Six ways marketplace measurement goes wrong before the numbers reach a seller

Marketplace seller software measurement mistakes usually begin before calculation: the decision, event or eligible population is vague. The remedy is an evidence record that another reviewer can reproduce, not a more elaborate chart.

Method: desk review of primary UK methodological, consumer, safety, privacy, tax and security sources. Research date: 6 September 2026. Scope: one fictional England seller, one authorised marketplace account and one approved product-stock-to-order journey. Inclusions: observable errors that can change a decision or conceal a non-compensating gate. Exclusions: prevalence estimates, vendor rankings and live-system tests. Ranking: non-ranked. Conflicts: none declared; no supplier or affiliate contributed.

1. Counting transmissions as rendered offers

An interface acknowledgement proves only its defined response. It does not establish what a customer could see. Keep source product evidence, transformation, transmission, marketplace response and inspected rendered state apart. GOV.UK online-selling guidance identifies information, price, ordering and confirmation subjects, so the inspected offer version matters.

2. Hiding eligibility inside the query

A result without a written population cannot be reproduced. Declare account, product set, event, period, Europe/London clock, exclusions and identity rule before running the query. The Government Data Quality Framework treats fitness for purpose, metadata and communicated limitations as important public-sector practices. Here it is a method reference, not a merchant standard.

3. Collapsing later order states

Placed, accepted, authorised, captured, fulfilled, cancelled, returned, refunded, disputed and settled are not synonyms. Combining them can make customer and money outcomes disappear. HMRC's VAT record guidance provides record categories for VAT-registered businesses; an accountant must approve the actual reconciliation and treatment.

4. Resolving duplicates by convenience

Taking the latest row can erase a correction or retain an out-of-order message. Store event, ingestion and correction times, key version and retry relationship. Route missing keys, late events and conflicting states to an exception owner. NCSC secure online-service guidance distinguishes logging, transaction monitoring and incident management, without certifying a particular implementation.

5. Treating privacy choice as data quality

A technically complete tracking stream may still conflict with a user's recorded choice; a privacy-respecting dataset may have deliberate gaps. Document both. The ICO's storage and access technologies guidance explains the PECR context for device storage and access. It does not say that permission makes a measure unbiased.

6. Averaging away a safety exception

A strong aggregate cannot compensate for an unresolved unsafe-product action. Preserve the product identifier, business role, instruction, evidence and suppression status in an urgent queue. OPSS product-safety advice explains that applicable duties depend on product and role. Specialist review owns the response.

7. Calling attribution causation

A provider-assigned order, rule-based credit or model association does not show what would have happened otherwise. A causal claim needs a credible counterfactual, defined exposure, outcome window, missingness assessment and uncertainty. HM Treasury's QPIE guidance is a public-policy evaluation reference, not proof about this merchant.

Apply the correction gate

For each error, record affected decisions, data versions, correction owner and rerun evidence. Keep thresholds blank. If the seller cannot reconstruct the eligible population or reconcile material states, withdraw the result and hold the related operational decision.