A marketing analytics report for a budget decision: an annotated one-page example

1. Define the decision before selecting the metrics

Imagine a small online homewares retailer with a $300,000 paid-media budget for a four-week period: $180,000 for search and $120,000 for social. An additional $50,000 is available but uncommitted. The CMO must decide whether to add it to social for the next comparable period. Finance wants positive incremental contribution after media cost; spending the full reserve is optional.

Here, finance has agreed a 50% contribution margin after product and variable fulfilment costs but before advertising. Revenue means net sales, excluding tax and after discounts and returns. Additional fixed costs and future repeat purchases are excluded. For this illustration, the same margin applies to incremental net sales. Those are assumptions to check, not universal accounting definitions.

The report compares two equal four-week periods, not two calendar months of different length. The proposed decision meeting is 5 October 2026, with a formal review on 19 October. These are example business dates, not this article's publication date. The CMO owns the spend decision; the analytics lead owns evidence quality; finance owns the margin definition.

2. The one-page report

The following is the decision page. The annotations and calculation that follow are its supporting detail. Values are rounded for reading; calculations use unrounded numbers.

Next evidence: by 9 October, the analytics lead supplies a design and feasibility assessment for current social spend versus a controlled increase, using all-channel net sales. At the 19 October review, the CMO records the chosen action or a specific remaining blocker. The report does not promise a completed experiment by then.

3. Read the rows as different answers

The sales row is the business outcome. It is deliberately not assigned to a channel. The $100,000 increase could include changes in demand, promotions, availability, customer mix, or media. This example has no causal design for that period-over-period comparison, so the report makes no claim about how much marketing caused.

The attribution row describes credited revenue. Google Analytics defines attribution as allocating credit across touchpoints. Some attribution models use counterfactual methods, so it would be inaccurate to dismiss all attribution as simple last-click counting. The operational question remains: does this particular estimate represent the proposed spend change, outcome, population, and period?[1]

The example's platform credit figures are inputs, not exports from a real account. Their lookback and view-through settings would need to travel with a live report. Do not add the $540,000 and $480,000 and call the sum unique marketing-generated revenue: this example has no cross-platform deduplication. A numerical total below finance's sales total would not establish that overlap is absent.

The experiment row concerns a defined intervention. Incremental revenue return divides estimated additional revenue by the treatment-control spend difference. Google Ads' geography-based lift documentation uses that same ratio for incremental conversion value and cost. Here, the outcome is explicitly net sales rather than an arbitrary conversion value.[2]

The model-based planning row records that no modeled estimate is available. A real MMM can estimate causal effects under assumptions. Predictive fit alone cannot establish that its channel effects are accurate. Any modeled recommendation belongs beside its specification, diagnostics, and causal assumptions, not in a row labelled simply 'truth'.[8]

4. Make the uncertainty inspectable

For the arithmetic illustration, stipulate ten comparable market pairs, with one geography in each pair randomly assigned to social on and the other to social off. Each pair has an exact $5,000 spend difference over the test period. Assume no spillover, no other treatment changes, and independent, approximately normal paired revenue differences. These are simplifying assumptions, not findings established from the constructed data.

The ten hand-built treated-minus-control net-sales differences are −$2,000, $0, $2,000, $4,000, $6,000, $8,000, $10,000, $12,000, $14,000, and $16,000. Their sum is $70,000 and their mean is $7,000 per pair. Because spend differences are fixed and equal here, dividing the mean difference by $5,000 equals dividing the total difference by $50,000.

The sample standard deviation is $6,055.30; the standard error of the mean is $1,914.85. Applying the standard paired-difference calculation and a two-sided 90% Student t interval with nine degrees of freedom gives $7,000 ± 1.8331 × $1,914.85 per pair. Dividing the endpoints by $5,000 yields 0.698x to 2.102x, displayed as 0.70x to 2.10x.[3][4]

Real geo experiments can have few units, unequal market sizes, heavy-tailed revenue, and strong time variation. The analysis needs to account for those design problems. The ten-pair t calculation is included to make the arithmetic easy to follow. A live analysis needs an estimator and validation suited to its randomization and data.[5]

The apparent tension is useful. Under the teaching assumptions, the interval is above zero, yet it still crosses the 2.0x profitability hurdle. Evidence of additional revenue is not sufficient to establish profitable additional revenue.

5. Convert the result into the decision's economics

With a 50% contribution margin before advertising, every $1 of additional revenue supplies $0.50 toward media cost. Break-even incremental revenue return is therefore 1 / 0.50 = 2.0x. The $70,000 lift contributes $35,000 before media and −$15,000 after the $50,000 spend difference.

Transforming the unrounded revenue interval with the same fixed 50% margin gives approximately −$32,551 to +$2,551 after advertising. That interval includes sampling uncertainty from the teaching calculation only. It excludes uncertainty in margin, returns, repeat purchases, market transfer, and the next period's media prices.

A decision rule also needs the cost of delay. Waiting could miss profitable growth; maintaining current spend could continue a loss. The example permits a short, reversible hold because the reserve is optional and the team has a near-term review. The short hold depends on those conditions; a different cost of waiting could call for a different action.

6. Test the proposed change, not an easier substitute

The existing test compares social on with social off. The proposed action adds $50,000 to a $120,000 run rate, a 41.7% increase. Those are different interventions. The test does not estimate the return on that increase, and multiplying 1.40x by the new budget would manufacture a forecast.

Meridian distinguishes average ROI from marginal ROI, which concerns a small spend change. For this sizeable increase, the useful modeled quantity would be the finite difference between response at $170,000 and $120,000, rather than the local slope alone. Hold other channels and the intended geographic and time allocation comparable, then compare that additional revenue with the $100,000 needed to cover $50,000 of media at the stipulated margin. Response curves also carry extrapolation and lag assumptions that need to be checked.[7]

An experiment and an MMM may legitimately disagree because they cover different periods, populations, campaign scope, or counterfactuals. Meridian's calibration guidance explicitly warns that transferring experimental evidence into a model adds uncertainty. An on/off experiment should not be relabelled as evidence of an increase at the current run rate.[6]

For this example, the next design should compare the current policy with the proposed higher-spend policy at comparable local intensity. It should use all-channel net sales so that purchases moving between touchpoints are not mistaken for new business. Before launch, the team needs the following:

• A clearly defined population and spend contrast. Representative test markets should reflect the audience and auction conditions to which the decision will apply; material limitations must be recorded.
• A feasibility analysis using actual pre-period variation. Determine whether the design can distinguish a commercially useful return from the 2.0x break-even hurdle, then choose duration and scale. The toy's ten pairs do not supply a power calculation.
• A fixed primary outcome, measurement window, return treatment, estimator, and analysis date. Include enough follow-up for the business's purchase cycle; do not select a winning observation window after seeing results.
• An agreed risk budget and action rule. State the largest acceptable test loss, the role of uncertainty, and what findings would justify expanding, reducing, or retaining spend. A safety stop is different from repeatedly checking for a favourable result.

If a credible test is too expensive or slow, the decision still belongs to the CMO. The report can compare a smaller reversible change, a model-informed estimate with explicit assumptions, and waiting. It should explain what each option would cost and what evidence is available.

7. A reusable blank decision page

Keep these fields on the decision page and put implementation detail in a linked appendix. The blank fields below are an intentional reusable template, not unfinished article copy.

At review, retain the original recommendation and record what changed. If the evidence remains inconclusive, name the consequence and the next choice. Replacing last month's slide with this month's numbers erases the judgment the team needs to learn from.

8. What this example establishes

The decision page brings together the budget question, evidence, economics, recommendation, and next review date.

The sources below explain the methods and their limits. The accompanying calculation lets you check the arithmetic.

The example intentionally leaves out long-run brand effects, customer lifetime value, changing marginal costs, and complex channel interactions. Where those matter, revise the outcome and time horizon before using the same report structure. The central discipline remains: state the quantity the decision needs, show what the evidence actually estimates, and make the remaining judgment visible.