How to measure incrementality

By Harry Burman · Updated October 20, 2026

Incrementality measurement estimates the sales ads cause by comparing a group that sees them with a group that does not, or by using a model checked by those tests. For Meta ads that sell on Amazon, geo holdouts are the practical default.

Incrementality measurement is the set of methods that estimate caused sales: sales with the ads minus sales without them. Platform attribution is not measurement of cause; it applies credit rules. The main experimental options are user-level holdouts inside an ad platform and geo holdouts that change ads by region. Models such as marketing mix modeling estimate channel effects from history and work best when experiments calibrate them. Brands that advertise on Meta and sell on Amazon usually cannot rely on Meta conversion lift alone, because Meta cannot see Amazon checkouts. A geo holdout that pauses Meta in matched US states and compares Amazon ship-to orders answers the budget question. Plan length and holdout share with a sample-size tool, lock the design, run for a fixed window, then report lift with a range. Read how lift is calculated for the geo holdout math.

What incrementality measurement means

Measurement here means estimating incrementality, not reading a dashboard. You need a credible stand-in for "sales without the ads." That stand-in is usually a holdout group, sometimes a model trained on past tests.

Marketing incrementality is the same idea applied to spend. Incremental sales and incremental revenue are the order and dollar forms of the result.

Methods at a glance

Common ways teams measure incrementality. Experiments answer cause most directly.
Method	What it compares	Best when
User holdout / conversion lift	People who see ads vs. people withheld	Sales the platform can track on-site
Geo holdout	Regions with ads on vs. ads paused	Amazon or other sales platforms cannot see
Matched markets + DiD	Pre-chosen regions that sold in step	You want a simple, transparent geo design
Synthetic control / GeoLift	Test region vs. a weighted blend of others	Few treated regions or uneven history
Marketing mix model	Spend and sales over time	Many channels; needs tests to stay honest

User holdouts inside Meta are clean for tracked site purchases. They miss Amazon orders. Geo designs only need sales by region, so they fit Meta → Amazon. Read geo testing, matched market testing, difference-in-differences and synthetic control for the building blocks. MMM vs. testing covers when a model is enough.

Steps for Meta ads that sell on Amazon

Export several months of Amazon orders by ship-to state (or ZIP3), plus Meta spend by day. Drop cancelled orders if your export includes them.
Pick test and control regions with matched market testing (or a synthetic control). Check that the groups sold in step before the test.
Size the holdout and length with the holdout test calculator so a realistic lift is detectable. Avoid Prime Day and other shocks unless the test is about those weeks.
Pause Meta in the holdout with location targeting. Keep creative and budget stable in the ads-on regions. Lock the design before day one.
After the fixed window, compare groups with difference-in-differences. Report incremental orders, revenue and iROAS with an uncertainty range.

The design rules and estimator are written out on how lift is calculated.

Mistakes that ruin the reading

Treating attributed revenue as caused. Credit windows are not experiments.
Skipping power. Small brands with noisy sales need longer tests or larger holdouts; otherwise every result looks like "no clear effect."
Changing creative mid-test in only one group. Keep the ads-on side stable so the contrast is about presence of spend.
Ignoring leakage across state lines. Spillover pushes measured lift toward zero; pick regions with that in mind.

Lewis and Rao (2015) showed how wide uncertainty ranges on return on ad spend can be even in large experiments. Plan for noise.

Where Spillmark fits

Spillmark runs geo holdout tests for Amazon brands on Meta. It plans matched state groups from your order history, tells you how long to run and which group to pause, locks the design before the test starts, compares Amazon sales with difference-in-differences, and keeps each result on record for your brand.

Frequently asked questions

What is incrementality measurement?

Any method that estimates sales caused by ads, usually by comparing a holdout that does not see the ads with a group that does, then reporting lift with a range.

How should Amazon brands measure Meta incrementality?

A geo holdout: pause Meta in matched US states, keep ads on elsewhere, and compare Amazon orders by ship-to state. User-level Meta lift misses Amazon checkouts.

Is attribution a form of incrementality measurement?

No. Attribution assigns credit by rules. Incrementality measurement estimates cause with an experiment or a model checked by experiments.

Can marketing mix modeling measure incrementality?

MMM estimates channel effects from spend and sales over time. It can be useful across many channels, but holdout tests should check the model.

How long should an incrementality test run?

Long enough to detect a realistic lift given your sales noise and holdout share. Use a sample-size calculator before you pause spend; many Amazon Meta tests run several weeks.