Spot the Outlier: Why Averages Hide Where Revenue Is Lost.
How outlier analysis reveals the channels, regions and segments that quietly decide growth.
Read the analysis →Attribution reports what marketing touched. It cannot say what marketing changed. Three kinds of test close that gap.
Only the gap between actual revenue and the modelled baseline is counted.
Every platform is happy to report how much revenue it drove. Add the claims together and the total is often larger than the revenue the business actually earned. Each platform counts every sale it touched, and most sales are touched more than once.
Attribution answers the question "which ads were near this sale". The question a board cares about is different: "what would revenue have been without this spend". That is a question about cause, and it needs a different kind of evidence.
The most direct evidence is an experiment. Run a campaign in some regions and not in comparable others, then compare. The gap between the two groups is the lift the campaign caused. Just as important is the interval around it. A lift estimated from a small test can be large and still be indistinguishable from zero.
Single regions are noisy. Pooling them gives an estimate narrow enough to act on, which is why a test should be designed across several regions from the start.
A result that could be zero is not a result. It is a reason to run a better test.
Not everything can be held back. When an experiment is impractical, the alternative is a baseline: a model of what revenue would have been, built from history, seasonality and markets that were not affected. Actual revenue is then measured against it, and only the gap is credited.
A baseline is an estimate and should be shown with its uncertainty. If the gap sits inside the band, the honest conclusion is that the effect is not yet proven.
Day-to-day decisions, such as which creative or landing page to keep, do not need a regional experiment. They need a clear rule. Reading a test as "the probability that B beats A" gives one: ship when the probability is high enough, stop when it is clear there is no difference, and keep testing in between.
Charts in this article are modelled examples and do not represent any client.