Amazon Analytics Engineer Interview Questions (2026)

1. 1Online Assessment (SDE OA)60 minMedium

   Two timed coding problems plus a workplace-simulation and logic section; the main gate for freshers and India volume hiring.

2. 2Phone screen45 minMedium

   One coding problem plus 1-2 Leadership Principle STAR questions with an SDE.

3. 3Coding loop round60 minMedium

   DSA problem to working code, followed by assigned-LP behavioral questions in STAR format.

4. 4System design loop round60 minHard

   Design an Amazon-scale service with capacity math, plus LPs; low-level/OOD design substitutes for junior candidates.

5. 5Hiring Manager round45 minMedium

   Team fit, project deep dives, and Deliver Results/Bias for Action stories with the manager you would report to.

6. 6Bar Raiser60 minHard

   An interviewer from outside the team stress-tests LP stories and overall bar with the hardest cross-examination of the loop; holds veto.

1. Q1

   Design an A/B test for a new Retail Marketplace ranking or recommendation change. Define hypothesis, primary metric, guardrails, randomization unit, and launch decision rule

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Context: Amazon wants to grow retail conversion while keeping delivery promises and inventory healthy.

   How to answer: A strong answer will define a clear, testable hypothesis, such as 'The new ranking algorithm will increase customer purchases.' The primary metric should directly reflect the hypothesis, like 'Add-to-Cart Rate' or 'Conversion Rate per User.' Guardrail metrics are crucial to ensure no negative impact on other key areas (e.g., 'Page Load Time,' 'Revenue per User,' 'Return Rate'). The randomization unit should be carefully chosen, typically 'User ID' or 'Session ID,' to avoid contamination. Finally, a clear launch decision rule, incorporating statistical significance (e.g., p < 0.05) and practical significance over a defined test duration, should be articulated.

2. Q2

   For Prime, should randomization happen at customer, session, device, seller, or region level? Explain the tradeoffs

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Consider cross-device behavior, interference, marketplace effects, and operational feasibility.

   How to answer: Randomization for Prime should primarily happen at the customer level. This ensures a consistent experience for an individual user across sessions and devices, preventing contamination and ensuring independent observations. Session or device level randomization could lead to a single customer being in both control and treatment groups, invalidating results. Seller or region level randomization might be appropriate for specific experiments targeting those entities, but not for core Prime features affecting the customer experience directly.

3. Q3

   Choose primary and guardrail metrics for a Buy Box experiment aimed at improving purchase conversion rate. What metrics would prevent a harmful launch?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Include user experience, partner health, revenue, reliability, and long-term retention considerations.

   How to answer: For a Buy Box experiment focused on purchase conversion, the primary metric would be 'Purchase Conversion Rate' (units purchased / unique product page views or sessions). Guardrail metrics are crucial to prevent negative impacts. Key guardrails include 'Average Order Value' (AOV) to ensure we're not just driving micro-conversions, 'Revenue per User' or 'Gross Merchandise Volume' (GMV) to monitor overall financial impact, and 'Add-to-Cart Rate' as an upstream funnel health check. Additionally, 'Page Load Time' or 'Error Rate' are essential technical guardrails to ensure the user experience isn't degraded.

4. Q4

   During a Sponsored Products experiment, the treatment/control split is 52/48 instead of 50/50. How would you diagnose sample ratio mismatch?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Assume assignment logs, exposure logs, and eligibility filters may disagree.

   How to answer: A strong candidate would first define Sample Ratio Mismatch (SRM) and its implications for experiment validity. They would then outline a diagnostic process starting with a chi-squared test on the observed counts to quantify the deviation from the expected 50/50 split. Further investigation would involve checking the bucketing logic for bugs, examining pre-experiment metrics for existing differences, and analyzing the SRM across different dimensions like device type, region, or user segment to pinpoint the source of the imbalance.

5. Q5

   The Alexa Shopping experiment is trending positive after two days. A PM wants to stop early and launch. How do you handle peeking and sequential testing?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Discuss pre-specified stopping rules, alpha spending, business urgency, and risk.

   How to answer: Explain that peeking early invalidates the experiment's statistical significance due to increased Type I error rates. Discuss the importance of pre-determining sample size and experiment duration based on desired MDE, power, and significance level. Propose methods to address the PM's request while maintaining statistical rigor, such as using sequential testing methodologies (e.g., AGILE, SPRT, or always-valid p-values) if designed upfront, or emphasizing the need to wait for the predetermined duration to ensure valid results and avoid false positives. Highlight the potential business impact of launching a false positive.

6. Q6

   A new Fulfillment by Amazon feature shows a large week-1 lift in purchase conversion rate, but the effect fades by week 4. What could explain this and how would you design the test duration?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Discuss novelty, learning effects, seasonality, and durable impact.

   How to answer: The fading lift could be due to a novelty effect, where new users engage more initially but then revert to baseline behavior as the novelty wears off. Another explanation is a selection bias if the feature was rolled out to a specific segment that naturally has higher initial engagement. It could also be a short-term incentive driving initial purchases that isn't sustainable. To design the test duration, I would consider the typical customer lifecycle and purchase frequency for FBA products, aiming for at least 2-3 full purchase cycles or a period long enough to observe habit formation and potential novelty decay, typically 6-8 weeks, with a power analysis to ensure sufficient sample size for the desired MDE over the full duration.

7. Q7

   In a marketplace-like Retail Marketplace feature, treatment users may affect control users. How would network effects or interference bias the experiment?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Examples include seller supply, content inventory, delivery capacity, or pricing pressure.

   How to answer: Network effects or interference in a marketplace A/B test primarily bias the experiment by violating the Stable Unit Treatment Value Assumption (SUTVA). This leads to an underestimation or overestimation of the true treatment effect, depending on the nature of the interaction. Specifically, positive network effects (e.g., more buyers attract more sellers) would dilute the observed treatment effect in the treatment group, making it appear less impactful. Conversely, negative network effects (e.g., treatment users consuming limited inventory, leaving less for control) would make the treatment appear more negative or the control group perform worse than it would in isolation, leading to an overestimation of negative impact or an underestimation of positive impact. The observed difference between treatment and control would not accurately reflect the true causal effect of the feature.

8. Q8

   purchase conversion rate is a low-frequency event for Prime. How would you set up an experiment with enough power without waiting too long?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Discuss proxy metrics, variance reduction, larger samples, longer windows, and risk of metric gaming.

   How to answer: To address low-frequency events like Prime purchase conversion, focus on choosing a more sensitive, higher-frequency proxy metric that correlates strongly with the ultimate business goal. This could involve metrics like 'add to cart' rate, 'proceed to checkout' rate, or even engagement metrics like 'product page views per session' if a direct conversion proxy is too low. Additionally, consider increasing the sample size by expanding the experiment's reach or duration, or employing techniques like CUPED to reduce variance and boost statistical power. Finally, clearly define the minimum detectable effect (MDE) and ensure the chosen proxy metric allows for detecting a meaningful change within a reasonable timeframe.

9. Q9

   Design a geo or region-level experiment for Buy Box. When is this better than user-level randomization, and what are the analytical downsides?

   MediumStatistics & Experimentation RoundA/B TestingAmazon-specific

   Context: Use matched markets, pre-period balancing, spillover checks, and fewer experimental units.

   How to answer: A geo-level experiment for Buy Box involves randomizing entire geographic regions (e.g., states, cities, or postal codes) into treatment and control groups, rather than individual users. This approach is superior to user-level randomization when there's a high risk of network effects or contamination, such as when changes to the Buy Box algorithm for one user might indirectly affect the experience or behavior of other users in the same local marketplace (e.g., seller behavior, inventory availability). The primary analytical downsides include reduced statistical power due to fewer, larger experimental units, increased variance, and potential for confounding variables if regions are not perfectly balanced, requiring more sophisticated statistical methods like difference-in-differences or hierarchical modeling.

10. Q10

    The Sponsored Products experiment lifts purchase conversion rate overall, but only for new users and only in one category. How would you evaluate heterogeneous treatment effects?

    HardStatistics & Experimentation RoundA/B TestingAmazon-specific

    Context: Balance pre-planned segments with exploratory slicing and multiple testing risk.

    How to answer: To evaluate heterogeneous treatment effects (HTE), I would first define relevant subgroups based on user tenure (new vs. existing) and product category. Then, I would perform a stratified analysis, calculating the lift in purchase conversion rate and its statistical significance for each subgroup. I'd use interaction terms in a regression model (e.g., OLS or logistic regression) with user tenure and category as covariates to formally test for HTE. Finally, I would consider power implications for smaller subgroups and potential multiple comparisons issues, adjusting p-values if necessary.

11. Q11

    Treatment improves purchase conversion rate but worsens on-time delivery rate for Alexa Shopping. Walk through a launch recommendation

    HardStatistics & Experimentation RoundA/B TestingAmazon-specific

    Context: Make a decision under conflicting metrics and quantify tradeoffs for stakeholders.

    How to answer: A strong recommendation balances the trade-offs between increased revenue (purchase conversion) and potential customer dissatisfaction/churn (on-time delivery). The candidate should propose a phased rollout, starting with a small segment and closely monitoring both primary and secondary metrics. Key considerations include quantifying the monetary impact of both metrics, identifying potential causal links between the treatment and delivery issues, and exploring mitigation strategies for the negative impact before a full launch. The recommendation should be data-driven, risk-aware, and customer-centric, potentially involving further experimentation.

12. Q12

    How would you design ramp-up, holdback, and post-launch monitoring for a successful Fulfillment by Amazon A/B test?

    HardStatistics & Experimentation RoundA/B TestingAmazon-specific

    Context: Include ramp stages, persistent holdback, alert thresholds, rollback criteria, and owner accountability.

    How to answer: A strong candidate would outline a phased ramp-up strategy, starting with a small percentage (e.g., 1-5%) of traffic, closely monitoring key operational metrics (latency, error rates, fulfillment times) and business metrics (conversion, revenue) for anomalies. For holdback, they would propose reserving a small, stable control group (e.g., 1-2%) from the experiment for an extended period post-launch to detect long-term novelty effects or persistent issues. Post-launch monitoring involves establishing a comprehensive dashboard with both operational and business KPIs, setting up automated alerts for significant deviations, and scheduling regular reviews to ensure sustained performance and identify any regressions or unexpected interactions.

13. Q13

    Midway through the Retail Marketplace test, tracking for Buy Box changed. How would you decide whether the experiment results are still usable?

    HardStatistics & Experimentation RoundA/B TestingAmazon-specific

    Context: Compare instrumentation versions, affected traffic share, raw logs, and sensitivity analyses.

    How to answer: Assess the nature and timing of the Buy Box tracking change relative to the experiment's duration and key metrics. Determine if the change introduced bias, data loss, or altered user behavior for the Buy Box metric specifically. Analyze pre- and post-change data for the Buy Box metric in both control and treatment groups to identify any significant shifts or discrepancies. If the impact is localized to the Buy Box metric and doesn't affect primary success metrics or overall user flow, consider excluding or re-evaluating only that specific metric, potentially using a difference-in-differences approach if the change was uniform across groups. If the change is fundamental or affects primary metrics, the experiment may need to be stopped and re-launched.

14. Q14

    Two overlapping experiments on Prime both affect gross merchandise sales. How would you detect and manage interaction effects?

    HardStatistics & Experimentation RoundA/B TestingAmazon-specific

    Context: Discuss experiment registry, factorial design, exclusion rules, and interaction terms.

    How to answer: Detecting interaction effects requires a factorial experimental design where treatments from both experiments are combined (e.g., A1B1, A1B2, A2B1, A2B2). Analyze the sales metric across these combined groups using ANOVA or regression, looking for a statistically significant interaction term. If detected, quantify the magnitude and direction of the interaction. Management strategies include sequential rollout, creating a combined treatment, or segmenting the user base to avoid overlap.

15. Q15

    Amazon's Retail Marketplace revenue suddenly drops 10% week over week. Structure a business case to diagnose the issue and identify the most likely drivers

    MediumProduct Analytics & Business CaseBusiness CasesAmazon-specific

    Context: Consider traffic, conversion, pricing, mix, supply/inventory, outages, marketing, and seasonality.

    How to answer: A strong candidate would structure their diagnosis by first confirming the data's accuracy and scope (e.g., specific marketplace, region, product category). They would then segment the revenue drop by key dimensions like product category, seller type (FBA vs. FBM), customer segment, and traffic source to pinpoint the affected areas. Next, they would investigate potential causes across three main pillars: internal system issues (e.g., payment processing, site errors), external market factors (e.g., competitor actions, economic shifts), and operational changes (e.g., pricing algorithms, shipping policy changes). Finally, they would propose specific data points and dashboards to monitor for each hypothesis and outline a prioritization strategy.

Practice these with instant AI feedback in a live mock interview → Start a Amazon Analytics Engineer mock

Prepare 8-12 STAR stories mapped to Leadership Principles; expect a Bar Raiser; quantify impact

Indicative Analytics Engineer pay in India: ~₹9–40 LPA (role-level range, not a Amazon-specific figure).

Other roles at Amazon

Analytics Engineer interviews at other companies