# A small aggregate improvement hides a weaker prospecting campaign

Fictional demonstration. Baseline: 1–14 August 2026. Comparison: 15–28 August 2026. Source: the actual offline run in `output.json`. This interpretation was drafted; no reporting model was executed.

| Metric | Baseline | Comparison | Change |
| --- | ---: | ---: | ---: |
| Spend | €6,000 | €6,520 | +8.67% |
| Attributed gross revenue | €31,600 | €34,880 | +10.38% |
| Conversions | 395 | 436 | +10.38% |
| CPA | €15.19 | €14.95 | −1.55% |
| ROAS | 5.27× | 5.35× | +1.58% |
| CTR | 2.143% | 2.029% | −0.114 percentage points |

Nonbrand search contributes the largest revenue increase: +€3,520, with 100 → 144 conversions. Its CPA improves from €20.00 to €15.28 and its ROAS from 4.00× to 5.24×. Both periods have a 4.00% CTR, while conversion rate rises from 2.50% to 3.00%. These observations support investigating the source of stronger post-click performance; they do not identify its cause.

Social prospecting requires attention. Spend rises from €1,800 to €2,100, while conversions fall from 45 to 35 and revenue falls from €3,600 to €2,800. CPA worsens from €40 to €60. CTR drops from 1.00% to 0.80%, and conversion rate drops from 1.50% to 1.25%. Those two declines justify checking creative, audience delivery, landing-page changes and tracking before attributing the result to any one factor.

Retargeting generates 99 conversions versus 100 with 10% less spend, improving CPA from €10.00 to €9.09. Brand search still has the highest reported ROAS, but it falls from 10.00× to 9.58×. Brand and retargeting audiences can contain existing demand, so their high reported ROAS should not be treated as proof of incremental value.

## Proposed decisions

1. Hold further prospecting budget increases while checking event tracking and recent creative/audience changes. Draft a controlled creative test once measurement is confirmed.
2. If nonbrand search has available relevant demand and sufficient margin, propose a limited 10% budget increase for a monitored test. This is a proposed test size, not a forecast or an executed budget change. Monitor CPA, query quality and conversion lag.
3. Keep retargeting stable while reviewing audience size and frequency. The dataset contains neither, so it cannot justify unrestricted scaling.

## Limits and validation

Only two short aggregate periods are present. There are no order-level records, confidence intervals, attribution windows, margins, refunds, audience overlap, conversion lag, seasonality controls, or experimental holdouts. Revenue per conversion is fixed at €80 in this deliberately simplified dataset. Reported revenue and conversion growth therefore move together. ROAS is not profit, causal lift, or promised business return.

The valid CSV passed 7 recorded execution checks. The deliberately corrupt CSV was rejected with exit code 2: one row has more clicks than impressions, and another has negative revenue. The resulting coverage errors are secondary consequences of rejecting the malformed row. See `verification.json` and `invalid-output.json` for the exact evidence.