Comparing marketplace attribution methods on one population and one window

Marketplace seller software attribution methods answer different questions. Observation describes a record, a rule assigns credit, and causal evaluation estimates what changed because of an intervention. Comparing them requires the same eligible seller records, order-state outcome, Europe/London window and unit. There is no universal winner.

Research for this comparison ended on 6 September 2026. No live marketplace record entered the work, and no effect has been estimated.

Set common comparison fields

For every route, state the decision, exact product or order event, eligible population, numerator and denominator, maturity window, identity rule, late and missing treatment, exclusions, source/query/version, privacy boundary and uncertainty. Keep order acceptance, payment authorisation, capture, fulfilment, refund, dispute and settlement separate.

The ICO's storage and access technologies guidance provides the current PECR context for storing or accessing information on user equipment. Permission and statistical validity are independent questions.

Compare the evidence routes

Route	Output	Main assumption	Proper use	Limitation
Direct observation	Defined records and states	Evidence captures the eligible event	Reconciliation and operations	Says what was recorded, not why
Declared source	Marketplace, supplier or participant statement	Source and definition are accurately preserved	Labelled context or status	Declaration may be incomplete or self-reported
Deterministic rule	Credit assigned by a buyer-written rule	Convention fits the decision	Repeatable allocation	Rule is not an estimated effect
Supplier or marketplace attribution	Output under the provider's current method	Method, eligibility and identity are understood	Bounded platform reporting	Opaque changes and unavailable counterfactual
Matched observational analysis	Difference among matched records	Measured matching fields address relevant differences	Exploratory adjusted comparison	Unmeasured confounding can remain
Statistical association	Modelled relationship	Model form and included variables are adequate	Description and hypothesis formation	Association alone does not establish cause
Experiment or credible quasi-experiment	Estimated effect for a defined exposure	Counterfactual and validity conditions hold	Causal question within studied boundary	Contamination, missingness and limited transportability

Demand more for causal language

Predeclare eligibility, allocation or identifying event, actual exposure, comparison, primary outcome and later-state window. Record non-compliance, contamination, missing outcomes, exclusions, stopping logic and uncertainty. A short change in accepted orders cannot stand for settled contribution or durable customer benefit.

HM Treasury's Magenta Book is guidance for central government evaluation, not a merchant compliance standard. Its impact-evaluation principles are useful when testing whether a design has a credible counterfactual. The companion QPIE guidance offers a structured way to examine evaluation quality. A qualified statistician must decide whether those ideas fit this case.

Protect transaction meaning

Do not label captured amounts as revenue, or marketplace-attributed order value as incremental profit. Reconcile refunds, disputes, settlement, costs and accounting treatment first. HMRC's VAT records guidance supplies recordkeeping context for VAT-registered businesses, not a transaction-specific tax conclusion.

Consumer, safety, privacy, accessibility and security gates also sit outside the comparison. An attractive attributed number cannot justify misleading price presentation, ignored tracking choice, unsafe product handling or an inaccessible journey.

Choose by question, not by apparent precision

Use observation for state control, a transparent rule for administrative allocation, and properly caveated supplier attribution for the supplier-defined view. Use adjusted analysis for exploration, not causal certainty. Reserve causal wording for a reviewed design with defensible comparison evidence. If population, state or identity cannot be aligned, mark the methods INCOMPARABLE and stop rather than forcing a common result.