Marketing Attribution Advertising Rules: Proving Accuracy Claims

What evidence supports an advert's claim that an attribution model is accurate? Define the output being tested, set a reference, validate the model and examine its errors and limits. Measured accuracy is also different from evidence that marketing caused an outcome. An accurate prediction is not, by itself, proof of campaign impact.

Define what the accuracy claim means

Accuracy is not self-defining. A dashboard might report conversions observed by a tracking setup, credit assigned by a platform, a model's prediction of an outcome, or an estimate of incremental effect. Those outputs answer different questions. Calling any one accurate without saying what was tested leaves the claim unclear.

Turn the proposed claim into a testable proposition. Is the model predicting an outcome, matching recorded events, or assigning credit across channels? What is the unit being tested: a conversion, a campaign, a channel total or a revenue amount? What evidence would count as a correct result for that unit?

Platform-attributed credit describes how a configured product assigned credit within its coverage and lookback rules. It does not establish that the assignment is accurate against an independent reference, or that the credited marketing caused the outcome. The attribution methods guide can help distinguish the different kinds of output; the claim still needs evidence suited to the particular accuracy statement.

For example, a model that predicts the total number of recorded outcomes is being tested against those recorded outcomes. That does not automatically test whether its distribution of credit between campaigns is correct. A matching overall total can conceal errors in the individual assignments. Keep the claim no broader than the exact model task that the evidence evaluates.

Specify the test and its reference

Before scoring a model, write down what it predicts and what counts as the reference result. Without that definition, a metric can look precise while answering the wrong question. Preserve the study definition alongside the test results, rather than trying to reconstruct it after seeing the output.

For an accuracy assessment, establish:

These are not administrative details: changing the population, outcome or reference changes what the test means. If the advert addresses a wider or different audience than the evidence covers, narrow the claim to the tested scope or do not make it.

The marketing attribution quality checklist in England is useful for the wider set of study-definition fields; here, the priority is making the accuracy test reproducible and relevant to the claim.

Validate on data that tests the model

A model should not be judged only on the same cases used to develop or tune it. Keep evaluation data separate from model development, and record how the validation split was made.

The test cases should reflect the intended use of the model; a result on one population or period does not automatically support a claim about another.

Choose a metric that corresponds to the stated task. For numerical predictions, mean absolute error (MAE) and root mean square error (RMSE) are examples of error metrics, but they summarise errors differently and neither is a universal measure of attribution quality.

State the metric, its unit and how to interpret it. If the claim concerns classification or allocation rather than a numerical forecast, use a measure suited to that task instead of presenting an unrelated score as proof of accuracy.

Interpret a score against a meaningful comparator where one is available. The comparison should use the same target, population and evaluation cases; otherwise, the difference may reflect a changed test rather than a better model. Decide what the assessment is intended to establish before reviewing the result, rather than choosing a favourable measure after the outputs are known.

A validation split is necessary evidence, but its label alone does not make a test reliable. Keep enough information to explain how the cases were selected, what was excluded and whether the data represent the use described in the advert. Do not turn a single score into a broad claim about performance outside the tested conditions.

Inspect the errors, not just the headline score

An average error can hide where a model performs poorly. Review the error distribution and failure cases as well as the headline metric. Check whether errors cluster by relevant groups, campaigns, periods or outcome types, and whether missing or excluded cases could change the interpretation.

Look for consistent overprediction or underprediction, large individual errors and cases where the model fails to return a usable result. For attribution outputs, ask whether a favourable aggregate result conceals unreliable assignments to particular channels or campaigns. The point is not to promise that every subgroup can be assessed.

The aim is to identify which parts of the tested scope the evidence actually supports.

State uncertainty and limitations in the evidence file, and use them to bound the claim.

If a model has only been checked against one reference source, say what that source represents; do not imply that it verifies an underlying causal truth unless the test does so. If the advertised population differs from the validation population, the test does not establish accuracy for the untested group.

There is no single accuracy threshold that makes every attribution claim defensible. A useful assessment explains the task, reference, metric, observed error and limits. If those pieces do not support the proposed wording, the appropriate response is to narrow or avoid the accuracy claim, not to rely on a dashboard screenshot as a substitute for validation.

Keep accuracy separate from lift and return

A model can predict recorded outcomes well without establishing that marketing caused them. Accuracy against an observed reference is not the same as a counterfactual estimate of what would have happened without the marketing. A change from an earlier period is an observation, not proof of incremental effect.

Where an advert claims that a campaign generated lift, the evidence needs to address the counterfactual and the study design, not just the model's prediction error. See counterfactuals in incrementality testing for that separate question. Likewise, a claim about return needs a defined financial measure; an accuracy score alone cannot substantiate it.

Keep the wording attached to the thing actually tested. A disclosure process can record any later change to that wording. Evidence that a model predicts recorded events within a defined scope supports a statement about that prediction and scope, not a broader claim that it correctly identifies causal contribution, proves incremental sales or establishes return on investment.

Apply the advertising substantiation rule

For an England-facing advert within its remit, CAP Code rule 3.7 is the relevant substantiation requirement for claims capable of objective proof.

The ASA's advice online page, "Substantiation", explains that documentary evidence should be held before publication. For an attribution accuracy claim, the supporting material should therefore show what was tested, the reference used, the validation design, the metric and the observed limitations.

A supplier brochure or dashboard export can document what a product reports; on its own, it does not independently validate the accuracy of the method.

The evidence should support the specific performance statement being made, rather than a different output or a broader population. If the test cannot establish the proposition the advert states, change the proposition or leave it out.

The CMA's unfair commercial practices guidance covers business conduct that the law on unfair business-to-consumer trading prohibits under the Digital Markets, Competition and Consumers Act 2024. A published accuracy claim sits within that wider conduct, and the UK rules and compliance guide maps the other publication checks.

Tracking that feeds the model is a separate matter, and the ICO's guidance on storage and access technologies has been finalised.

Before publication, hold the evidence file for the claim: