Privacy first measurement: attribution methods and what each output really means

Privacy first measurement attribution methods do not answer one interchangeable question. Last interaction allocates credit by a rule. A platform model allocates credit using its data and assumptions. A controlled evaluation estimates a causal difference against a comparison. Choose by the decision, not by which output assigns the most credit.

What to take away

Comparison method

Attribution comparison research covers three methods against five criteria: question answered, event coverage, identity or path dependency, output and causal strength. Privacy input, uncertainty and permitted use follow.

Covers England organisation, UK privacy duties, a global product record and UK government evaluation guidance. No account or experiment was tested, and prices, rankings, performance claims and deprecated models are excluded.

The comparison is non-ranked. Qualified privacy, PECR and causal-inference reviewers must inspect the live design.

Last-interaction allocation

This method gives all credit to the final eligible interaction under the system's rule. It requires a recorded outcome, an eligible path and a defined lookback window. Its output is credited outcome value by channel. It is transparent enough to reproduce when the path is observed, but it ignores earlier eligible interactions and unobserved paths.

Google Ads currently documents last click and data-driven models and notes that several older rule-based models are no longer supported on its attribution models page. That is a global first-party product record, not evidence about another platform or the buyer's configuration.

Use last interaction for a bounded reporting convention when readers understand the rule. Do not call the credited channel the cause of the outcome.

Platform-modelled allocation

A modelled approach uses recorded paths and statistical assumptions to distribute fractional credit. Required data, eligible channels, reattribution windows and treatment of missing observations depend on the product. The output remains attributed credit, even when the algorithm uses counterfactual modelling internally.

Google Analytics describes data-driven, paid-and-organic last click and Google-paid-channels last click in its first-party attribution documentation. The page states that models vary by event and account data and that credit can be reassigned. No property was accessed for this comparison.

This method may help compare reporting allocations inside its defined ecosystem. Keep observed outcomes separate from modelled credit, publish the model and report version, and recheck after eligibility or product changes. Do not export its credit as a universal causal estimate.

Controlled incrementality

An experimental or credible quasi-experimental design compares an exposed or intervention group with a counterfactual group. It needs a pre-specified population, assignment unit, intervention, outcome, period, exclusions, analysis method and uncertainty. Its output is an estimated difference under the study design, not a division of every recorded conversion among touchpoints.

The 2026 government impact-evaluation quality guidance says monitoring an outcome does not show that an intervention caused it and explains the role of a matched or randomised comparison. It concerns policy evaluation; commercial validity still depends on specialist design and the actual setting.

Use this route when the decision genuinely requires incrementality and the design is ethical, lawful and adequately powered. Record interference, attrition, non-compliance and generalisability limits.

Compare outputs without false equivalence

Define one outcome event and population before running any method comparison. Keep the observation period, time zone, inclusion rule and source data fixed. Then show allocated credit and experimental effect in different columns because their numerators and denominators differ.

Illustrative figures, not measured data. A period records 1,000 conversions. Last interaction credits Channel A with 550 and Channels B and C with 450. A holdout test on Channel A estimates an incremental 90 conversions, interval roughly 40 to 140. The columns use different units.

For any displayed rate, publish the event, numerator, denominator, population, unit, period, time zone, inclusions, exclusions, owner and caveat. If one method cannot supply those fields, mark it not comparable. Stop when permission behaviour fails, path coverage changes materially, a model version is unknown or the counterfactual is compromised.

Attribution can organise reporting. Only an appropriate causal design can support an incremental claim, and even that claim remains bounded by its population, period and uncertainty.

Consent and permitted use

UK rules constrain what these methods may collect. Under PECR regulation 6, storing or reading information on a user's device generally needs consent unless it is strictly necessary for a service the user requested.

UK GDPR requires a lawful basis, a specified purpose and data minimisation. Consent must name the measurement purpose, and withdrawal must stop the processing.

Check the consent banner, the tags fired and the retention setting before comparing outputs, because no method widens the permitted use of the data.

Before you act

Common questions

What does a controlled incrementality design actually output?

It outputs an estimated difference under the study design, not a division of every recorded conversion among touchpoints. It needs a pre-specified population, assignment unit, intervention, outcome, period, exclusions, analysis method and uncertainty. Record interference, attrition, non-compliance and generalisability limits.

When should last interaction be used?

Use it for a bounded reporting convention when readers understand the rule. It gives all credit to the final eligible interaction under the system's rule and requires a recorded outcome, an eligible path and a defined lookback window.

What must be published for any displayed rate?

Publish the event, numerator, denominator, population, unit, period, time zone, inclusions, exclusions, owner and caveat. If one method cannot supply those fields, mark it not comparable. Stop when permission behaviour fails, path coverage changes materially, a model version is unknown or the counterfactual is compromised.