Articles

Marketing Analytics Case Study Interview: Turn Conflicting Channel Metrics into a Budget Decision

Practice a marketing analytics case with conflicting channel metrics, a reconciled revenue table, incremental-lift questions, and a defensible budget memo.

The paid-search dashboard says one thing, the social platform says another, and finance trusts neither. In a marketing analytics case study interview, the useful answer is not to pick the prettiest return-on-ad-spend number. It is to define what each system measures, reconcile the money that can be reconciled, identify what remains causal uncertainty, and make a budget decision with a testable condition.

This worked case gives you a small, original dataset and a decision memo to rehearse. For a related prompt, try Determine Optimal Budget Allocation for Maximum Profit, then explain why platform-attributed revenue alone cannot determine the optimal allocation.

Evidence boundary: All company, channel, and dollar figures below are invented for interview practice. Official Google Analytics and Google Ads documentation supports the measurement distinctions; it does not endorse this toy budget or prove incrementality for a real advertiser. No specific employer's interview format is implied.

The case brief: three reports, one next-month budget

You advise a subscription retailer with an $80,000 monthly acquisition budget. The growth lead wants to add $10,000 to social because its platform shows $105,000 in attributed revenue. Finance says the internal order ledger is lower and asks whether the extra spend would create profitable orders. You have 30 minutes to recommend the next-month plan. Assume a 45% contribution margin on net booked revenue before media cost for this exercise; it is a simplifying assumption, not a channel-specific margin estimate.

The candidate should answer six urgent questions: Are the reports even comparable? What is the finance-reconciled revenue? Which orders are duplicated or refunded? Does attributed revenue indicate incremental return? What decision can be made now? Which observation would change that decision? State these questions before calculating so the analysis has a destination.

Channel	Spend	Platform-attributed revenue	Web-analytics attributed revenue	One-touch ledger assignment	Refunds in assigned orders	Net assigned orders
Paid search	$40,000	$120,000	$96,000	$90,000	$5,000	$85,000
Paid social	$30,000	$105,000	$70,000	$65,000	$8,000	$57,000
Email	$10,000	$65,000	$60,000	$52,000	$2,000	$50,000
Total	$80,000	$290,000	$226,000	$207,000	$15,000	$192,000

The ledger assignment is intentionally a one-touch management view: each eligible order is assigned once, then its refund is removed. The channel totals therefore add to the exercise's $192,000 net booked revenue. The platform columns can credit the same customer journey differently and should not be added or interpreted as independent sales. In a real case you would first confirm whether email spend belongs in the same acquisition budget and whether the 45% margin applies after returns and fulfillment. Do not silently make those assumptions for the interviewer.

Reconcile definitions before ranking channels

Start with a metric contract, not a chart. Ask for each report's date window and time zone, order versus conversion-event grain, gross versus net value, currency, attribution model, lookback window, customer and order deduplication, refund treatment, tax and shipping inclusion, and cross-device identity. Google Ads documentation distinguishes conversion counting choices such as “one” and “every” and describes transaction-ID deduplication. Google Analytics attribution and conversion settings likewise affect which channels receive credit. Those are plausible sources of mismatch, not proof that any particular row here is wrong.

Build a reconciliation at the order ID grain. Start with paid, non-test, eligible orders in finance. Remove canceled orders and refund amounts according to the stated recognition policy. Join each order to its web event and platform export where an ID exists. Mark missing joins rather than dropping them. Then group by event date and purchase date separately to see whether latency explains part of the gap. If the platform has only aggregate metrics, record that limitation; do not manufacture a row-level bridge.

A useful reconciliation output has columns for order ID, amount, purchase timestamp, refund timestamp, channel assignment, analytics transaction ID, platform conversion IDs, and inclusion reason. If two conversion events reference one order, count the order once in the ledger. If an order has no click record, it remains a finance order; the lack of attribution is not evidence that the order did not happen. Conversely, a click or conversion event with no settled order is not automatically booked revenue.

For this exercise, the platform total exceeds net ledger revenue by $98,000, while the analytics total exceeds it by $34,000. Those differences are diagnostic gaps, not fraud or lost cash. Platform-reported return ratios are 3.00 for search, 3.50 for social, and 6.50 for email. Net one-touch revenue divided by spend is 2.13, 1.90, and 5.00 respectively. Neither set is incremental ROAS. Email's high ratio may partly reflect existing customers who would purchase anyway; social's lower ledger ratio may reflect a different attribution rule. The next question is causal, not a new arithmetic ranking.

Translate the ledger into contribution, without overstating causality

The exercise's net assigned revenue is $192,000. Applying the assumed 45% margin yields $86,400 contribution before media; subtracting $80,000 media spend leaves $6,400 descriptive contribution. This calculation checks whether the portfolio is close to a contribution threshold under the stated assumptions. It is not evidence that marketing created all $192,000. Some customers would have purchased without these impressions, and the channel assignment allocates orders rather than estimating that counterfactual.

At the channel level, the same rough exercise would produce $38,250 search contribution before its $40,000 spend, $25,650 social contribution before $30,000 spend, and $22,500 email contribution before $10,000 spend. It would be a mistake to cut search immediately because the first difference is negative: the uniform margin and one-touch assignment may be wrong, and a channel's marginal effect can differ from its average attributed ratio. It would also be a mistake to double email because its assigned contribution looks largest. This is exactly the distinction an interviewer wants to hear.

The incremental question is: if we spend one more dollar on a channel, how much additional net contribution occurs relative to not spending it, after spillover and displacement? A controlled holdout, geo experiment, or carefully designed lift test can answer a scoped version. For large multi-channel planning, a marketing-mix model may help, but it brings assumptions and uncertainty; Google's Meridian materials explicitly discuss calibrating with experiment evidence and scenario planning. Do not present a model as a replacement for measurement discipline.

Produce a decision memo, not a dashboard tour

Here is a defensible one-page answer for the toy case:

Decision for next month: Hold the $80,000 base budget while validating the order bridge. Do not approve the requested $10,000 permanent social expansion from the platform ratio alone. If a separate $10,000 test budget is available, allocate it to a pre-registered social holdout or geo test with enough traffic to detect a decision-relevant lift. If that budget must come from the existing $80,000, protect core campaigns and negotiate a smaller pilot with an explicit opportunity cost.

Why: The systems disagree materially, but the internal net-order ledger is reconcilable. Its $6,400 portfolio contribution is descriptive under a 45% assumed margin. No row yet identifies social's incremental return. The case also lacks a marginal response curve, customer lifetime value by cohort, and capacity constraints.

Decision gate: Before launch, define the eligible audience, randomization unit, exposure and purchase windows, net-contribution outcome, minimum detectable effect, contamination checks, and stop rule. Expand only if a credible estimate of incremental contribution after added spend clears the business hurdle with acceptable uncertainty. If the test is underpowered, report that limitation rather than declaring no effect.

That is a budget decision, even though it is conditional. “Need more data” alone is not one. The memo says what will be funded, what will not, why, and what evidence can change the answer. The exact test size and threshold cannot be set responsibly from this toy table; the candidate should ask for baseline conversion, variance, unit economics, and available geographies or users.

How to investigate a conflicting number under interview pressure

Use a short sequence. First, anchor the denominator and period: same currency, same date range, same recognized order population. Second, inspect the biggest gap, not every dashboard setting. Here social's $105,000 platform figure versus $57,000 net assigned ledger figure is the largest proportional controversy for the requested increment. Third, request the few exports needed to distinguish causes: order IDs and refunds, conversion IDs and timestamps, click/impression records, campaign cost, and attribution settings. Fourth, calculate sensitivity to reasonable alternative rules. Fifth, separate bookkeeping certainty from causal uncertainty in the recommendation.

Consider a delayed refund. A platform report may show a conversion in the click month, while finance later recognizes a partial refund. Neither system is necessarily “broken”; they answer different questions at different times. A robust analysis specifies whether the budget decision uses gross orders, net orders observed through a fixed lag, or projected ultimate net value. If recent cohorts have not had time to refund, compare equivalent maturity windows or flag a provisional estimate.

Consider an order touched by email and search. The web analytics report may allocate credit fractionally, the ad platform may take credit based on its own eligible touch, and a one-touch ledger can assign the same order to one channel for reporting. The $290,000 platform sum is therefore not a cash total. Show one illustrative order across the systems to make this concrete. Do not try to “correct” an attribution rule until the business question is stated.

Finally, check whether spend and revenue are both net of taxes, credits, coupons, agency fees, and shared campaign cost. A 3.5 platform ROAS can be numerically correct within its own definition while useless for an incremental profit decision. The interview strength is in making that boundary visible without becoming paralyzed by imperfect data.

What would change your recommendation?

Suppose the order bridge finds that social conversions were duplicated because one purchase emitted both a web and server event without reliable transaction IDs. Then fix instrumentation and backfill the comparable series before ranking. Suppose instead the bridge is clean but a randomized test estimates a positive incremental contribution beyond the added spend. Then increase social within the measured audience and monitor saturation, with a holdout or staggered rollout. Suppose the test shows little lift or confidence remains too wide to exclude an uneconomic result. Then maintain or reduce the experimental allocation, while improving sample design rather than proclaiming all social advertising ineffective.

Also ask what happens if margin differs by channel or customer cohort. If social brings higher-retention customers, a first-order margin view understates value. But lifetime-value assumptions must be based on mature cohorts and net of retention costs; speculative LTV should not be inserted to rescue a weak short-run result. If search mainly captures brand demand, its attributed revenue can be high while incremental lift is small. A brand holdout can test that mechanism.

In a live interview, distinguish a decision from a promise. You can recommend a bounded experiment today, not guarantee the result. You can set a pre-specified threshold and responsible owner. You cannot infer marginal return from three aggregate channel rows.

A 90-second answer you can adapt

“I would first make the reports comparable at order, date, currency, and refund grain. The platforms sum to $290,000, but the one-touch internal ledger nets to $192,000 after $15,000 of refunds; I would not sum platform credit as booked revenue. At the assumed 45% margin, the ledger implies $86,400 before media and $6,400 after the $80,000 spend, but that is descriptive, not an incremental return. I would hold the base allocation, decline an untested permanent $10,000 social increase, and fund a bounded lift test if the opportunity cost is acceptable. I would pre-register net incremental contribution and expand only if the estimated marginal benefit clears the hurdle with enough precision. Meanwhile, I would reconcile IDs, windows, and settings so future decisions use one auditable order bridge.”

Practice saying that without reading the table. Then change one assumption—margin, refund lag, or the experimental estimate—and explain whether the decision changes. That ability to revise a budget when the evidence changes is stronger than a polished dashboard narration. Start with Determine Optimal Budget Allocation for Maximum Profit and defend the incremental objective.

Sources and Further Reading

Research checked September 23, 2026. Platform settings and model guidance can change; verify the current configuration for the advertiser in front of you.