Measurement
Part of Attribution measurement built for England around five separate jobs
Marketing attribution measurement mistakes and controls
Marketing attribution measurement mistakes across GA4, Google Ads and Meta Ads: controls for credit, coverage, consent, data changes and reporting.
Marketing attribution measurement mistakes usually become visible when a decision-maker asks what the number includes. The remedy is a complete evidence record, not another unexplained metric.
Marketing attribution reporting in Google Analytics 4 (GA4), Google Ads, Meta Ads and Adobe Analytics can differ because platforms use different attribution models, conversion definitions, identity signals and reporting windows. Treat each platform's output as a scoped view, not a mutually exclusive count; keep the platform, account, conversion action, model, click-through and view-through windows, consent state and extraction date with every comparison. The controls below help identify where those choices make results incomplete or incomparable.
What to take away
- A numerator alone cannot show a rate, so record the eligible denominator and exclusions.
- First-touch, last-touch and multi-touch rules allocate credit under a convention, not causal evidence.
- Adding platform totals can double count the same outcome and conceal missing coverage.
- A single clean total can erase rejected rows, delayed imports, refunds and later corrections.
- A new event trigger, identity rule or source schema can make periods incomparable without a marked break.
Reporting a rate without its eligible population
A numerator alone cannot show a rate. Record the event, entity, eligible denominator, exclusions, geography, period and Europe/London timezone. If eligibility cannot be reconstructed, publish a bounded count or HOLD.
The Government Data Quality Framework is public-sector guidance that separates completeness from accuracy and asks teams to assess quality for user need. It does not approve a commercial metric.
Treating a credit rule as causal evidence
First-touch, last-touch, last non-direct click, data-driven attribution and position-based models allocate observed credit under a convention. A platform label does not establish what would have happened without the marketing activity. For each platform and conversion action, preserve the selected model, configuration, account scope, click-through and view-through windows, and extraction version beside the result.
HM Treasury's QPIE guidance explains in government evaluation why monitoring change differs from estimating cause against a credible counterfactual.
To assess whether marketing caused additional outcomes, use an incrementality design where feasible rather than treating attributed conversions as the counterfactual. Geo holdouts and lift tests compare outcomes with a control condition; document the design, eligible population and limitations alongside the platform attribution report.
Adding incompatible Google Ads and Meta Ads totals
Google Ads and Meta Ads may report overlapping outcomes through different identities, click-through and view-through windows, and eligible populations. For example, the same purchase may appear as an attributed conversion in both platforms even though the underlying order record contains one purchase. Walled-garden reporting does not necessarily provide a shared event-level ledger for reconciling those claims. Reconcile against underlying records and report conflicts before comparing channels.
The Office for Statistics Regulation's quality standard governs official statistics, but its requirement to explain representativeness and comparability gives a useful admission test for combined data.
Hiding late, missing and corrected states
A single clean total can erase rejected rows, delayed imports, unmatched identifiers, refunds and later corrections. Give each state a definition and show its eligible population. Where Google Analytics 4 or Google Ads reports modelled conversions, identify those estimates separately from directly observed event records where the reporting view allows it; do not present them as reconciled transactions. A missing record is not zero, and a refund is not merely a negative conversion without an approved accounting rule.
Government guidance on communicating quality, uncertainty and change targets official statistics, while explicitly noting possible value for other authors of numbers. It supports making limitations visible to users.
Changing definitions without marking a break
A new event trigger, identity rule, source schema, query, attribution model or lookback window can make periods incomparable. Preserve versions, state the effective time and recalculate only under a declared method. Do not splice old and new values into an uninterrupted trend line.
The Government Analysis Function's RAP strategy promotes version control, input validation and documented manual steps for government analysis.
Collecting extra personal information for convenience
More identifiers may improve matching inside a chosen system, but that operational benefit does not settle necessity or permission. Record the consent signals used for measurement, including Google Consent Mode where applicable, and relevant iOS limits such as Apple's App Tracking Transparency (ATT). These affect what can be observed or matched and do not, by themselves, establish that collection is necessary or permitted. The ICO's data minimisation guidance says personal information should be limited to what the specified purpose needs. Qualified reviewers must assess the actual design.
Publishing an inaccessible or uncorrectable view
A colour-only signal, unlabeled chart or unusable export can block scrutiny. GOV.UK's assistive-technology testing guidance is for government services, yet it demonstrates the need for task-based checking beyond automated tests.
Attach a correction route and expiry to every release. If a reader cannot find the definition, coverage, uncertainty or superseded version, the decision view stays on HOLD.
Before you act
- Record the platform, account and conversion action alongside the eligible denominator, exclusions, geography, period and timezone.
- Preserve the credit rule, attribution model, click-through and view-through windows, configuration, account scope and extraction version.
- Reconcile underlying records across platforms, including any modelled conversions, and report conflicts before comparing channels.
- Record relevant consent signals and identity or measurement limits, including Google Consent Mode and iOS ATT where applicable.
- Give each data state a definition and show its eligible population.
- Preserve versions and state the effective time of any definition change.
- Attach a correction route and expiry to every release.
Common questions
Why is a numerator alone not enough to report a rate?
A numerator alone cannot show a rate. The article says to record the event, entity, eligible denominator, exclusions, geography, period and Europe/London timezone. If eligibility cannot be reconstructed, publish a bounded count or HOLD instead of an unexplained metric.
Why can adding platform totals be misleading?
Platforms may observe overlapping events through different identities, lookback rules and eligible populations. Adding their credits can double count the same outcome and conceal missing coverage. The article says to reconcile underlying records and report conflicts before comparing channels.
What should happen when a definition changes?
A new event trigger, identity rule, source schema, query or model can make periods incomparable. The article says to preserve versions, state the effective time and recalculate only under a declared method. Do not splice old and new values into an uninterrupted trend line.
