Measurement
Attribution measurement built for England around five separate jobs
Build attribution measurement for England with defined events, populations, data states, uncertainty, causal boundaries and decision-led reporting.
What to take away
- Defensible attribution starts with a written decision, a named owner and a review date.
- Five measurement jobs stay apart, and the table below names each one.
- Every metric card carries its decision, eligibility rule, counting entity and geography.
- Timezone, source versions and uncertainty belong on the same card.
- Thresholds stay deliberately blank pending the buyer's own decision, so the gap is method, not omission.
This guide covers an England operating context. Its privacy sources have UK scope, while government analytical standards are cited as transferable method rather than rules for private business. Research closed on 6 September 2026. No dataset or platform account was available, so every threshold stays blank by design.
Write the decision before the metric
Write the decision first: which choice is being made, whose call it is, and when it is reviewed. Then name the evidence that would change that choice. The business outcome must stay separate from the reporting proxy. An accepted sales record is not the same event as a platform-labelled conversion, even on one journey.
Teams need different views, so build separate reports rather than one overloaded headline. Blank thresholds stay blank until the buyer explains which decision they control. Attribution measurement mistakes, from missing populations to unmarked breaks lists the errors an unexamined metric card tends to hide.
Keep five measurement jobs apart
What it reports
- Platform credit
- Credit under a documented platform rule
- Cross-system reconciliation
- Agreement, conflict and missing links
- Descriptive analytics
- Events inside a collection boundary
- Marketing mix modelling
- Modelled association of aggregate inputs and outcomes
- Experimental analysis
- Difference against a designed comparison
What it cannot establish
- Platform credit
- Cross-platform totals or causal incrementality
- Cross-system reconciliation
- Why an outcome happened
- Descriptive analytics
- Marketing contribution
- Marketing mix modelling
- Truth outside its specification
- Experimental analysis
- Effects beyond the population tested
First-touch, last-touch and multi-touch rules belong in the first row when they assign credit under a stated rule. They are not smaller versions of marketing mix modelling. Never add platform credits together as though they shared one eligible population.
Five measurement jobs
Job
- Platform credit
- Credit under rule
- Reconciliation
- Agreement and gaps
- Descriptive analytics
- Observed events
- Marketing mix modelling
- Modelled association
- Experiment
- Difference vs comparison
Reports
- Platform credit
- Cross-platform totals
- Reconciliation
- Why outcome happened
- Descriptive analytics
- Marketing contribution
- Marketing mix modelling
- Universal truth
- Experiment
- Beyond tested population
Cannot establish
- Platform credit
- Reconciliation
- Descriptive analytics
- Marketing mix modelling
- Experiment
HM Treasury's Magenta Book separates monitoring an observed change from estimating an effect against a credible counterfactual. That distinction transfers to commercial work.
Example: a completed metric card
The card below is illustrative; none of its entries were measured.
Field / Completed entry (illustrative)
- Decision
- Whether to renew branded paid-search next quarter
- Owner
- Paid media lead, with a separate release reviewer
- Counting entity
- Accepted order, deduplicated on order reference
- Geography and clock
- England orders, cut at 23:59 Europe/London
- Eligible population
- Sessions from England in the period
- Metric and formula
- accepted orders / eligible sessions
- Uncertainty
- Tracking refusals, stated rather than adjusted
- Expiry
- Recheck after any source or model change
Where a denominator is missing, keep the number as a platform-scoped count and leave the rate blank. Blank formulas such as rate = eligible events / eligible population stay blank until both terms exist in approved data.
Reconcile records before interpreting them
- eligible, received and accepted
- rejected, duplicate or late
- missing, unmatched or refunded
- corrected or deleted, with the earlier value kept
Keep the original source value beside any standardised version. When two systems disagree, publish the disagreement and its treatment rather than editing one total to match the other. Record link keys, precedence rules and the effect of consent-state changes.
The Government Data Quality Framework treats quality as fitness for a user's purpose and separates completeness from accuracy. Both ideas transfer to commercial reporting.
Lay out the decision view in order
- State the question, owner, period, population and method class.
- Show source coverage and exception states.
- Present the metric with its uncertainty and comparison.
- Finish with the decision, action, review date and correction history.
Avoid traffic-light colour without accessible text and a declared threshold. The roles inside an attribution team explain who signs off a release and who reviews it afterwards.
Score attribution methods on identical fields
Assess candidates on the same fields: decision, required variation, data granularity, identity dependence and assumptions. A rule-based model is reproducible but arbitrary. Marketing mix modelling depends on its specification, and an experiment carries causal weight only inside its design.
Treat collection permission as a separate gate
The ICO's guidance on storage and access technologies applies to tags, pixels and similar mechanisms and asks for a purpose-specific assessment. Data minimisation rules ask that personal information stay adequate, relevant and limited to the stated purpose.
Do not narrow the eligible population after seeing who permitted collection. Report the coverage boundary and let qualified reviewers choose the lawful route. On iOS, App Tracking Transparency changes what is observable at all, so a ledger gap may be a device-level refusal.
Research benchmarks by admission, not appetite
No reliable England-wide attribution benchmark was identified for this guide. A candidate needs the same population, geography and period as your own data, plus a matching event definition and adjustment method.
Log each candidate as accepted, rejected or unresolved with the exact reason. If none passes, publish that the benchmark is not established, as the attribution benchmark question sets out.
Common questions
What should a metric card contain?
A decision, an owner, an eligibility rule, a counting entity, a geography rule, a period and a timezone. Add source versions, uncertainty, expiry and a correction route.
How do I reconcile two systems that disagree?
Build a state ledger at each join, keep both original values, and report the disagreement alongside its treatment.
Is consent needed before collecting attribution data in England?
Tags and pixels that store or read information on a device fall under UK PECR, so a purpose-specific assessment is needed. The lawful route depends on the processing.
In this guide
- Marketing attribution key metrics: a documentation frameworkA framework for marketing attribution key metrics, covering event definitions, eligible populations, platform credit, modelled estimates and decision rules.
- Designing an attribution dashboard with a decision strip and an exceptions panelDesign an attribution dashboard around a real decision, with metric definitions, exception states, uncertainty, corrections and accessible evidence.
- Attribution methods on one evidence grid, with causal language held in reserveCompare attribution methods fairly on one evidence grid covering inputs, assumptions, outputs, uncertainty, causal limits and failure conditions.
- Marketing attribution measurement mistakes and controlsMarketing attribution measurement mistakes across GA4, Google Ads and Meta Ads: controls for credit, coverage, consent, data changes and reporting.
- Six tests before a marketing attribution benchmark is usedSix admission tests decide whether an outside figure can serve as a marketing attribution benchmark for England, and an internal series if none passes.
