Quick answer: Incrementality is the share of conversions, leads or revenue that a marketing campaign actually caused, measured by comparing an exposed test group against a withheld control group. Any sale the control group makes anyway is not incremental, even if the ad platform claims credit for it.
Every ad platform reports conversions it was involved with. Incrementality asks a harder question: how many of those would have happened anyway? A customer who searched your brand name, clicked your ad, and bought would probably have bought through the organic result too. That sale sits in your dashboard as a win, but it added nothing.
The gap is usually large. Across 663 large-scale randomised experiments run on Facebook, the median true lift on lower-funnel outcomes such as purchases was just 6%, while standard observational methods missed the experimental answer by a median of 62 percentage points on those same outcomes (Gordon, Moakler and Zettelmeyer, Marketing Science, 2023). Reported performance and real performance are not the same number.
Why incrementality matters
Budget decisions are made on reported numbers. If the reported numbers include sales you were going to get for free, you scale the wrong channel and starve the one doing the actual work.
The classic case is branded search. In a series of large field experiments at eBay, brand-keyword ads showed no measurable short-term benefit: when the ads were switched off, that traffic simply arrived through the free organic listing instead (Blake, Nosko and Tadelis, Econometrica, 2015). A campaign can show a strong ROAS and still be worth zero, because ROAS measures credit claimed while incrementality measures behaviour changed.
The problem got bigger, not smaller, after iOS 14 and third-party cookie decay. Attribution now leans on modelling to fill gaps in tracking, and every platform models in its own favour. Retargeting and Performance Max are the usual suspects in Indian accounts: both are very good at finding people already heading to checkout and billing you for the introduction.
Incrementality is also the only honest way to settle a channel argument. When Google and Meta together claim more revenue than your accounting software recorded, no attribution model will resolve it. A holdout test will.
Incrementality vs attribution vs MMM
These three answer different questions and are not substitutes for each other.
| Attribute | Attribution | Incrementality | MMM (Marketing Mix Modelling) |
|---|---|---|---|
| Question it answers | Which touchpoints did the buyer see? | Did the ads cause the sale? | How does total spend across channels move revenue? |
| Method | Credit rules applied to tracked journeys | Controlled experiment, test vs holdout | Statistical regression on historical data |
| Proves causation? | No | Yes | Correlational, improved when calibrated by tests |
| Speed | Real time | 2 to 6 weeks per test | Needs 2 to 3 years of data |
| Cost | Included in the platform | Lost revenue from the holdout | Analyst time or a paid tool |
| Best used for | Daily optimisation and pacing | Big budget calls: scale, cut, or reallocate | Annual planning and offline media |
Use attribution to steer week to week, use incrementality to decide where the money goes, and use MMM to plan the year. Treat attribution as a hypothesis and incrementality as the test of it.
How to measure incrementality
Every method comes down to one comparison: a group that saw the ads against a comparable group that did not.
Incremental lift % = (Test conversion rate – Control conversion rate) / Control conversion rate
Incremental conversions = Test conversions – Control conversions (after scaling both groups to the same size)
iROAS = Incremental revenue / Ad spend
The three test designs worth knowing
- Platform holdout (user-based). Google Ads Conversion Lift and Meta Conversion Lift split your audience into an exposed group and a holdback group and report incremental conversions and iROAS inside the platform. Google defines iROAS as incremental conversion value divided by incremental cost, so an iROAS of 2 means every ₹1 invested produced ₹2 of value that would not otherwise have existed (Google Ads Help). Easiest to run, but each platform only grades its own homework.
- Geo holdout. Switch ads off in a matched set of cities or states and keep them on elsewhere, then compare total sales. Google supports this as Conversion Lift based on geography across App, Demand Gen, Display, Video, Search, Shopping and Performance Max campaigns (Google Ads Help). This is the design to use in India, where a Pune-vs-Nagpur or Bengaluru-vs-Hyderabad split is easy to construct and it captures offline and phone sales too.
- Scaled spend test. Instead of going to zero, cut spend by 40% to 50% in the test markets. Less revenue at risk, but you need more volume to read the result.
Worked example, in rupees
A Mumbai D2C brand runs a 4 week geo holdout on Meta. Ads stay on in the test markets and go dark in matched control markets of the same size.
- Orders in test markets: 4,000
- Orders in control markets: 3,400
- Incremental orders = 4,000 – 3,400 = 600
- Incremental lift = 600 / 3,400 = 17.6%
- Average order value: ₹1,500, so incremental revenue = ₹9,00,000
- Ad spend in test markets: ₹3,00,000
- iROAS = 9,00,000 / 3,00,000 = 3x
Meta’s dashboard, meanwhile, claimed 1,200 orders worth ₹18,00,000, a reported ROAS of 6x. So half the orders Meta took credit for were genuinely caused by the ads and half were already coming.
Now the part most reports skip. If this brand runs a 30% gross margin, its break-even ROAS is 1 / 0.30 = 3.33x. At 6x reported ROAS the campaign looks like a winner, and at 3x iROAS it is slightly under break-even. That does not automatically mean switch it off. It means the real decision is now visible: fix creative and audience overlap, cut the retargeting that is double-counting, or accept the loss on first order because repeat purchase makes it back. All three are defensible. Scaling blindly on the 6x is not.
Common mistakes
- Running the test during a festival or sale. A holdout that overlaps Diwali, Big Billion Days or a site-wide discount measures the promotion, not the ads. Pick a boring 4 week window with stable pricing, or the seasonality will swamp your signal.
- Leaking the holdout. If the control group still gets your email, WhatsApp broadcasts, YouTube ads or an influencer post, it is not a clean control. Geo tests leak too: neighbouring districts and OTT or connected TV delivery spill across boundaries.
- Testing with too little volume. Small tests produce wide confidence intervals, and a result of “somewhere between minus 5% and plus 30% lift” tells you nothing. Check the confidence interval before you act, and if it crosses zero, treat the test as inconclusive rather than as proof.
- Reading a flat result as proof the channel is useless. No detected lift often means the creative or offer is not persuasive, or the test was underpowered. It is a prompt to change the input and retest, not a verdict on the channel.
- Treating one test as permanent truth. Incrementality moves with creative fatigue, audience saturation, competitor spend and season. A number from March does not govern your October budget. Re-test the big channels twice a year.
Frequently asked questions
What is incrementality in marketing?
Incrementality is the portion of conversions, leads or revenue that a campaign actually caused, rather than the portion it received credit for. It is measured by comparing an exposed test group with a withheld control group and taking the difference. Sales the control group makes anyway are not incremental.
What is the difference between incrementality and attribution?
Attribution assigns credit for conversions that already happened, using rules such as last click or a data-driven model. Incrementality runs an experiment to find out whether those conversions would have happened without the ads. Attribution is fast and correlational; incrementality is slower and causal.
How do you calculate incremental ROAS (iROAS)?
Divide incremental revenue by the ad spend that produced it. If a holdout test shows 600 extra orders at an average order value of ₹1,500, that is ₹9,00,000 of incremental revenue; on ₹3,00,000 of spend the iROAS is 3x. Compare that figure against your break-even ROAS, which is 1 divided by your profit margin.
How long should an incrementality test run?
Most tests need 2 to 4 weeks at minimum so that delayed conversions land inside the measurement window, and longer if your sales cycle runs past a month. Avoid promotional periods, keep creative and budgets stable throughout, and let the test finish rather than stopping it the moment the numbers look favourable.
Are branded search ads incremental?
Often only partly. Large field experiments at eBay found that brand-keyword ads produced no measurable short-term sales benefit, because traffic shifted to the free organic listing when the ads were paused. That result will not hold for every brand: if competitors bid on your name or your organic listing is weak, branded ads can be genuinely incremental. Test it on your own account before cutting or keeping the budget.
Can small advertisers run incrementality tests?
Yes, but usually with geo tests rather than user-level holdouts. User-level lift studies need large audiences to reach statistical significance, while a geo holdout compares regional sales totals and works at lower volume. If your account converts fewer than a few hundred times a month, run longer tests on your biggest channel rather than many small ones.
How Uginix helps
Uginix builds a holdout test into the account before recommending a scale-up, because scaling a channel that is claiming credit rather than creating demand is the fastest way to burn a budget. We run geo and platform lift tests across Google and Meta, reconcile the results against real revenue, and report incremental ROAS alongside the platform number so you can see the gap. This is the same discipline behind our view on how much control to hand Meta Advantage+ and on what Performance Max is really doing with your budget. If a channel does not clear the bar, we tell you and reallocate. Not sure where your ad spend is leaking? That is exactly what our free growth audit finds.
Related terms
- Attribution: the set of rules that assigns credit for a conversion across the touchpoints a buyer saw.
- iROAS (Incremental Return on Ad Spend): incremental revenue divided by ad spend; the causal counterpart to reported ROAS.
- Holdout group: the randomly selected audience or set of geographies deliberately withheld from seeing your ads, used as the control.
- Conversion Lift: the native experiment tools in Google Ads and Meta Ads Manager that run a holdout and report incremental conversions.
- Cannibalisation: paid conversions that simply replace organic ones, the main reason attributed conversions exceed incremental ones.
- MMM (Marketing Mix Modelling): a top-down statistical model of how total spend by channel moves revenue, best calibrated with lift test results.
Sources: Google Ads Help, About Conversion Lift, Google Ads Help, Conversion Lift based on geography, Gordon, Moakler and Zettelmeyer, Marketing Science (2023), Blake, Nosko and Tadelis, Econometrica (2015).