Resources
ROAS vs incrementality: which metric for which decision
Short answer
ROAS answers "how much attributed revenue did I get per dollar spent?" Incrementality answers "how much of that outcome happened because of the advertising and wouldn't have happened anyway?" ROAS is useful for tactical optimization inside a channel; incrementality is the metric that should govern budget allocation across channels, countries and funnel stages. Using ROAS to decide where the next dollar goes is the most expensive mistake in performance marketing.
The conflict, in one sentence
The channels with the highest ROAS are usually the least incremental.
Branded search, retargeting and demand-capture formats post excellent ROAS precisely because they harvest demand that already existed. The ad appears in front of someone already looking for you, takes the credit, and your dashboard congratulates you.
The canonical eBay experiment (Blake, Nosko and Tadelis, Econometrica, 2015) is the reference case: brand keyword ads showed no measurable short-run benefit. In non-brand, positive impact concentrated among new and infrequent users, while frequent users absorbed a large share of spend without changing behavior.
Running the other direction, other experiments show display campaigns can lift total purchases — including offline and click-free purchases — even when in-platform ROAS fails to reflect it.
The pattern: ROAS systematically overvalues the bottom of the funnel and undervalues the top.
Definitions and formulas
| Metric | Formula | What it answers | Caveat |
|---|---|---|---|
| ROAS | Attributed revenue / ad spend | Attributed efficiency | Useful for operating, not causal on its own |
| Target ROAS | (Conversion value / spend) × 100% | Value-based bidding goal | Set too high, it throttles volume |
| Absolute lift | Y_treatment − Y_control | Total causal effect | Requires a comparable counterfactual |
| Relative lift | (Y_treatment − Y_control) / Y_control | Percentage causal effect | The comparable metric across campaigns |
| Incremental value | Value_treatment − Value_control | Genuinely new revenue | Basis for iROAS |
| iCPA | Spend / incremental conversions | Causal cost per conversion | Always higher than attributed CPA |
| iROAS | Incremental value / incremental cost | Causal return | The budget allocation metric |
| mROAS | Δ revenue / Δ spend | Return on the next dollar | What matters for scaling, not average ROAS |
The distinction between average ROAS and marginal ROAS (mROAS) deserves its own attention. Optimal budget allocation is defined at the margin: what past spend returned matters less than what the next dollar will return. A channel averaging 4x ROAS can have an mROAS of 1.2x if it's already saturated.
Causal measurement methods
| Method | How it works | When to use it | Minimum requirement |
|---|---|---|---|
| User-level holdout | Randomized control group that sees no ads | Channels with user identity (Meta, Google) | Enough conversion volume for power |
| Geo-experiment | Treated regions vs. control regions | Identity-less channels, TV, OOH, or where holdouts aren't available | Enough comparable, stable geos |
| Ghost ads / PSA | Control group sees a neutral ad | When you need to control for audience selection | Platform support |
| Switchback | On/off across time periods | Small budgets, fast decisions | Low within-period seasonality |
| MMM | Time-series model over spend and outcomes | Macro allocation across channels and planning | 2–3 years of history and spend variation |
The current technical consensus: MMM and experiments don't compete — they calibrate each other. Modern frameworks (Google's Meridian, Meta's Robyn) are explicitly designed to take experimental results as model priors. An MMM with no experiments anchoring it is elegant curve fitting.
The constraint nobody warns you about: statistical power
Here's the uncomfortable one. Lewis and Rao ("The Unfavorable Economics of Measuring the Returns to Advertising," QJE, 2015) showed that estimating advertising ROI with useful precision requires enormous samples, because the variance in individual consumer spending dwarfs the effect of the ad.
In practice this means two things:
- An underpowered test isn't "a small test" — it's noise. A non-significant result does not prove the channel doesn't work.
- Start with the channel where the most budget is at stake, not the easiest one to test. One good test on 40% of your spend beats five tests on 3%.
Calculate power before running anything. If detecting the minimum effect you care about would take six months, the test isn't viable and you should decide via MMM or explicit judgment instead.
What it looks like in practice
A typical portfolio after its first round of tests:
| Channel | Reported ROAS | Estimated iROAS | Read |
|---|---|---|---|
| Branded search | 12.0x | 1.4x | Mostly existing demand. Reduce, don't eliminate |
| Retargeting | 8.5x | 2.1x | Overvalued, but incremental on cold cohorts |
| Prospecting social | 2.2x | 2.0x | ROAS was already close to honest |
| Display / video | 0.9x | 1.6x | Undervalued by last-click attribution |
Note the asymmetry: closing channels lose a lot when you switch to a causal metric; discovery channels gain. That asymmetry is why reallocating budget on ROAS alone tends to narrow the funnel until growth stalls.
Important caveat: the figures in this table illustrate the pattern — they are not benchmarks. Your iROAS depends on your category, brand awareness, purchase cycle and saturation level.
How ADEX approaches it
We run measurement in three layers, kept deliberately separate:
- Operating layer (daily). Platform metrics, CPA, ROAS, spend by campaign. Used to manage, not to set budget.
- Experimental layer (quarterly). One lift test or geo-experiment per quarter on the channel with the most budget at stake. Results documented with confidence intervals, not just point estimates.
- Allocation layer (semi-annual). Budget reallocation driven by iROAS and mROAS, not attributed ROAS.
ADEX has managed $50M+ in ad spend across 18 countries, and most of the value we add isn't campaign optimization — it's stopping a team from moving $200K into a channel that only looked profitable.
When you don't need incrementality
Be honest about the threshold:
- Under roughly $20–30K/month in a channel: the test won't have power, and the holdout's opportunity cost is proportionally high.
- A single channel: there's no allocation decision to make.
- Unvalidated product or offer: first confirm someone wants to buy.
- Broken tracking: fix measurement before measuring causality. An experiment on dirty data produces a clean, false number.
FAQ
What is incrementality in marketing? The causal effect of advertising: how many conversions happened because of the ad and wouldn't have happened otherwise. Measured by comparing an exposed group against a comparable control group.
How much budget do I need for a lift test? It depends on your baseline conversion rate and the minimum effect you want to detect, not on a fixed figure. As practical guidance, below roughly $20–30K/month in the channel, most tests lack the power to give a useful answer.
Doesn't a holdout cost me sales? Yes — that's the price of the experiment. The control group is money you don't spend on people who might have converted. That cost is usually far smaller than running a channel with a true iROAS of 1.0x for a year.
Does MMM replace experiments? No. An MMM without experiments to calibrate it is prone to unreliable results. Current Google and Meta frameworks incorporate experimental results as priors for exactly this reason.
Why is my branded search ROAS so high? Because it captures demand that already existed. A large share of those users would have found you through the organic result anyway. It's the number-one candidate for an incrementality test.
How often should I run tests? One well-designed test per quarter on your largest-spend channel is a sustainable cadence for most teams. Higher frequency rarely justifies the operational cost.
Next step
If you've never validated your largest-spend channel, that's your first test. Design my first incrementality test →
Related reading
- Broken attribution and CAC
- Cheap leads vs real pipeline
- Service: performance marketing agency
- Service: marketing data dashboards
External sources
- Blake, Nosko and Tadelis (2015). Consumer Heterogeneity and Paid Search Effectiveness. Econometrica. https://doi.org/10.3982/ECTA12423
- Lewis and Rao (2015). The Unfavorable Economics of Measuring the Returns to Advertising. Quarterly Journal of Economics. https://doi.org/10.1093/qje/qjv023
- Gordon, Zettelmeyer, Bhargava and Chapsky (2019). A Comparison of Approaches to Advertising Measurement. Marketing Science. https://doi.org/10.1287/mksc.2018.1135
- Google documentation on Conversion Lift, geo experiments and Meridian; Meta documentation on Conversion Lift and Robyn.
Numeric tables in this article illustrate the described pattern and are not industry benchmarks.