Incrementality is the part of your sales that your marketing actually caused: the conversions that would not have happened if you hadn’t run the campaign. You measure it by comparing what happened with a counterfactual (what would have happened otherwise), usually a control group or a set of regions that didn’t see the ads, and the gap between the two is the incremental lift. That sounds like a technicality, but it’s the difference between the number your ad platform reports and the number your finance team should care about.
In my former life as the co-founder of a 50-person growth marketing agency that spent over $10 million a year on Facebook ads, I learned that this gap is where most arguments about marketing results start. A dashboard can show a healthy return while the bank balance tells a more modest story, and both can be accurate at the same time, because they’re answering different questions. The platform is answering “which conversions came after someone saw an ad?”, while the business wants to know “which conversions happened because someone saw an ad?” I’ll show you how to put a number on the second question with your own data.
What incrementality means: observed sales are not the same as caused sales
The idea at the heart of incrementality is that every impact claim has a missing half. You can observe the sales you made while a campaign was running. You can’t simultaneously observe the same business, in the same week, without that campaign. That second, unobservable quantity is the counterfactual, and incremental impact is the observed outcome minus the counterfactual outcome under some specified alternative (no campaign, a smaller campaign, or a different one).
The Vexpower course makes this concrete with its invented brand, Northstar. Northstar runs a campaign and records £135,000 of sales. If the business would have made £100,000 anyway, the campaign caused £35,000. If it would have made £120,000 anyway, the campaign caused only £15,000. The observed sales are identical in both cases; the causal claim changes because the counterfactual changes. This is why “the campaign drove £135,000 of sales” is never a complete sentence, and why the first question to ask of any marketing result is “compared with what?”
It’s worth being precise about the alternative, because two reasonable comparisons can produce two different numbers that are both correct. An experiment that switches a campaign off entirely estimates the effect of the campaign against nothing. An experiment that adds 20 percent to the budget estimates the effect of that extra 20 percent, which on a saturating channel will be smaller per dollar than the average. A marketing mix model reports a channel’s contribution against a modeled baseline, which is yet another comparison. None of these is “the” incrementality of a channel; each is the incrementality of a specific change, and you shouldn’t treat a marginal lift from a small budget increase as if it were automatically the average return on all historical spend.
Incremental lift, iROAS and the incrementality factor
Once you have a counterfactual, the vocabulary follows from simple arithmetic. Incremental lift is the outcome with the marketing minus the outcome without it. In a holdout test that’s usually expressed as conversion rates: if 0.30 percent of the test group converted and 0.24 percent of the control group did, the absolute lift is 0.06 percentage points and the relative lift is 25 percent (0.06 divided by 0.24). Multiply the absolute lift by the number of people in the test group and you have incremental conversions:
Incremental return on ad spend, or iROAS, is then the revenue those incremental conversions generated, divided by what you spent. It’s the same shape as the ROAS number in your ad platform, with one important difference: the numerator only counts revenue that wouldn’t have happened without the ads. If you want the profit view, use margin per conversion instead of revenue and compare the result with your breakeven ROAS rather than with 1.0.
The third number, which practitioners often call an incrementality factor (or incrementality multiplier), is incremental conversions divided by the conversions the platform attributed to the campaign over the same period. A factor of 0.48 says that for every 100 conversions the platform claims, about 48 were caused by the ads. Teams use the factor to discount platform-reported ROAS on similar campaigns they haven’t tested, which is a reasonable shortcut as long as you remember it was measured on one campaign, one audience and one time period, and can drift as any of those change.
There’s also a fourth number, marginal ROAS, which is the return on the next dollar rather than the average dollar. Incrementality tests that switch a campaign off tell you about the average; to decide whether to scale up, you need to know where you sit on the response curve, which is the subject of the marginal ROAS guide.
A counterfactual calculator for your own campaigns
The calculator below takes the three numbers you can read off any ad platform (spend, attributed conversions and value per conversion) and compares them with what a lift test found. If you have test and control conversion rates from a conversion lift study or your own holdout, use the first mode. If you don’t, use the second mode to see how sensitive your ROAS is to the share of attributed conversions that are really incremental, which is a useful exercise before you decide whether a test is worth running.
With the default numbers the platform reports a ROAS of 2.0, which looks comfortably profitable, but the test implies an iROAS of 0.96, meaning the campaign returned slightly less revenue than it cost. Notice that the test group had 3,000 conversions in total, of which the platform attributed 1,250 and only 600 were incremental. Those three numbers measure different things, and much of the confusion in marketing measurement comes from treating them as if they were interchangeable.
Why platform-reported conversions overstate incrementality
An ad platform’s attributed conversion is assigned by an attribution rule: someone saw or clicked an ad, and then converted within a window. That rule alone doesn’t tell you whether the conversion would have occurred without the advertising, and in practice there are a few predictable reasons the platform’s number runs higher than the incremental one. None of them requires anyone to be acting in bad faith; they’re built into how targeting and attribution work.
View-through credit for people who were going to convert anyway
View-through attribution gives a campaign credit when someone saw an ad (without clicking) and converted later. On a platform with enormous reach, a large share of your customers will have scrolled past one of your ads in any given week simply because they use the platform, so a view-through window will sweep up conversions that had nothing to do with the ad. The longer the window and the broader the reach, the more of your existing demand gets relabelled as ad-driven. Shortening the window reduces the problem but doesn’t solve it, because even a one-day view-through window can’t tell the difference between an ad that changed someone’s mind and an ad that happened to be on screen the day before they bought.
Retargeting people who would have bought anyway
Retargeting is the cleanest example of selection bias in marketing. The algorithm’s job is to find people likely to convert, and the people most likely to convert are the ones who were already on your checkout page. A retargeting campaign aimed at cart abandoners will report spectacular ROAS, because its audience was selected for being close to purchase, not because the ads did much persuading. Delivery optimization makes this worse in prospecting campaigns too: an algorithm rewarded for attributed conversions will learn to show ads to people who were going to convert, which is exactly what you’d ask it to do if attributed conversions were the goal and exactly what you wouldn’t if incrementality were. Meta now offers an “incremental attribution” setting that, in Meta’s words as reported by Social Media Today in September 2025, optimizes delivery using models that predict whether a conversion is caused by an ad. That’s a step in the right direction, but it’s still the platform’s own model of its own effect, so it’s a hypothesis to check with a test rather than a replacement for one.
Brand search that captures demand you’d have got for free
Paid search on your own brand name is the other classic case. When someone types your brand into Google, the brand ad is often the first thing they click, so it gets the credit, but many of those people would have clicked the organic listing directly underneath it. The best-known evidence here comes from eBay: in a large field experiment, Blake, Nosko and Tadelis found that brand keyword ads had no measurable short-term benefit, and that for non-brand keywords the effect was concentrated among new and infrequent users, while frequent users (who received most of the spend) would have bought anyway. That’s one company with an exceptionally strong brand, so it doesn’t mean brand search is never incremental (competitors bidding on your name change the picture, for instance), but it’s a good reason to test it rather than assume. If you want to practise this, the old Vexpower simulator Should we turn off brand ads? walks through a deprivation test on exactly this question.
Correlation-based methods tend to overstate it too
It would be convenient if a clever enough model could recover incrementality from observational data alone, and sometimes it can get close. However, the evidence suggests caution. Gordon, Zettelmeyer, Bhargava and Chapsky compared observational methods with large randomized experiments run on Facebook and found that the observational methods often failed to recover the effects the experiments measured, even with rich user-level data. The underlying reason is the same as with retargeting: the people who see ads are different from the people who don’t, and if the model can’t fully account for that difference, it attributes some of their pre-existing propensity to buy to the ads.
There’s also a noise problem. Lewis and Rao showed that the effect of advertising on an individual’s purchases is tiny relative to how much individual purchases vary anyway, which means you need very large experiments to measure returns precisely. That doesn’t make measurement hopeless; it means you should expect wide confidence intervals, plan tests with enough power, and be suspicious of any method that reports precise answers from small amounts of data. I’d rather show a range that honestly includes “not profitable” than a point estimate that pretends it doesn’t. For a fuller treatment of why correlation misleads in marketing data, see the guide to causal inference for marketers.
How to measure incrementality
There are three broad families of methods, and they’re complements rather than rivals. Each estimates a slightly different quantity under different assumptions, so the strongest measurement programs use more than one and investigate when they disagree.
Experiments
A randomized experiment is the most direct way to estimate a counterfactual: you randomly split people (or regions, or time periods) into a group that gets the marketing and a group that doesn’t, and because the split is random, the control group’s outcome is a fair estimate of what the test group would have done without the ads. User-level holdouts, such as Meta’s and Google’s conversion lift studies, randomize individuals. Geo experiments randomize regions, which works for channels you can’t target at the individual level, like TV or out-of-home, and captures offline sales. Each has its own design choices and pitfalls, which I cover in the guide to incrementality testing, and geo designs specifically in geo lift testing.
Marketing mix modeling
A marketing mix model estimates incrementality from aggregate data (usually weekly sales and spend by channel) by modeling sales as a baseline plus the contribution of each channel, with adjustments for carryover and diminishing returns. Its strength is coverage: one model gives you an estimate for every channel, including ones you can’t easily test, and lets you run budget scenarios. Its weakness is that the estimates rest on the model’s assumptions about what else drives sales, so a confounder you left out shows up as a biased channel effect. The guide to marketing mix modeling covers how these models work, and calibrating MMM with experiments covers how to tie them back to test results.
Quasi-experiments
When you can’t randomize, you can sometimes find a natural comparison: a region where a campaign launched later, a period when a channel went dark by accident, or a policy change that affected one market and not another. Methods like difference-in-differences and the synthetic control method build a counterfactual from the units that weren’t treated. They’re weaker than a true experiment because you have to argue that the comparison group would have tracked the treated group without the change, but they’re often far better than nothing, and they’re how most “we turned off brand search for a month” analyzes should be read.
Using incrementality without fooling yourself
The practical value of incrementality is that it changes decisions. A channel with a platform ROAS of 4 and an incrementality factor of 0.2 is earning less than one with a platform ROAS of 2 and a factor of 0.8, and you’d only know that from a test. However, an incrementality number is only as good as its definition, and the most common mistakes I see aren’t statistical; they’re mismatches between what was measured and what’s being decided.
If you want to work through these ideas end to end on one invented business, the Vexpower marketing mix modeling course starts with the define-incrementality lesson and carries the same counterfactual thinking through to calibration and budget decisions. The MMM vs multi-touch attribution guide is also worth reading next, because attribution is where most teams’ numbers come from today, and knowing what it can and can’t tell you is the first step towards measuring what your marketing actually caused.
