Instacart's 2023 S-1 reported 2022 revenue of $2,551M, split $1,811M transaction and $740M advertising and other. You are the architect for the ad-serving path inside the shopping app. The interviewer asks which business KPI your next six months of work moves, and how you would prove it moved. Talk me through your answer.

What the interviewer is testing

Whether you can name a KPI you can actually move, and then say how the movement would be attributed to your change rather than to seasonality, a pricing experiment or a retailer signing on. Most candidates name a metric. Very few can describe the measurement that would survive a finance review.

The revenue split is the whole prompt. A business earning roughly 29% of revenue from advertising has a different architecture job from one earning all of it on transactions, and the published numbers say which one this is.

The clarifying questions that change the answer

• Which KPI is the ad business actually run on? Revenue per session, fill rate, advertiser retention and cost per incremental order are different targets with different owners. Ask before answering, because they conflict.
• What is the lag between my change and the KPI? Latency moves within hours. Advertiser retention is a quarterly measure, so a six-month project cannot be judged on it.
• Is there a holdout? If every user gets the change on the same day, the KPI moves and I cannot prove it was me.
• Who signs off that the benefit was real? If the answer is nobody, the benefit will be claimed in a deck and never checked.

A strong answer's arc

Name a chain, not a metric. Architecture change → an operational metric I own → a product metric with a short lag → the business KPI finance reports. For an ad-serving path:

1. The change: cut ad-decision latency inside the search response, and raise the share of search responses that carry a relevant sponsored slot from 68% to 85%.
2. The operational metric: ad-request timeout rate and decision p99. Mine, measured in minutes, moves the day I ship.
3. The product metric: sponsored impressions per session and click-through, with a lag of days.
4. The KPI: advertising revenue per session, measured as a 1% holdout against treatment over four weeks.

Then state the arithmetic that makes it worth doing. On a business with $740M of annual advertising revenue, a 1% lift is roughly $7M a year, which is what buys the six months — and say what you would need to see in week two to abandon it.

Common weak answers

• "GMV." Nothing an ad-serving engineer does moves gross merchandise value in a way that can be isolated from demand. Naming an unattributable KPI is the standard failure.
• "Latency, because latency affects conversion." True in general, worthless here: without a holdout, the elasticity is borrowed from someone else's study on someone else's funnel.
• "Revenue." No lag, no denominator, no counterfactual. Finance will ask which part of a $2.55B number you are claiming and the conversation ends.

What a strong answer adds

The second-order effects. More sponsored slots per session is a tax on the organic result set, so the holdout has to carry a guardrail metric — order completion, basket size, or reorder rate — and a pre-agreed threshold at which the lift is refused even though ad revenue rose. The architect who names the guardrail before the lift is the one who gets the next budget. Also name the instrumentation cost honestly: a holdout means a second serving path, a consistent bucketing key, and the discipline not to ship anything else into that surface for four weeks.