What is incrementality in marketing?

Every ad sits on top of sales you would have made anyway. Some shoppers already know your brand, others find you in search, and some are about to reorder. If an ad reaches them right before they buy, the ad platform counts the sale, but the ad didn't make it happen.

Incrementality asks a simpler question: how many sales happened because of the ads? In numbers:

incremental sales = sales with ads − sales without ads
incremental ROAS   = incremental sales ÷ ad spend
lift               = incremental sales ÷ sales without ads

The hard part is the second line of that first formula. You never see "sales without ads" for the same people at the same time. Every method for measuring incrementality is a way of estimating it. Those extra orders are incremental sales; in dollars they are incremental revenue. When the question is specifically about ad dollars, see marketing incrementality.

Attribution is a set of rules for handing out credit. Meta credits a sale to an ad if the buyer clicked or saw it within a set window. Amazon Attribution credits a sale if the shopper clicked a tagged link and bought within 14 days. Google, TikTok and the rest have their own rules. Add them up and the total is usually more than your real sales, because each platform counts the same buyer.

Attribution gets cause wrong in both directions:

• Too much credit. It counts buyers who would have bought anyway, like loyal customers who happen to see a retargeting ad.
• Too little credit. It misses buyers it can't track. Someone sees your Instagram ad, doesn't click, and searches for you on Amazon the next day. Meta never sees that sale, and Amazon Attribution doesn't either.

Researchers have tested this directly. A study of large ad experiments at Facebook compared the experiments' answers with what common non-experimental methods would have said, and found those methods often got the effect wrong (Gordon and others, 2019).

To estimate "sales without ads", you need a group that doesn't see the ads but is otherwise like the group that does. That group is called a holdout group. There are two main ways to make one:

• Hold out people. The ad platform randomly keeps some users from seeing your ads and compares their purchases with everyone else's. Meta's conversion lift studies work this way. They can only count purchases the platform can see, which matters if you sell on Amazon. See conversion lift vs. brand lift.
• Hold out places. You turn ads off in some regions and compare their sales with regions where ads stay on. This is a geo test. It works with any sales data that records where orders ship, including Amazon's.

Both are experiments. There are also models that estimate incrementality from history, like marketing mix models. Models are useful for many channels at once, but they work best when experiments check them. For a methods overview aimed at Meta → Amazon, see incrementality measurement.

A made-up brand spends $6,000 on Meta ads in a month. Two reports and one test say three different things:

1. Meta's report credits the ads with $18,000 of sales on the brand's own site: a ROAS of 3.0.
2. Amazon Attribution credits $3,600 of Amazon sales to tagged ad clicks: a ROAS of 0.6.
3. A geo holdout test finds that, across both channels, the ads caused about $11,000 of sales the brand wouldn't have had: an incremental ROAS of about 1.8.

None of the reports is "lying". They answer different questions. Only the test answers the one that matters for the budget: what would happen to sales if the ads stopped? The attribution gap calculator lays the three numbers side by side for your own brand.

Incrementality changes decisions. If a campaign's reported ROAS is 4 but its incremental ROAS is 1, cutting it barely moves sales. If reported ROAS is 0.6 but the ads drive Amazon sales no report can see, cutting them can hurt more than you expect.

It also changes how you judge profit. An extra sale only helps if it makes money after product costs, fees and the ads that caused it:

• Compare incremental ROAS with your break-even ROAS, not reported ROAS.
• Use incremental results to split budget between channels, since each platform over-credits itself by a different amount.

• One test is one moment. It measures your ads at one spend level, with one set of creative, in one season. Re-test when those change.
• Small effects need big tests. Sales are noisy. In 25 large ad experiments at US retailers and brokerages, the median uncertainty range on return on ad spend was more than 100 percentage points wide (Lewis and Rao, 2015). A small brand needs to plan the test size first, with something like the holdout test calculator.
• Tests cost something. If the ads work, the holdout sells less during the test. That's the price of knowing.

Spillmark measures the incrementality of Meta ads on Amazon with geo holdout tests across US states. It splits the states into two groups that normally sell in step, tells you which group to pause Meta ads in, and compares Amazon orders against how the groups usually compare. Every result comes with its uncertainty range, and the test design is locked before it starts. The full method and math are on how lift is calculated.