What is incrementality?

A channel reports a thousand conversions. The question incrementality asks is how many of those thousand would have happened anyway — through organic search, direct traffic, a different channel, or simply because the customer had already decided. The answer is the only number worth budgeting against, and it is never the reported one. The reported figure counts people who saw an ad and converted; incrementality counts people who converted *because* they saw it.

The gap is structural rather than a measurement error. Retargeting is the clearest case: it shows ads to people who have already visited your site, browsed a product and often added it to a basket — that is, people who were already the most likely to buy. The platform then claims the conversions it observes, and the arithmetic looks superb. Branded paid search has the same shape: someone typing your brand name into a search engine was going to reach you regardless, and the paid click frequently just substitutes for the organic one directly beneath it.

Attribution modelling does not fix this, and it is important to see why. Last-click, first-click, linear, time-decay and data-driven models all take a fixed pot of observed conversions and divide it among the touchpoints that preceded them. Every one of them assumes the conversion happened because of the marketing and argues only about which channel gets the credit. Incrementality asks whether the conversion would have happened at all — a different question, and one no division of observed data can answer.

Measuring it requires withholding. Somebody has to not see the advertising, so their outcomes can serve as the counterfactual. At user level that is a holdout in a platform's own lift study or a public-service-announcement control, which is clean where available and limited to what the platform will run. Above that it is a geo experiment, turning spend off or up in randomly chosen regions and comparing — the workhorse for anything that cannot be hidden from an individual. And across all channels at once it is marketing mix modelling, which infers the counterfactual from variation over time rather than by withholding.

The results reliably surprise people, and reliably in the same direction. Studies across large advertisers routinely find incremental lift far below platform-reported conversions on lower-funnel channels, and it is common for the sum of channel-reported conversions to exceed total company conversions — because several platforms claim the same customer. Once measured, incrementality changes budgets rather than dashboards: the point is that a channel with a lower reported ROAS and higher incrementality is worth more than one with the reverse, and only one of those numbers appears in a platform UI. We work through the practical consequences in measuring the true effect of your ads with geo tests.

A retailer spends £400,000 a month on retargeting. The platform reports 20,000 conversions at a £20 cost per acquisition. To check it, they run a geo holdout: retargeting is switched off in 20 randomly chosen markets for six weeks, with 20 matched markets left on.

Ninety-three per cent of the conversions the platform claimed would have happened without the advertising, which is a normal result for retargeting and still lands badly in a meeting. Note what the holdout did not do: switching retargeting off did not cut conversions by 93% — sales in the off markets fell by only 7.4%, because those customers had already chosen and simply arrived by another route. The decision this unlocks is not necessarily "stop retargeting": at £270 it may still clear the margin on a high-value basket. It is that the £400,000 was being compared against a £20 CPA when the real one is thirteen times higher, and every budget decision made on that comparison was made on the wrong number.