Measurement
Part of Marketing automation measurement: a reporting framework
Marketing automation attribution methods for platform reviews
A framework for using marketing automation attribution methods to review platform claims and match each method to decisions it can support.
Marketing automation attribution methods are useful criteria in a marketing-automation platform review, but they answer different questions. Platform rules distribute credit among recorded touchpoints; observational analysis describes patterns; experiments and carefully designed quasi-experiments can address incremental effect. Review each output against the decision it can support, rather than treating all four as equivalent forms of attribution.
How to assess attribution claims in platform reviews
- Platform attribution, observational analysis, experiments and quasi-experiments answer different questions and should not all be called attribution.
- Platform outputs assign credit under a named model but do not validate causation or comparability between products.
- Observational analysis describes associations and needs an explicit exposure rule; it cannot remove confounding by adding fields.
- Experiments estimate incremental effect but require documented assignment, non-delivery and crossover reporting.
- Use a metric card to keep the outcome constant while the evidence design changes.
Comparison method
This review framework is for organisations operating in England. It compares methods on question, population construction, counterfactual, assumptions, uncertainty and defensible decision. No platform, model or result was tested. It is not a ranking, and there are no affiliate or supplier relationships.
Four attribution methods compared
Platform-reported credit
- Question answered
- Which interaction gets credit
- Main limitation
- Missing touchpoints, identity, model
- Suitable wording
- Attributed by named platform model
Observational association
- Question answered
- How outcomes differed by exposure
- Main limitation
- Selection, seasonality, concurrent activity
- Suitable wording
- Associated with exposure
Randomised experiment
- Question answered
- What changed under assignment
- Main limitation
- Feasibility, contamination, power
- Suitable wording
- Estimated incremental effect
Quasi-experimental evaluation
- Question answered
- What causal account fits
- Main limitation
- Identification assumptions remain central
- Suitable wording
- Wording tied to exact design
Question answered
- Platform-reported credit
- Which tracked interaction receives credit under this model?
- Observational association
- How did recorded outcomes differ across exposed records or periods?
- Randomised experiment
- What changed under the predefined assignment and analysis?
- Quasi-experimental or theory-based evaluation
- What causal account is supported without random assignment?
Main limitation
- Platform-reported credit
- Missing touchpoints, identity rules and model choice shape the output
- Observational association
- Selection, seasonality and concurrent activity can explain differences
- Randomised experiment
- Feasibility, contamination, power and implementation affect inference
- Quasi-experimental or theory-based evaluation
- Identification assumptions and alternative explanations remain central
Suitable wording
- Platform-reported credit
- "Attributed by the named platform model"
- Observational association
- "Associated with exposure"
- Randomised experiment
- "Estimated incremental effect" with uncertainty
- Quasi-experimental or theory-based evaluation
- Wording tied to the exact design and sensitivity
These examples show how to classify supplier-reported methods in a platform review, not how to rank products. Google Analytics 4 last click, first click, linear, position-based, time decay and data-driven models sit under platform-reported credit. HubSpot contact-create, deal-create and revenue attribution reports also sit under platform-reported credit. The Magenta Book supports theory-based evaluation and counterfactual design under quasi-experimental or theory-based evaluation.
Judge Google Analytics 4 last click for channel reporting within its 30-day acquisition lookback. Judge HubSpot revenue attribution for lifecycle reporting under its documented model. Judge Adobe Analytics Attribution IQ for within-platform credit across named models. Judge Marketo Measure full path for multi-touch diagnosis. Treat a 30 to 180 day lookback as a typical range.
The Magenta Book discusses counterfactual and theory-based evaluation and separates observed change from attribution. It is government guidance, not evidence that a particular marketing design works.
Read platform outputs by their own definitions
Google explains Analytics attribution models as rules, sets of rules or data-driven algorithms that assign credit to touchpoints. HubSpot's first-party guide describes contact-create, deal-create and revenue attribution reports with model and data requirements. These pages define supplier-reported outputs. They do not independently validate causation or comparability between products.
Preserve the platform, property settings, model, lookback treatment, eligible channels, identity rules and extraction date. A changed setting creates a new series even if the chart title stays the same.
Observational analysis needs a similarly explicit exposure rule. State whether a person counted after delivery, an impression, a click or journey entry, and keep unknown exposure separate. Compare pre-existing differences between groups before interpreting the outcome. A large adjusted model does not remove confounding simply because it contains many fields.
For an experiment, document assignment rather than observed delivery as the primary group definition unless the approved protocol specifies another estimand. Report non-delivery and crossover rather than deleting them. For a quasi-experiment, publish the identification assumption and a sensitivity that would overturn the decision.
Use one metric card across methods
For "qualifying order rate", record:
Metric card fields to record
- Populationeligible people before exposure
- Numeratorunique finance-reconciled qualifying orders
- Denominatorall eligible people, non-delivery included
- Unit, period, timezoneEurope/London cohort
- Inclusions and exclusions under frozen rules
- Sourceeligibility, exposure log, order ledger
- Uncertaintymissing links, contamination, selection
- Owner and update trigger with finance approval
- Populationall eligible people in the approved England evaluation before exposure or assignment.
- Numeratorunique people with a finance-reconciled qualifying order in the fixed outcome window.
- Denominatorall eligible people, including non-delivery and no-response cases.
- Unit, period and timezonepercentage for the named cohort and follow-up period in Europe/London time.
- Inclusions and exclusionsnamed products, refunds, cancellations, staff and test orders under frozen rules.
- Sourceeligibility snapshot, assignment or exposure log, order ledger and versioned analysis.
- Uncertaintymissing links, contamination, selection, sampling variation and model assumptions as applicable.
- Owner and update triggerevaluation owner with finance approval; re-version after eligibility, identity, margin, tracking or design changes.
The card keeps the outcome constant while the evidence design changes. That makes it harder to mistake a different denominator for a better method.
Choose by decision risk
Use platform attribution for channel reporting within its documented boundary. Use observational analysis for diagnosis and hypothesis generation. Reserve causal wording for a design reviewed by a qualified specialist, with assumptions and uncertainty shown. If the decision would commit substantial money or affect people materially, weak evidence should narrow the action rather than inflate the claim.
Match method to decision risk
What does the decision commit?
reserve causal wording for specialist-reviewed design
platform attribution for channel reporting, observational for diagnosis
Before you act
- Preserve platform, property settings, model, lookback, channels, identity rules and extraction date.
- State the exposure rule for observational analysis and keep unknown exposure separate.
- Document assignment as the primary group definition for experiments unless the protocol specifies another estimand.
- Publish the identification assumption and a sensitivity for quasi-experiments.
- Record population, numerator, denominator, unit, period, timezone, inclusions, exclusions, source and uncertainty on a metric card.
- Reserve causal wording for a design reviewed by a qualified specialist with assumptions and uncertainty shown.
Common questions
What does the Magenta Book say about attribution?
The 2026 Magenta Book discusses counterfactual and theory-based evaluation and separates observed change from attribution. It is government guidance, not evidence that a particular marketing design works.
