What is a halo effect?

Advertise one product and customers buy others. Improve one page and satisfaction rises across the journey. Run a brand campaign and a competitor's paid search gets more expensive because more people are typing your name. In each case the benefit is real and lands somewhere other than the thing that was changed — so a metric drawn tightly around the change reports less than actually happened.

This is cannibalization with the sign reversed, and the two share one root cause: a metric scoped more narrowly than the effect. Cannibalization means the narrow metric flatters the result because gains were taken from elsewhere. A halo means it understates the result because gains landed elsewhere. Both are failures of scope rather than of statistics, and both are found the same way — by measuring at a level wide enough to contain the spillover.

The consequences fall unevenly, and predictably. Changes whose effects are concentrated and immediate measure well: a checkout fix shows up in checkout conversion. Changes whose effects are diffuse measure badly: brand advertising, content, trust signals and quality improvements all tend to produce small distributed gains across many surfaces, none of which is individually significant. An organisation that funds only what measures cleanly will systematically over-invest in the first kind and under-invest in the second, without anyone deciding to.

In media measurement this is the strongest argument against channel-level attribution. Upper-funnel advertising works partly by making later touchpoints more effective — people who saw the brand campaign convert better on search — and a model that assigns credit to the last click books that entire benefit to search. The brand campaign then looks like a poor performer and gets cut, after which search performance quietly declines. Only a method that sees all channels at once catches this: a geo experiment that withholds the campaign from some markets, or marketing mix modelling fitted across the whole portfolio.

The practical difficulty is that widening the metric costs power. Total revenue is far noisier than one product's revenue, so an experiment honest enough to capture the halo may be unable to resolve it. That is a real constraint rather than an argument against trying: the usual resolution is to measure the narrow metric for the decision and the wide one for the programme, checking periodically whether the sum of individually measured effects matches the aggregate.

A retailer runs a four-week brand campaign for its outerwear line. Last-click attribution is used alongside a geo experiment in which the campaign is withheld from 18 matched markets. The question is what the campaign was worth, and the two methods disagree substantially.

On the attribution number this campaign gets cut. On the geo number it returns £1.67 for every pound and should be expanded. The gap is almost entirely halo: £372,000 of the lift appeared in categories the campaign never mentioned, which last-click has no mechanism for seeing because those purchases arrived through search and direct visits with no campaign touchpoint. Two cautions before treating the geo number as settled. The 18-market design has its own uncertainty and the interval around £870,000 is wide — this is one measurement, not a constant. And a four-week campaign may pull forward purchases that would have happened later, which a four-week measurement window would count as incremental; extending the post-period is what distinguishes genuine growth from timing. What the comparison does establish firmly is that the attribution figure is a floor rather than an estimate.