Metrics and analysis
Incrementality testing: proving the ad caused the sale
How to prove advertising generated sales that would not have happened otherwise, with the three test designs available and what each one requires.
Also available in: Português · Español
Every ad platform reports a return, and all those returns added together usually exceed the company's real revenue. The explanation is simple: attribution measures contact, not cause. Incrementality is the attempt to measure cause.
The difference, in one sentence
Attribution answers: which ad did the person touch before buying?
Incrementality answers: would they have bought if the ad did not exist?
Different questions, and only the second decides whether budget is creating revenue or buying revenue that was already yours.
The classic example is remarketing. It shows the best return in the account, because it shows ads to people who already demonstrated intent. A good share of them would buy anyway. The attributed return is high; incrementality can be near zero.
The same applies to brand campaigns on search: they harvest intent created elsewhere. The mechanism is in last-click attribution.
The three possible designs
| Design | How it works | Precision | Cost |
|---|---|---|---|
| On and off | Pause the channel for a period and compare | Low | Low |
| Geographic | Advertise in some regions and not others | Medium | Medium |
| Control group | Part of the audience never sees the ad | High | High |
The third is the most correct and the least accessible: it needs the platform's own tooling and high volume. The first two are within reach of any operation.
The on-and-off test
The simplest and the one almost nobody runs.
- Pick the channel and a two-to-four week period.
- Pause it completely.
- Change nothing else. Not promotions, not prices, not another channel.
- Compare total revenue against the equivalent previous period.
- Resume and watch the recovery.
What informs you is not the paused channel's sales — those go to zero by definition. It is the company's total revenue. If it falls 8% when you pause a channel attribution claimed was responsible for 35% of sales, you just learned something expensive about that number.
The limitations are real: seasonality, competition and other channels move at the same time. That is why step 3 matters so much, and why two weeks is the minimum.
The geographic test
More precise and still accessible. You split the country or region into two comparable groups, advertise in one and not the other, and compare each group's revenue.
What makes it work:
- Comparable regions in sales history and profile, not only in size.
- Enough time to accumulate conversions in both groups.
- No other difference between them: same promotion, same price, same service.
- Revenue measurable by region in your own system, not in the platform.
The last item is what usually makes it impossible. If your system does not split revenue by region, the test cannot be read — and fixing that is more valuable than the test itself.
What you need before any test
Source tagged at signup. Without knowing where each sale came from, no design works. The standard is in UTM parameters for ads.
Total revenue accessible. The test reads the operation's revenue, not the ad panel's.
Conversion volume. Below a few dozen per period, the difference between groups does not separate from noise.
Discipline not to touch anything. The test dies the day someone uses the window to also try a promotion.
A worked example
An operation billing US$ 400,000 a month. Meta's attribution claims US$ 180,000, or 45% of revenue.
The test: remarketing paused for two weeks, nothing else changed.
| Two weeks before | Two test weeks | |
|---|---|---|
| Remarketing spend | US$ 12,000 | US$ 0 |
| Revenue attributed to remarketing | US$ 84,000 | US$ 0 |
| Total operation revenue | US$ 200,000 | US$ 188,000 |
Attribution said remarketing generated US$ 84,000. Switching it off, the operation lost US$ 12,000 of revenue.
Reading: the channel's incrementality is roughly 14% of what it claimed. The other 86% were sales that would have happened anyway, with remarketing serving as the last click before them.
What that does not mean: that remarketing should be switched off. It generated US$ 12,000 of incremental revenue on US$ 12,000 of spend — break-even, and it probably improves with a shorter window and lower frequency. What changes is the number that enters a comparison against other channels: US$ 12,000, not US$ 84,000.
Without that test, this campaign would be the account's champion in any report, and budget would flow to it.
What to do with the result
The number that comes out is an estimate of how much attributed revenue is real. It is almost never 100%, and almost never zero.
High incrementality. The channel is creating demand. Worth scaling, and the attributed return is understating it.
Medium incrementality. Normal for most channels. Use the factor to discount platform returns when comparing channels.
Low incrementality. The channel is harvesting what was already yours. It does not mean switching it off: it means the budget there has a ceiling and growing requires another channel.
The most common low-incrementality case is aggressive remarketing, and the fix is rarely switching it off — it is shortening the window and lowering the frequency. The broader view is in Meta Ads remarketing.
The mistakes that invalidate the test
Too short a period. A week captures neither the sales cycle nor normal variation.
Changing something else at the same time. A promotion, a price, stock or another channel moving contaminates everything.
Reading it from the platform panel. The paused channel's panel goes to zero. That is not a result, it is a definition.
Testing during a peak. A seasonal spike distorts any comparison.
Concluding from one test. One result is a measurement; a decision asks for repetition.
When testing is not worth it
A small operation. Without conversion volume, the result does not separate from noise, and you lost real revenue for a number that informs nothing.
A single channel. If you only have one, pausing it means pausing the business.
When the decision is already made. Testing to confirm what you decided is expensive theatre.
The cheap alternative for those cases: tracking the relationship between total spend and total revenue across months, which is what ROAS or TACoS shows with no test required.
Where to start
Pick the channel whose reported return you doubt most — usually remarketing or brand. Pause it for two weeks, changing nothing else, and compare total revenue against the two weeks before.
The result will not be precise, and it will be more honest than any platform report you have today. The document that records the before and after is on the paid media reporting page.
Frequently asked questions
What is incrementality?
The share of sales that only happened because the ad existed. Attribution says which ad someone touched before buying; incrementality says whether they would have bought anyway.
Why does attribution not answer that?
Because it credits contact, not cause. Remarketing to people already intending to buy shows an excellent return in attribution and can have incrementality close to zero.
What is the simplest test?
Turning a channel off for a period and comparing total revenue against the equivalent previous period. It is imprecise and informs more than any platform report.
Do I need a large budget to test?
You need conversion volume, not a large budget. Below a few dozen conversions per period, the difference between groups does not separate from noise.