Measurement
Part of Attribution measurement built for England around five separate jobs
Attribution methods on one evidence grid, with causal language held in reserve
Compare attribution methods fairly on one evidence grid covering inputs, assumptions, outputs, uncertainty, causal limits and failure conditions.
Mistake: Calling allocated credit incremental return
The highest-cost error is to present allocated credit as incremental return. First-touch, last-touch and multi-touch rules divide observed credit according to their definitions; none compares observed outcomes with a no-marketing outcome. Label each output rule-based credit, and use causal language only when the supporting design can justify it.
Those allocations can differ for the same observed sequence because event definitions, eligible populations, identity scope and configuration vary. Reproducibility makes a rule checkable, but does not make it causal. Keep AI attribution claims separate from causal findings, and record the measurement and reporting method used.
Record Before Comparing Credit Rules
- Event sequence
- Eligible journey
- Lookback rule
- Identity scope
- Configuration and version
- Do not sum cross-platform credits
Mistake: Combining platform totals before reconciling them
Cross-platform totals can mislead when sources differ on identities, eligible populations or event definitions. Reconcile matched, unmatched and conflicting records before comparison; do not combine separate platform allocations into a channel-wide measure. Overlapping records and changing source definitions can change what is counted.
Keep the event sequence, lookback rule, identity scope, configuration and version for each rule. Check the relevant measurement job before comparing outputs. Where no public figure is available, treat lookback ranges as typical planning ranges rather than universal settings, and keep any boundary between sources visible.
Mistake: Treating record agreement as evidence of an effect
A reconciliation result can establish agreement, disagreement or gaps between source records, but it cannot explain why an outcome occurred. Report matched, unmatched and conflicting states plainly, and keep them separate from descriptive counts and causal estimates. The record describes alignment, not an effect.
| Evidence job | Output it permits | Warning sign and repair |
|---|---|---|
| Platform credit rule | Credit under a stated rule and version | Unknown eligibility or duplicate treatment: document the scope and rule |
| Cross-system reconciliation | Matched, unmatched and conflicting states | Unresolved collisions or shifting source definitions: show the unresolved states |
| Descriptive analytics | Counts, rates or journeys within collection coverage | Missing denominator or coverage gap: disclose the boundary |
| Marketing mix modelling | Modelled associations and uncertainty | Unstable estimates or weak variation: inspect the specification and diagnostics |
| Experiment or quasi-experiment | Estimated effect for the studied population and period | Broken assignment, contamination or implausible counterfactual: withhold the causal claim |
These categories do not provide interchangeable answers. GA4 and Meta are examples of credit rules; Northbeam, Triple Whale and Rockerbox are examples of reconciliation; Robyn, Meridian and LightweightMMM are modelling examples; Google Ads Conversion Lift, Meta Conversion Lift, GeoLift and CausalImpact are experiment or quasi-experiment examples. The evidence review distinguishes the claims each class can support.
Five Method Classes, One Grid
Platform credit
- Required data
- Observed events in scope
- Output
- Credit under rule
- Main causal limit
- Allocation is not incrementality
- Failure signal
- Unknown eligibility or duplicates
Reconciliation
- Required data
- Source records plus rules
- Output
- Matched and conflicting states
- Main causal limit
- Agreement does not explain cause
- Failure signal
- Unresolved collisions
Descriptive
- Required data
- Defined events and period
- Output
- Counts, rates, journeys
- Main causal limit
- Observation cannot isolate effect
- Failure signal
- Missing denominator
Causal model
- Required data
- Intervention and comparison
- Output
- Estimated effect
- Main causal limit
- Generalisation remains bounded
- Failure signal
- Broken assignment
The same grid, applied to the named methods
Each method below is read against the same six columns: inputs, assumption, output, uncertainty, causal limit and failure condition.
- First-touch.Inputs: ordered touchpoints per person. Assumption: the first recorded interaction deserves all credit. Output: whole credit to that touch. Uncertainty: none is reported. Causal limit: it cannot say the first touch caused the sale. Failure: missing early touches or split identities.
- Last-touch.Inputs: the same touchpoints. Assumption: the final interaction deserves all credit. Output: whole credit to the closing touch. Uncertainty: none is reported. Causal limit: it favours channels that close sales, whatever started them. Failure: direct or branded visits absorbing credit.
- Multi-touch (linear, time decay, position-based, data-driven).Inputs: ordered touchpoints. Assumption: a stated rule or fitted model splits credit. Output: shares of credit per touch. Uncertainty: model-based versions may offer none. Causal limit: shares are not effects. Failure: gaps in the path, as the model sees only tracked touches.
- Marketing mix modelling.Inputs: aggregate spend and outcome series. Assumption: the specification captures the main drivers. Output: modelled contributions with intervals. Uncertainty: stated by the model. Causal limit: association unless validated. Failure: weak variation in spend.
- Geo or conversion-lift experiment.Inputs: a designed treatment and comparison. Assumption: assignment holds. Output: an estimated effect for the studied population and period. Uncertainty: stated. Causal limit: this is the class that can support causal language. Failure: contamination between groups.
No row wins overall. First-touch and last-touch differ only in which touch they credit, so compare them with each other on the same journeys before comparing either with a model or an experiment.
Mistake: Relabelling a modelled association as return
Calling a marketing mix model estimate incremental return overstates what the model establishes: it produces modelled associations, not automatically causal effects. State the outcome, input series, geography, period, controls, transformation, model version and uncertainty before presenting an estimate. That record lets readers inspect the scope of the claim.
MMM can use aggregate data without person-level paths, but that does not remove data-quality or identification problems. Review the MMM comparison without treating its results as causal proof. Report the assumptions and diagnostics that shape the estimate, including the choices behind the model specification.
Declare Every MMM Specification Element
- Outcome
- Input series
- Geography and period
- Controls
- Transformation
- Model version
- Uncertainty
Mistake: Treating observed movement as its own comparison
An observed rise or fall alone cannot show what marketing caused. An experiment relies on a designed comparison, while a quasi-experiment relies on an identification strategy using available variation. Check implementation, outcome quality, assumptions and uncertainty before making a causal claim; the method comparison explains these distinctions.
Monitoring observed change cannot, by itself, identify an intervention's effect. Ask whether a counterfactual is plausible for the specific question, then limit any estimate to the population and period studied. Choose between attribution and incrementality according to the evidence question, rather than a preference for stronger wording.
Mistake: Choosing a method before defining the decision
A method selected before the buyer states the decision may produce a technically tidy answer to the wrong question. Set the decision and evidence question first, then choose the method that fits the required claim and its evidence limits.
Record the owner, population, event, numerator, denominator or unit, geography, period, timezone, source and method versions. Agency reporting templates can make those fields explicit. Include late, duplicate, missing, unmatched, refunded and corrected states, and compare required work, uncertainty and failure conditions without inventing weights or a winner.
Use a planning template to hold the acceptance boundary, and keep the marketing strategy plan tied to a buyer need. The decision should determine which evidence is relevant; a method should not be selected simply because its output is convenient to report.
Mistake: Publishing after every candidate fails
Publishing a preferred result after every candidate fails its evidence conditions is not a defensible compromise. Mark the decision HOLD, name the failed condition, and state what must change before reconsideration. A missing record should remain visible rather than being turned into confidence.
The Office for Statistics Regulation's Code of Practice for Statistics offers a check for an official-statistics quality claim; its guidance on administrative data supports transparent treatment of data sources, methods, comparability, strengths and limitations. This is a useful discipline, not a substitute for testing a method's specific assumptions.
This comparison is a desk review researched on 6 September 2026 for attribution decisions in organisations operating in England. It excludes supplier ranking, price, performance and implementation recommendations; no supplier or method funded the draft. Keep quality analysis, uncertainty and the decision boundary visible rather than collapsing them into a single score.
