Strong answers question the data first, decompose the funnel into stages, locate the specific leaking stage by looking at conversion rates rather than absolute counts, and tie the fix to a stage-level metric. Weak answers jump straight to "run more ads" or "redesign the site" with no funnel decomposition and no distinction between traffic volume and conversion.

D2

difficulty: medium
concept: biggest-leak Here's a funnel: 100,000 visitors → 8,000 sign-ups → 1,200 activated → 300 paying. Where's the worst leak, and what would you fix first?
Framework: Compute the conversion rate between each stage (not just the totals) → compare each rate to a reasonable benchmark for that stage → identify the stage with the steepest or most abnormal drop → reason about why that stage leaks → propose the highest-leverage fix for that stage → state the metric you'd watch.
Model answer: "Stage rates: visit→sign-up 8%, sign-up→activation 15%, activation→paid 25%. The 8% sign-up rate is roughly normal, and 25% activated-to-paid is healthy. The ugly one is sign-up→activation at 15% — 85% of people who sign up never reach value. That's the worst leak and the highest leverage, because everything downstream depends on it. I'd focus on onboarding and time-to-aha: instrument the activation event, find where new users stall, and test a guided first-run that gets them to value faster. I'd measure sign-up→activation rate, and watch that it doesn't just push the leak downstream."
Rubric: Strong answers convert the absolute numbers into between-stage rates, compare against what's normal, and pick the leak by the worst rate rather than the biggest absolute count (the 92,000 lost at the top is a trap if 8% is a normal visit→sign-up rate). They tie the fix to that stage and note the risk of just moving the leak. Weak answers point at the largest raw drop, ignore benchmarks, or propose fixes for a stage that isn't actually the bottleneck.

D3

difficulty: medium
concept: ab-experiment-design You think a shorter sign-up form will lift conversions. Design an A/B test to find out.
Framework: State a falsifiable hypothesis ("if [change], then [metric] will [move], because [reason]") → define control vs. variant changing one thing → pick the primary metric and a guardrail metric → decide who's split and how (random, concurrent) → state roughly how much data / how long before you trust it → define the decision rule (ship, kill, iterate) up front.
Model answer: "Hypothesis: if we cut the sign-up form from 6 fields to 3, completion rate rises, because less friction. Control = current form, variant = 3-field form — one change only. Primary metric: sign-up completion rate. Guardrail: downstream activation and paid rate, because easier sign-up can bring lower-quality users — I don't want to win the form and lose the customer. Split users randomly and concurrently to remove time and seasonality effects. I'd run until each arm has enough completions that the difference isn't noise, not stop the moment it looks good. Decision rule set in advance: ship only if completion improves and activation holds."
Rubric: Strong answers give a falsifiable hypothesis, change exactly one variable, pick a primary metric and a guardrail (catching the quality-vs-quantity trap), split randomly and concurrently, respect sample size over early peeking, and pre-commit a decision rule. Weak answers change several things at once, track only the surface metric, "stop when it's winning," or have no notion of significance or guardrails.

D4

difficulty: medium
concept: cac-ltv A channel costs $200 to acquire a customer (CAC). Customers pay $25/month and stay 8 months on average. Is this channel worth scaling?
Framework: Compute LTV from the inputs (revenue × lifetime, ideally on margin not just revenue) → compare LTV to CAC and state the ratio → judge against the rough health bar (LTV well above CAC, payback period reasonable) → name what's missing or assumed → give a conditional recommendation, not a blind yes/no.
Model answer: "Revenue LTV = $25 × 8 = $200, so on revenue the LTV:CAC ratio is 1:1 — you exactly break even, which is not worth scaling; the rule of thumb is you want LTV comfortably above CAC, often around 3:1. And that's before costs: on gross margin LTV is below $200, so you'd lose money on each customer. Payback is also slow — 8 months of revenue just to recoup acquisition. I wouldn't scale it as-is. I'd ask whether retention can be improved (longer lifetime lifts LTV), whether CAC drops at scale or rises, and whether margin changes the picture. Scale only if we can move LTV well above CAC or cut CAC."
Rubric: Strong answers compute LTV correctly, express it as a ratio against CAC, know the ~3:1 health bar and that margin (not just revenue) matters, consider payback period, and give a conditional answer that names the levers (retention, CAC at scale). Weak answers see "LTV equals CAC" and call it fine, ignore margin and payback, or give a flat yes/no with no economic reasoning.

D5

difficulty: hard
concept: channel-economics You have a fixed $50K monthly budget across three channels with different CAC and different retention of the users they bring. How do you allocate it?
Framework: Reject CAC-alone as the deciding metric → evaluate each channel on both CAC and the quality (retention/LTV) of the users it brings → rank by LTV:CAC, not cheapest acquisition → account for saturation (CAC rises as you scale a channel) and volume limits → propose a concentrated-but-tested allocation → state what you'd measure to reallocate next month.
Model answer: "Cheapest CAC isn't the answer — a $30 channel whose users churn in a month can be worse than a $120 channel whose users stay a year. I'd compute LTV:CAC per channel using that channel's retention, and rank by ratio and payback, not by CAC. Then I'd weight toward the one or two channels with the best economics rather than spreading $50K thin across all three, while respecting that CAC usually rises as I scale a channel and that each has a volume ceiling. I'd hold back a small test slice to probe a new channel or a saturating one. Next month I reallocate based on realized cohort LTV:CAC by channel, not first-week vanity numbers."
Rubric: Strong answers refuse to optimize on CAC alone, segment LTV/retention by channel, rank by LTV:CAC and payback, account for diminishing returns at scale and volume caps, concentrate rather than spread thin, and set up a measurement loop on real cohort data. Weak answers dump budget into the lowest-CAC channel, treat all acquired users as equal quality, ignore saturation, or split evenly "to be safe" with no economic basis.

D6

difficulty: hard
concept: data-honesty Activation jumped 40% the week after your team shipped a new onboarding flow. Your VP wants to roll it out everywhere. What do you say?
Framework: Welcome the result but pressure-test causation before celebrating → check for confounders (a/b vs. before/after, seasonality, a concurrent campaign, a traffic-mix or tracking change) → ask whether the lift is statistically real or small-sample noise → check whether activation's definition changed → verify the lift holds downstream (retention, paid) not just at the activation event → recommend the responsible next step (controlled validation) without killing momentum.
Model answer: "Great signal — but a 40% week-over-week jump right after a launch is exactly when I'm most skeptical, because lots of things move at once. Was this a clean A/B, or before-vs-after where a campaign, a price change, or seasonality could explain it? Is the sample big enough that 40% isn't noise? Did we change how 'activation' is counted when we shipped? And does the lift survive downstream — are these users actually retaining and paying, or did we just make the activation event easier to trigger? I'd validate with a held-back control or a clean A/B before a full rollout. I'm not slowing us to be cautious — I'm making sure we scale a real win, not a measurement artifact."
Rubric: Strong answers separate correlation from causation, name concrete confounders (before/after vs. A/B, seasonality, concurrent campaigns, metric-definition changes), question sample size, and insist the lift hold downstream before rollout — all while staying constructive about momentum. Weak answers take the 40% at face value, recommend rolling out immediately, or are vaguely skeptical without naming a confounder, a validation method, or a downstream check.