# Ambiguous Campaign Performance and Measurement

**Synthetic practice case — learner version.** All businesses, records, amounts and events below are fictional. They are not a real campaign, client, platform finding or assurance engagement. Currency units (CU) are invented accounting units, not a benchmark.

## Your decision

A fictional retailer sells desk organisers at CU100 per order. On 29 September 2026 its growth lead asks to move CU3,000 of a proposed CU12,000 next-fortnight budget from paid search to paid social. The reason given is that social's attributed ROAS increased. Finance has authorised no reallocation yet. The decision deadline is tomorrow; the analyst has only the following extracts. Prepare an evidence-based challenge and the smallest useful next check.

## Evidence A — platform summaries

Extracted 29 September, 09:00 UTC. P1 = 1–14 September; P2 = 15–28 September, inclusive. Spend is CU, excluding tax, on delivery dates. Each row is the relevant platform's own report; cross-channel credit may overlap. Clicks and impressions are supplied delivery counts. Attributed orders/revenue are report credit, not deduplicated business sales.

| Channel / period | Spend | Impressions | Clicks | Attributed orders | Attributed revenue |
|---|---:|---:|---:|---:|---:|
| Search P1 | 6,000 | 100,000 | 5,000 | 120 | 12,000 |
| Search P2 | 6,000 | 120,000 | 4,800 | 132 | 13,200 |
| Social P1 | 4,000 | 200,000 | 4,000 | 80 | 8,000 |
| Social P2 | 6,000 | 400,000 | 4,000 | 150 | 15,000 |

The search export is on an interaction-date basis with a seven-day click window in both periods. Social P1 uses seven-day click/one-day view; P2 uses 28-day click/seven-day view after an export-setting change. Both social exports use interaction dates. No matched-window re-export is supplied. Late credited conversions may still change recent values. These fictional settings describe this exercise, not any platform's default policy.

## Evidence B — analytics and order summaries

Analytics purchase-event extract, generated 29 September 09:30 UTC, uses event dates and session-channel grouping. It is not the same attribution surface as Evidence A. The analyst supplied the following counts: Search P1 112, Search P2 125; Social P1 65, Social P2 160. Order-level reconciliation and modelling/filter details were not supplied. Do not treat all purchase events as unique orders.

The CRM order export below was generated 29 September 10:00 UTC. It counts unique confirmed orders by order date, using the last captured source recorded at checkout. Other includes direct/unknown sources. Source assignment is not a causal measurement. Each order has CU100 gross booked value, before the refunds shown; no partial refunds occur in this fictional extract.

| CRM source / period | Confirmed orders | Refunds recorded by extract time | Gross booked value |
|---|---:|---:|---:|
| Search P1 | 100 | 10 | 10,000 |
| Social P1 | 60 | 6 | 6,000 |
| Other P1 | 40 | 4 | 4,000 |
| Search P2 | 110 | 2 | 11,000 |
| Social P2 | 65 | 2 | 6,500 |
| Other P2 | 45 | 1 | 4,500 |

Finance reports a preliminary 40% contribution rate on net sales before media, service overhead and fulfilment-capacity changes. Its calculation detail is absent. Refunds can arise up to 30 days after an order under this fictional business's supplied policy. Neither period is fully mature at the extract time; P2 is especially recent. No channel-level variable-cost or lifetime-value evidence is supplied.

## Evidence C — change log and actual limited QA

| Date | Supplied record |
|---|---|
| 15 September | Social spend allocation increased; broader placements introduced. Placement-level outcome export is missing. |
| 18 September | Purchase-tag release added a browser callback. Release note says deduplication was planned; no full release QA record is attached. |
| 20 September | Consent-banner text changed. No before/after acceptance-rate or event-coverage analysis is attached. |
| 22 September | A delivery-delay warning was added to one landing-page variant. Exposure and order effects have not been analysed. |
| 28 September | Controlled staging checkout T17 generated two browser purchase events and one server event for one test order; all three lacked a shared event identifier in the supplied debug trace. This is one staging test, not a production error-rate estimate. |

The creative inventory lists the same three social concepts in P1 and P2. Comparable frequency, reach, exposure by asset and outcome-quality data are absent. The media lead says “people are tired of the ads”; this is an assertion, not a supplied study. No randomised incrementality or controlled creative experiment was run in this scenario.

## Your working output

1. State the decision and which quantities can be compared or calculated safely. Show arithmetic and definitions.
2. Separate observed delivery, attributed outcomes, business records, hypotheses and unsupported causal claims.
3. Identify conflicts and plausible explanations, including what the staging check does and does not show.
4. Request a small, prioritised set of evidence that could change the decision; do not request every possible report.
5. Compare a hold, investigation and bounded test/change proposal within actual human authority. Address the deadline without pretending certainty.
6. Complete the Marketing Measurement and Experiment Decision Record with unanswered fields visible. Record the human decision as Pending.

Useful methods: AMO-P-000365; AMO-P-000254 for a fatigue branch, AMO-P-000226 for tracking checks, AMO-P-000227 for attribution scope, AMO-P-000192 for a proposed experiment, AMO-P-000059 only when an actual experiment has results. None executes a campaign or authorises spend.

There is no hidden correct budget allocation. Your task is to make a defensible next decision under uncertainty. Do not invent an experiment, causal lift, permission or profitability proof. Discuss your reasoning before consulting a separately controlled debrief.