Incrementality

Why incrementality matters

Platform ROAS (return on ad spend) counts conversions the platform can see. Many of those people would have bought anyway. Incrementality asks the real question: did the ads cause extra sales? That is the only number that tells you whether to spend more.

Let's take a drink brand as an example. They sell a low-sugar sports drink for active women over 50, mostly from their own website. The drink brand runs Instagram, Google, and YouTube. In June, Meta (Facebook and Instagram) credits Instagram with $40,000 of website orders on $10,000 of spend (a 4x ROAS, or return on ad spend). Company revenue is almost flat. The usual suspects when a channel looks perfect in-platform and the bank account does not move:

• Brand search, meaning women already looking for the drink brand by name.
• Retargeting, meaning ads shown to women who already had the drink in the cart.
• Always-on social ads that follow people who would have bought from the site anyway.

You do not need an incrementality test for every campaign. A campaign is one group of ads with its own budget. You need a test when the spend is large and the platform number is the only argument for keeping it. A $2,000 test campaign can live on attribution (the credit rule). A six-figure retargeting line that has never been held out cannot.

How to measure incrementality

A holdout is the clean version of an incrementality test: some users or regions see no ads. Teams also call this lift testing. A geo lift (geo test) holds out regions instead of people.

Let's take a drink brand as an example. They sell a low-sugar sports drink for active women over 50, mostly from their own website. The drink brand pauses Meta (Facebook and Instagram ads) in Utah, Colorado, Idaho, Montana, and Wyoming for four weeks. Those states still sell $8,000 of drink (people who would have bought anyway). The rest of the country, scaled to the same population, sells $20,000. Extra sales are $12,000. The drink brand spent $10,000 on Meta in the exposed regions (scaled). Meta had claimed $40,000 of credited sales on that same $10,000 of spend. Extra is $12,000. Credited was $40,000. Both can be true. Only extra answers "should we spend more?"

The output of an incrementality test is lift, often written as iROAS (incremental return on ad spend: extra revenue divided by spend). Let's take a drink brand as an example. They sell a low-sugar sports drink for active women over 50, mostly from their own website. $12,000 extra on $10,000 spend is an iROAS of 1.2. It will usually look worse than in-platform ROAS (return on ad spend: credited sales divided by spend). That gap is the point.

Marketing mix modeling (MMM) estimates channel contribution without a person-level holdout. Use MMM for quarterly mix, tests to check it, and attribution for in-channel tactics. One model cannot do all three jobs. Lemonado is not an MMM product.

Write down the incrementality design: who was held out, for how long, and what counted as a conversion. A messy test still beats a confident last-click number, but only if you write the caveats down. Let's take a drink brand as an example. They sell a low-sugar sports drink for active women over 50, mostly from their own website. A one-week geo test on a drink people already buy every month can work. The same week on a product people take months to choose will lie.

Common incrementality mistakes

• Calling last-click lift. Let's take a drink brand as an example. They sell a low-sugar sports drink for active women over 50, mostly from their own website. The drink brand's Google last-click number is a credit rule, not extra sales.
• Running a one-week geo test on a product people take months to buy.
• Holding out a tiny, unlike region (one college town) and treating it as the whole United States.