What GA4 Data-Driven Attribution Assumes, and Where It Breaks

Teams often treat GA4's data-driven attribution (DDA) as the final word on channel value. It is an improvement on last click and linear rules: it uses machine learning to spread credit across touchpoints rather than handing all of it to the final one. But it remains a correlational model. It assigns value from observed paths; it does not measure what a channel causes or the incremental lift it provides.

Google describes DDA as comparing the conversion probability of paths that include a touchpoint with similar paths that do not. That comparison assumes the two groups of users differ only by that touchpoint. In practice they rarely do: people who click a brand search ad or a retargeting banner were already more likely to convert. DDA cannot correct for this selection bias, nor for unmeasured confounders such as offline media, seasonality or word of mouth. A channel can therefore score highly in DDA while a controlled experiment shows little incremental contribution, and budget moved on DDA's evidence follows correlation rather than causation.

The model also learns only from journeys GA4 can stitch together. When consent refusals, ITP cookie limits and device switching break journeys, early touchpoints drop out of the training data and credit drifts towards channels that sit close to conversion. Because the model's weights cannot be inspected, this drift is invisible in the interface.

Moving beyond DDA starts with reconstructing journeys outside it: transforming the raw GA4 export in BigQuery, stitching users across devices and sessions where a deterministic identifier exists, and enriching the result with CRM records, offline conversions and cost data. Modelling these relationships in dbt keeps them consistent and auditable. A server-side collection layer such as Robin Signals can load behavioural events into the same warehouse under the organisation's own control.

With a clean, unified dataset, analyses can be designed to mimic experiments instead of fitting credit to observed paths. Options include uplift modelling, incrementality tests (with quasi-experimental designs where randomisation is not feasible) and survival analysis of how touchpoints affect time to conversion. Libraries such as CausalImpact in R and Python support the quasi-experimental work.

DDA is a reasonable tool for distributing credit among observed touchpoints. Budget decisions, however, need incrementality, which DDA does not measure.