Causal vs Rule-Based vs Data-Driven Attribution in GA4

GA4 offers two kinds of attribution model: rule-based last click, in its paid-and-organic and Google-paid-channels forms, and data-driven attribution, the default. Both answer the same question, how to share credit for a conversion among the sessions GA4 could resolve, and they differ only in the sharing rule. Causal attribution answers a different question, what would have happened without the channel, and it comes with two things the GA4 models cannot supply: an interval, and a statement of how much of the journey the data could see.

What GA4 actually offers

Since 2023, when first click, linear, time decay and position-based models were removed from GA4 and Google Ads, the attribution settings hold data-driven attribution as the default, plus last click in two variants. Lookback windows are set per property, 7 or 30 days for acquisition conversions and 30, 60 or 90 days for other conversion events. Is GA4's data-driven attribution a black box you can trust? covers the model's internals; this article is about the question each model answers.

Rule-based: a rule, applied

Last click gives the whole conversion to the last non-direct session in the window. It is transparent, stable, and wrong in a predictable direction: it hands credit to whatever sits closest to the purchase, which is branded search, retargeting and email, and starves whatever created the demand. The eBay experiment the book quotes is the canonical case: switching off brand-keyword ads lost no measurable sales because 99.5% of the forgone paid clicks came back through natural search, and last click had been scoring those ads as the best channel in the account.

Data-driven: a learned rule, applied to the same journeys

Data-driven attribution uses a model fitted on the property's observed conversion paths to share credit among touchpoints according to how much each changed the probability of conversion, compared with paths that lacked it. It is a better sharing rule. It is still a sharing rule over journeys GA4 could see, and it inherits three limits.

First, coverage. The Price of Being Found's own census found 48.6% of sessions at one company with no resolvable source; whatever the share is on your property, data-driven attribution distributes credit over the other part and says nothing about the missing half.

Second, selection. Paths that contain a touchpoint were chosen by the platform that served it, to people more likely to convert. A model that fits observed paths faithfully reproduces that targeting. In the experiments Gordon and colleagues ran at Facebook, observational methods were off by a factor of three in half the studies, and the naive comparison fit the data beautifully while being wrong by more than four times.

Third, no counterfactual. The model compares paths with and without a touchpoint. It never observes what a converting customer would have done had the touchpoint not existed, which is the only question a budget needs answered.

Causal: a different question

Causal attribution estimates the counterfactual: how much revenue would not have happened without the channel. Two honest designs observe it, the individual holdout and the geographic holdout, and an observational causal read estimates it from natural variation in the data (weeks a channel was scaled or paused, promotions, seasonality) and states an interval that widens where the data cannot tell.

What distinguishes a causal number from a GA4 number is not a better algorithm. It is what it states: its coverage, its design, and its interval with the floor of the design that produced it. "Channel X: incremental ROAS 1.9, interval 1.2 to 2.6, observational read on a 40-day export, coverage 0.61" is a sentence that can be checked. "Channel X: 34% of conversions, data-driven" is a share of a total that may itself be half visible.

Side by side

What to do this week

• If you own the budget: keep GA4 as the free description it is, and stop reading its channel shares as effects. Compute coverage for last quarter first.
• If you have to defend the number: label every GA4 figure in the deck "observational, data-driven" and put an interval beside anything you call incremental. Data-driven attribution vs causal attribution has the longer treatment.

A causal read on the same GA4 export produces the per-channel estimates with intervals, in 5 to 10 minutes, and the demo shows the output on a sample store with no signup.

GA4 attribution settings as documented on 9 September 2026. The experimental findings and the coverage census are quoted as The Price of Being Found (Edition 2.10) presents them in Chapters 8 and 9, with the book's scope notes.