Skip to content

For GA4 usersFrustrated with GA4 attribution? Upload your GA4 export, see causal insights in 5–10 minutes for €99 pay-per-use.

Causal Inference

3 min read

How to decompose what actually sells

A practical sequence for separating concept from execution across a product line, using batches, a holdout and an interval rather than a hunch about which feature mattered.

Share
Quick Answer·3 min read

How to decompose what actually sells: A practical sequence for separating concept from execution across a product line, using batches, a holdout and an interval rather than a hunch about which feature mattered.

Read the full article below for detailed insights and actionable strategies.

Attribution by the numbers

Avg ad spend wasted

30%

Meta ROAS inflation

2.3x

Cost to find out

€99

Setup time

2 min

Separating concept from execution takes a batch design, not a better eye. Here is the sequence.

Step one: write the hypothesis down first

Name the property you believe is doing the work, before you design anything. "The subject matters more than the style" is a hypothesis. "This design is great" is not.

Writing it down first is what stops the result from being reinterpreted afterwards to fit whatever happened.

Step two: build the batch to vary one thing

Take one concept and produce several executions of it. Take one execution style and apply it to several concepts. Two batches, each holding one axis fixed.

The batch has to be large enough that a single item's luck does not decide the result. Small batches produce confident nonsense.

Step three: keep the traffic allocation fixed

This is the step that fails most often. If you push more traffic to the items that start well, you have replaced your experiment with the feedback loop it was designed to avoid.

Fix the allocation in advance and leave it alone for the whole window, even when one item looks like it is winning.

Step four: read the interval

ResultReading
Concept batch separates, execution batch does notConcept is doing the work
Both separateBoth matter, size them before choosing
Neither separatesUnderpowered, or the effect is small
Confidence interval spans zeroNot resolved, do not conclude

The last row is the one to respect. An unresolved result is a real outcome and it means run it longer or bigger, not pick the higher midpoint.

What this does not cover

Channel effects. If one batch happened to get more of a particular traffic source, the comparison is contaminated. That is the same incrementality question the channel work answers.

A causal read on a Google Analytics export returns per channel an estimate, an interval and a coverage share for 99 euro once, refunded if it does not move a budget decision, which tells you whether your traffic mix shifted under the test. The interactive demo shows it with no signup.

Key Terms in This Article

Related Articles

Sixty-second versions of these ideas: Causality Engine on YouTube Shorts.

Ready to see your real numbers?

Own the budget? Upload your GA4 export and see which channels drive incremental sales, with confidence intervals, in minutes. Have to defend it? Start with the live demo and take the read to your CFO.

Full refund if you don't see value.

Stay ahead of the attribution curve

Weekly insights on marketing attribution, incrementality testing, and data-driven growth. Written for the person who owns the budget and the person who has to defend it.

Which one are you? Optional.

No spam. Unsubscribe anytime. We respect your data.

Frequently Asked Questions

How do I separate concept from execution in a product line?

Build two batches. One holds the concept fixed and varies execution; the other holds execution fixed and varies concept. Each must be large enough that one item's luck does not decide the outcome.

Why fix traffic allocation during the test?

Because pushing traffic toward early winners recreates the feedback loop the test exists to avoid. Fix the allocation in advance and leave it for the whole window, even when one item looks like it is winning.

What if neither batch separates?

The test was underpowered or the effect is small. Run it longer or larger. Picking the higher midpoint from an unresolved result is not a conclusion.

Related reports

Real reports on this topic.

Anonymised reports from the Attribution Report Library tagged with causal inference.

Browse all related reports

Find your wasted ad spend in 5–10 minutes.

Watch the model work on a sample store first, no signup. Then upload your last 40–90 days of GA4 sessions and get incremental ROAS with confidence intervals. No pixel, no SDK. €99 per read.

Prefer to talk it through? Book a 20-min call, or read how it works.

Last-click guesses.We run the math.

Causal attribution for ecommerce brands. Watch the model work on a sample store first, then upload your GA4 export and see which channels really drove revenue in 5–10 minutes. €99, pay-per-use. Pro at €299/mo when you want it continuous.

No signup for the demo. Book a 20-min call or compare plans.