Attribution Is a Credit Rule, Not Proof That an Ad Caused a Sale

AI-generated conceptual illustration; no real customer journey, attribution result or causal measurement is shown.

A sale can appear beside an advertising interaction in a report without answering whether the sale would have happened without that interaction. Attribution organizes credit inside a measurement framework. Causal incrementality asks a different question: what additional outcome occurred because of the intervention? Keep the recorded business event, the allocation rule and the causal claim separate.

Begin with the event and the eligible path

An attribution model determines how conversion credit is assigned to eligible interactions. Google’s attribution-model explanation describes last click as giving credit to the last-clicked ad, while data-driven attribution uses account data. A model describes the allocation; it does not, by its name alone, establish that every influence on the customer was observed.

Imagine one fictional purchase on day 10 following ad interaction A on day 0 and ad interaction B on day 6. The exercise stipulates that the two interactions belong to the same journey and that the purchase is one accepted business event. Those are teaching assumptions, not a customer identity match, an imported record or a path assembled across platforms.

Before interpreting a real path, identify what the reporting system can credit. Google documents that creditable channels can differ with the conversion’s source and measured activity. A result scoped to particular channels is not automatically a reconciliation of every paid, organic, offline or otherwise unobserved influence. State the conversion action and the report’s channel scope alongside the model.

The same purchase can receive different allocations

For the fictional exercise, define two credit rules: give all credit to the latest eligible interaction, or divide credit equally between eligible interactions. The second rule is a deliberately simple teaching rule, not a claim that Google currently offers a linear model and not a simulation of data-driven attribution. Both rules operate on the same stipulated purchase.

At purchase, A is ten days old and B is four days old. Both fit the illustrative 14-day interval; neither fits the illustrative 3-day interval. In the first two rows, total credit is one unit, distributed differently. In the last row, this rulebook assigns no credit to either interaction. It does not erase the business event or prove the ads had no effect.

Data-driven does not mean a made-up split

Data-driven attribution is not the equal split in the table. Google’s data-driven-attribution explanation describes an advertiser-specific model that compares converting and non-converting paths to identify patterns and allocate credit. Its weights cannot be inferred from this single invented journey. Google also notes that data-driven and last-click models can sometimes produce the same result.

The provider currently distinguishes eligibility from the amount of data it recommends for more effective modelling; a recommendation should not be rewritten as a universal hard minimum. Keep the conversion action, available data and supported scope explicit. A credit allocation reported by a model is also not evidence that the reader has independently run a controlled incrementality experiment.

A conversion window is not the report’s date label

A conversion window is the interval after an eligible ad interaction during which a conversion can be recorded for that action. It concerns timing eligibility. A report’s selected period and date field concern how records are placed in the report. Do not assume that an interaction date, an outcome date and a window answer the same question.

The 14-day and 3-day intervals above are invented for arithmetic, not recommended settings or a statement of available windows for a particular Google conversion action. Real interpretation requires the interaction type, action settings and the report’s date semantics. A window from one product should not silently be treated as identical to another product’s lookback setting.

Explain disagreements before combining totals

When two reports disagree, first write down their definitions. Are they describing the same business action and reporting period? Do they include the same eligible interactions, channels and identities? Which events can be linked, and which remain unknown? Have timing rules, updates or exclusions changed? These questions form a reconciliation brief; they do not justify joining records merely because totals appear similar.

In the fictional case, imagine one document counts the accepted purchase while another lists only credited ad interactions under the short rulebook. Calling one document wrong would be premature: they answer different questions. Keep an uncredited business event visible as an event, and a credited interaction visible as an allocation. Neither category should silently replace the other.

Incrementality needs evidence for a different comparison

The causal question concerns the outcome under an intervention compared with what would otherwise have happened. A controlled comparison can help investigate it when its assignment, outcome definition, interference and uncertainty are handled appropriately. Microsoft Research’s experimentation principles stress planning the hypothesis, metrics and randomization. That methodological requirement is separate from choosing how to allocate an already recorded conversion.

For a fuller explanation of that comparison, continue with Did the Change Help? Metrics, Baselines and a Trustworthy A/B Test. The fictional path in this chapter has no control group, sample or observed lift. It cannot establish how many additional purchases advertising caused, even though its credit arithmetic is fully specified.

To distinguish a valuable business action from its signal and recorded conversion, read From a Valuable Action to a Measured Conversion. That definition comes before interpreting the credit attached to it. A reported allocation can be useful while remaining bounded by the action definition, the data the report covers and the evidence available for a separate causal claim.

A clear attribution statement names the event, eligible path, model, timing rule and reporting scope. A clear causal statement additionally explains the comparison and its uncertainty. If the evidence supports only credit allocation, report that result precisely. The reader should not have to guess whether “the ad got credit” was quietly changed into “the ad caused the sale.”