What is incrementality in marketing? How to tell if your spend actually caused the sale

Incrementality in marketing measures whether an ad exposure actually caused a sale, not just whether it happened before one. A channel is incremental when turning it off reduces total conversions; if conversions hold steady without it, that spend was capturing demand that would have converted regardless. This is a different question than attribution asks, and it usually produces a very different number than a platform's own reported ROAS.

Key terms in this article

Incrementality

The causal effect of an ad exposure on a conversion that wouldn't have happened otherwise.

Attribution

A method for assigning credit for a conversion across touchpoints, based on correlation, not causation.

Incremental lift

The percentage difference in conversions between an exposed group and a matched control group.

Geo split test

An incrementality test that holds out entire geographic regions from a campaign.

Holdout test

An incrementality test that withholds ads from a randomized subset of users within the same geography.

PSA test

An incrementality test that replaces ads with public service announcements for a control group, isolating the ad effect without a full media holdout.

What incrementality means

Attribution answers "which touchpoint gets credit for this sale?", while incrementality answers a harder question: "would this sale have happened anyway?" The two aren't interchangeable, and treating them as if they were is how advertisers end up overpaying for demand they already had.

The practical version of this problem: a channel can show a strong ROAS in a dashboard and still be worth nothing incrementally, if the people it reaches would have bought anyway through a different, cheaper path.

Incrementality vs. attribution — what’s the difference

Picture a shopper who searches a brand name, clicks a branded search ad, and buys. Attribution gives that search ad full credit for the sale. Incrementality asks whether the shopper would have typed the brand's URL directly and bought anyway, in which case the ad spend changed nothing.

Both have a role. Attribution helps you understand the journey; it can't tell you what would have happened without the ad. That's what incrementality testing is for.

How to calculate incremental lift

A simplified example: if the exposed group converts at 4.2% and the matched control group converts at 3.5%, incremental lift is (4.2 − 3.5) / 3.5 × 100, or roughly 20%. That means about one in five conversions in the exposed group can be attributed to the ad exposure; the rest would likely have happened anyway.

The result only means something if the control group is genuinely comparable to the exposed group in size, geography, and baseline behavior. A mismatched control group produces a lift number that looks precise and means very little.

The three ways to run an incrementality test

Each method trades off cleanliness of the result against how much reach or budget it requires you to sacrifice during the test window. A geo split gives the clearest read but needs enough markets to split meaningfully. A holdout is easier to run but can leak between exposed and control users if targeting isn't tight. PSA testing avoids a full blackout but depends on the platform supporting it.

What it takes to run a valid test

This is where most incrementality advice stops short: it explains the methods without saying what they require in practice.

Minimum volumes

There's no universal minimum; what matters is your test's Minimum Detectable Effect (MDE) — the smallest lift the test can reliably distinguish from ordinary week-to-week noise, given your baseline conversion volume, how much that volume naturally fluctuates, and how long the test runs.

Meta's open-source GeoLift framework, one of the standard tools for this kind of power analysis, is built around exactly this calculation: more test regions and longer windows lower your MDE, letting you detect smaller lifts, while volatile sales data raises it.

In practice, a test aiming to detect a 5–15% lift needs a baseline steady enough that a change of that size stands out from normal variation — a question worth running through a power calculation before launch, not guessing at.

Test window duration

Longer windows increase confidence but also increase the chance that something unrelated shifts demand mid-test: a seasonal spike, a competitor promotion, a pricing change. Shorter windows reduce that risk but need higher-volume traffic to still reach significance.

What to do when the result is inconclusive

This happens more often than test case studies suggest, especially at lower traffic volumes. An inconclusive result isn't a failed test; it usually means the lift, if any, is smaller than the test was powered to detect. The options from there are either to extend the window, widen the test to more geographies or a larger user pool, or accept that the channel's incremental contribution is likely marginal rather than run the same test again expecting a different answer.

Incrementality under pressure: what Black Friday breaks

Everything covered so far assumes a reasonably stable baseline. Black Friday and Cyber Monday remove that assumption, which makes BFCM a useful stress test for the whole framework.

Why BFCM baselines are unstable

Test windows depend on a control group behaving the way it normally would. During BFCM, nothing behaves the way it normally would. Conversion rates spike across every channel at once, including channels not running any promotion, because overall purchase intent is elevated. A control group's baseline from October is close to meaningless in the last week of November.

Measuring incremental sales during Black Friday

The formula doesn't change; what changes is the risk of a mismatched control group. A geo split run entirely within BFCM week needs regions with genuinely comparable holiday shopping behavior, not just comparable behavior the rest of the year. Where possible, running a shorter pre-BFCM baseline test and a separate BFCM-week test gives a more honest read than trying to isolate incrementality mid-surge.

ROAS vs. incremental revenue during BFCM.

This is where the attribution-versus-incrementality gap widens the most. Reported ROAS during BFCM tends to look strong across nearly every channel, because conversion rates are up broadly. That doesn't mean every channel is contributing incrementally; some of that ROAS is capturing demand that a lower-intent shopper would have converted on regardless, on a day when almost everyone converts at a higher rate. Optimizing purely for in-platform ROAS during BFCM week tends to reward channels that were already going to get the sale, not channels that created it.

What KPIs actually signal real incrementality

Total ROAS and conversion volume are the two metrics every dashboard already surfaces, and neither one distinguishes incremental from non-incremental demand on its own. The more useful signal is directional: conversion rate lift in a held-out group versus the same group's pre-BFCM baseline lift, if a test was already running before the surge started. Broader BFCM performance tracking, beyond this incrementality lens, is its own subject.

Why incrementality matters most for channels claiming incremental reach

Most advertisers don't have an attribution problem. Their issue is paying for demand that would have materialized anyway, something they only discovered after turning off the channel.

This matters most for any channel positioned as reaching audiences other channels can't. That claim is either true and provable, or it's a story a dashboard is telling. Commerce media platforms sit in exactly this position: the pitch is exposure to shoppers not already reachable through the walled gardens an advertiser is already running.

Commerce Media Tech’s inventory is built to be incremental to walled garden spend rather than overlapping with it, reaching shoppers in commerce and retail environments outside Google and Meta's ad surfaces.