Definition
Incrementality is the portion of sales that happened only because an ad ran, rather than sales the store would have received anyway. Attribution divides credit among the touchpoints a buyer passed through; incrementality asks the harder question of what would have happened with no touchpoint at all. It is measured by holding a comparable group of customers or regions out of the campaign and comparing what they bought with what the exposed group bought.
Formula
Incremental conversions = Conversions in the exposed group - Expected conversions from the holdout, scaled to the same size
- Exposed group
- Customers, regions or audiences that saw the campaign during the test window
- Holdout
- A comparable group deliberately excluded from the campaign for the same window, matched on size and pre-test sales behaviour
- Incremental ROAS
- Incremental conversion value divided by the spend that produced it, which is usually well below the ROAS the platform reports
A geo holdout run over six weeks
- Regions in the exposed group
- Six provinces and states, matched on prior revenue
- Regions in the holdout
- Six comparable regions, campaign paused
- Spend in the exposed regions
- $96,000
- Revenue in the exposed regions
- $430,000
- Revenue in the holdout, scaled to the same base
- $302,000
- Incremental revenue
- 430,000 minus 302,000 = $128,000
- Incremental ROAS
- 128,000 ÷ 96,000 = 1.33
- Platform reported ROAS for the same campaign
- 4.5
Illustrative figures. The gap between 4.5 and 1.33 is not an error in the platform. It is the difference between all revenue the platform can associate with its ads and the revenue that existed only because the ads ran.
Why it matters
Incrementality matters because the channels that report the strongest numbers are often the ones closest to a purchase that was already going to happen. Branded search, remarketing and audiences built from recent site visitors all look excellent in platform reports and can be the least incremental spend in the account. Testing tells a store which budgets are buying growth and which are buying credit for existing demand, and the answer frequently changes how a budget is split. It also gives an independent check on attribution: when reported ROAS and tested incremental ROAS move in the same direction over time, the attribution setup can be trusted for day to day steering.
Where it goes wrong
- Running a test too short to cover the purchase cycle, so the holdout regions simply delay their orders and the measured lift is larger than the real one
- Choosing holdout regions that differ in seasonality, store density or shipping speed, which puts a difference into the comparison that had nothing to do with the ads
- Changing something else during the test, such as a promotion, an email campaign or a price change, which makes the result unattributable to the variable under test
- Testing with too little spend or too few regions to see past normal weekly variation, and then treating an inconclusive result as proof the channel does not work
- Reading one test as permanent: incrementality shifts with creative, seasonality and how much of the audience already knows the brand, so the question is revisited rather than settled
Questions about incrementality
How do you run an incrementality test without stopping revenue?
Hold out a small share of the market rather than a channel in full. Geo splits are the usual approach because they are clean to implement and simple to measure from store data: pause the campaign in a matched set of regions, keep everything else identical, and compare. Google Ads and Meta also offer built-in conversion lift and geo experiment tools that run the split inside the platform.
Why is incremental ROAS lower than the ROAS in the platform?
Because platform ROAS includes every conversion the platform can connect to an ad, including shoppers who already intended to buy and would have arrived through search, email or a bookmark. Incremental ROAS counts only the difference the spend created. Both figures have a use: one steers bidding day to day, the other decides whether the budget deserves to exist.
How large does a store need to be to test incrementality?
Large enough that normal weekly variation in orders is smaller than the effect you are looking for. Stores with modest order volume can still learn something from a full channel pause with a clear before and after, provided nothing else changes and the pause runs long enough to cover the usual purchase cycle. The trade is a less precise answer, so treat it as directional.