Incremental ROAS per Channel From GA4 Without a Geo Test

Incremental ROAS per channel can be estimated from a GA4 export without a geo test, by reading the natural variation already in the data: weeks a channel was scaled, weeks it was paused, promotions, seasonality. A causal read does that in 5 to 10 minutes and returns an estimate per channel with a confidence interval and the coverage of the data stated. It is an observational estimate, not an experiment, and the honest version says so; the channel that matters most still earns a holdout, and the read is what tells you which one.

What the export contains that a model can use

A 40 to 90 day GA4 export is not a flat line. Spend moved, campaigns launched and ended, a promotion ran, a weekend behaved differently from a Tuesday. Each of those is a small natural experiment, and a causal model uses them the way a geo test uses regions: as variation in the input that lets it estimate what revenue would have been without each channel. The more the window varied, the tighter the interval; a window in which nothing changed gives the model little to read, and the interval says so by widening.

The how it works page states the assumptions this rests on: spend and sales data complete and accurate, enough variation to identify each channel's effect, and no unobserved external factor moving sales in lockstep with one channel. Where an assumption is uncertain, the interval widens rather than the number pretending.

What comes back

Per channel: an incremental ROAS estimate, its interval, and the platform-reported figure beside it. Across channels: a ranked reallocation, what to scale, what to cut, what to leave alone, and an exportable report. Summed, the platform-reported figures against orders give the claim ratio, which The Price of Being Found treats as the most useful single number when three systems disagree. How to use GA4 exports with Causality Engine has the export steps.

What it cannot do, stated up front

Three limits, and they are the same limits any observational method has.

It cannot see what the data cannot see. GA4 attributes sessions it could resolve; the rest sit in Direct and Unassigned. The book's own census found 48.6% of sessions at one company with no resolvable source. The read states its coverage so that the estimate is read against the visible share, not mistaken for the whole.

It is not an experiment. No observational estimate observes the counterfactual directly. The read's value is that it reports an interval instead of a point, and labels its design, which is what the book's four properties of a defensible number require. A channel whose interval includes zero has not been shown to add revenue at this data volume; it has not been shown to add nothing either.

It has a floor. Below roughly 5,000 euros a month in paid spend, or with fewer than 40 days of history, the intervals are usually too wide to move a decision, and the read says so rather than inventing a number. Two of the five published customer cases were declined for that reason.

When you still need the geo test

For the largest channel, the read is the map and the holdout is the territory. Qualify it first: spend share times honest incremental return has to clear the smallest lift a geo test can detect at your scale, about 8.3% for a typical DTC brand on 26 weeks of history and 8 weeks live. If it clears, hold that channel out of a slice of regions with the pre-period fixed and the design registered, and let the read cover the other channels between anchors. Incrementality testing for ecommerce is the playbook, and the book's cadence is one qualified holdout per quarter with the anchor date on the dashboard.

Most brands cannot qualify more than one or two channels for a holdout at all. That is exactly the case the observational read is for: an interval on every channel, a stated floor, and a rank order for which channel earns the next experiment.

What to do this week

• If you own the budget: export the last 40 days from GA4 and run the read. Sort the channels by the lower bound of their intervals, not by the point estimate.
• If you have to defend the number: label the read "observational, with intervals" on the slide, and put the one holdout you can qualify next to it as the anchor.

The interactive demo runs the read on a sample store with no signup; a read on your own export is EUR 99 at app.causalityengine.ai, refunded if it does not pay for itself in one reallocation decision.

Product facts as stated on causalityengine.ai on 9 September 2026. The coverage census, the four properties and the measurability arithmetic are from The Price of Being Found (Edition 2.10), Chapters 8, 15 and 19, with the book's caveats.