TNOMADS / Services / Measurement & Analytics

Measurement & Analytics

Incrementality — the only proof
that ads actually work.

71% of advertisers rank it as their #1 retail media KPI (ANA, January 2024). Here's how to measure it, interpret it, and act on it.

What is incrementality?

Incrementality is the share of sales caused by advertising that would not have happened otherwise. It is measured by comparing an exposed group against a held-out control group — through geo tests, audience holdouts or matched-market experiments — rather than by reading platform-attributed conversions, which credit sales that would have occurred anyway.

Method: Controlled experimentDesigns: Geo, audience holdout, PSAOutput: Incremental lift and iROASTypical duration: 4–8 weeks

What is incrementality?

The only metric that proves your ads are working.

Incrementality measures the additional sales generated by advertising compared to what would have happened without it. It answers the fundamental question every advertiser must ask: did my ad spend create new demand, or did it capture sales that would have occurred anyway?

A campaign can show a strong 5x ROAS and still be almost entirely non-incremental — if most of those attributed sales would have happened organically through branded search, repeat purchase, or existing category awareness. This distinction is what separates efficient advertising from expensive reporting.

In a January 2024 survey by the Association of National Advertisers, 71% of advertisers ranked incrementality as their most important retail media KPI.[1] US retail media spend reached $60.32 billion in 2025 and is forecast at $71.09 billion in 2026, which makes the need to prove causal impact more urgent than ever. Adoption is following: around 52% of US brand and agency marketers now use incrementality testing[2] — yet roughly 75% say their measurement systems still lack the speed, accuracy or trust they need.[3]

iROAS

Incremental ROAS

iROAS = incremental revenue ÷ ad spend. Unlike standard ROAS, iROAS removes conversions that would have occurred organically — giving you the true return on each advertising dollar.

RCT

Randomised Controlled Trials

The gold standard. Randomly assign audiences to test (exposed) and control (not exposed) groups. The difference in conversion rates represents true incremental lift. Ghost bidding is the preferred RCT methodology in retail media — it maintains targeting without serving impressions to the control group.

GEO

Geo-Based Holdout Tests

Divide target markets geographically into test and control regions. Run campaigns in test markets only. Compare sales velocity between regions to calculate market-level incremental lift — useful when audience-level control is not available.

MMM

Marketing Mix Modelling

Statistical modelling of historical spend and sales data across all channels to estimate incremental contribution. Best for long-term budget allocation decisions — less useful for campaign-level optimization.

Why it matters

The gap between ROAS and reality.

THE PROBLEM

Attribution Overclaims Real Impact

Standard last-click and even multi-touch attribution models credit ads for conversions they did not cause. A loyal customer who would have reordered regardless of seeing your ad still gets attributed to the campaign. The size of that gap is not a fixed number — it varies enormously by channel, category and audience, which is precisely why it has to be measured for your account rather than assumed from a published figure.

Last-Click BiasHalo InflationOrganic Cannibalisation
THE SOLUTION

Controlled Experiments Prove Causation

Incrementality testing isolates causation from correlation. Published iROAS results vary by an order of magnitude across advertisers, which tells you something important: some campaigns create substantial new demand while others run largely on recycled demand, and reported ROAS cannot tell you which of the two you are buying. Only measurement that controls for organic behaviour can distinguish between them.

Ghost BiddingHoldout GroupsCausal Lift
THE OUTCOME

Smarter Budget Allocation

The practical value of a geo test is that it reorders your channels. Platforms that look comparable on attributed ROAS routinely diverge once each is measured against a held-out control, because attributed return is partly a function of how close a channel sits to the transaction. Reallocating budget on the strength of that reordering is where incrementality testing pays for the revenue it costs to hold out.

Budget ReallocationChannel MixProven ROI
71%of advertisers rank incrementality as their number one retail media KPI (ANA, January 2024) [1]
~52%of US brand and agency marketers now use incrementality testing, up from niche status two years earlier [2]
~75%of marketers say their measurement systems lack the speed, accuracy or trust they need [3]
50%increase in retail media networks offering MMM access between Q1 and Q3 2024 [4]

The TNOMADS approach

We build incrementality into the reporting framework, not as a one-off test.

🔬

Always-on iROAS tracking

We build AMC queries and reporting frameworks that track new-to-brand rate, halo attribution, and media overlap on an ongoing basis — so incrementality insight informs every weekly optimization decision, not just quarterly reviews.

🗺

Geo holdout test design

For brands with sufficient scale, we design and execute geo-based holdout tests that measure market-level incremental lift — controlling for seasonality, distribution changes, and organic demand shifts.

💡

Separate promoted vs. total ROAS

We maintain two ROAS metrics for every campaign: same-SKU ROAS (for bid and campaign optimization) and total ROAS including halo (for budget decisions) — because confusing the two leads to systematic misallocation.

The vocabulary

What are the core concepts in incrementality measurement?

Incremental sales
Sales that would not have occurred without the advertising. The difference between what happened and what would have happened anyway — which is why it can only be estimated by comparison, never read from a report.
Attributed sales
Sales the platform credits to an ad based on its attribution rules. Always higher than incremental sales, because attribution cannot distinguish causing a purchase from being present before one.
iROAS
Incremental return on ad spend — incremental revenue divided by ad spend. The number that answers the question executives are actually asking, and typically far lower than reported ROAS.
Holdout group
A randomly selected group deliberately not exposed to the advertising. The comparison against this group is the measurement; without a holdout there is no causal claim, only correlation.
Geo test
Splitting geographies into treatment and control markets. The most practical design in retail media, because media can usually be controlled by geography even where it cannot be controlled by user.
Matched markets
Control geographies selected to resemble treatment geographies on the outcome metric before the test begins. Poor matching is the most common reason a geo test produces an unusable result.
PSA test
Serving a public service announcement to the control group instead of nothing, so both groups experience an ad. Cleaner in principle, more expensive in practice, since you pay for impressions with no commercial return.
Statistical power
Whether the test can detect an effect of the size you care about. Underpowered tests are common and dangerous, because they return ‘no significant lift’ for campaigns that were genuinely working.
Marketing mix modelling (MMM)
Statistical modelling of aggregate marketing inputs against business outcomes. Broader scope than experiments but far lower granularity, and best used to calibrate rather than replace testing.
Clean room
Privacy-safe environments such as Amazon Marketing Cloud where event-level exposure and conversion data can be analysed to support experimental measurement.

Comparisons

How does this compare to the alternatives?

Attributed ROAS vs incremental ROAS

The distinction that determines whether a media budget survives its first serious review.

Attributed ROASIncremental ROAS
What it measuresSales the platform credits to the adSales the ad actually caused
SourcePlatform reporting, available immediatelyControlled experiment, requires design and time
Typical magnitudeHigherLower, sometimes dramatically
Most inflated forBranded terms, retargeting, high-frequency buyers—
Right useIn-flight optimisation and pacingBudget decisions and channel-level investment cases
When to use whichUse attributed ROAS to run campaigns day to day — it is available immediately and directionally useful for pacing. Use incremental ROAS to decide how much to spend and where. Treating attributed ROAS as if it were incremental is how brands end up paying to reach customers who were already buying.

Geo test vs audience holdout vs MMM

Three approaches, with genuinely different costs, timelines and answers.

Geo testAudience holdoutMMM
What it needsGeographic media control and matched marketsPlatform support for randomised exclusionTwo-plus years of clean historic data
TimelineFour to eight weeksFour to eight weeksWeeks to months to build
GranularityChannel or campaign levelAudience or campaign levelChannel level, aggregate
Main weaknessRegional differences can confound resultsNot supported on every platform or formatLow granularity; cannot answer tactical questions
CostHeld-out revenue in control marketsHeld-out revenue in control groupAnalyst time and data engineering
When to use whichGeo tests are the practical default in retail media because media can usually be controlled geographically. Audience holdouts are cleaner where the platform supports them. MMM answers whole-business allocation questions that neither experiment can, but cannot tell you which campaign to change on Monday.

Incrementality testing vs new-to-brand as a proxy

What to do when you cannot yet afford a proper test.

Incrementality testNTB as proxy
RigourCausal, if properly designedDirectional only
CostHeld-out revenue plus analyst timeNone — already in your reporting
What it tells youHow much revenue the ads causedWhether you are reaching new customers
LimitationRequires scale and design disciplineSays nothing about whether those customers needed the ad
When to use whichNew-to-brand is not incrementality and should never be presented as it. But for brands without the scale to run a powered test, tracking NTB as a directional trend is far better than assuming attributed ROAS is causal. Start there and graduate to testing when volume allows.

Real results

What does this work look like in practice?

Industry · Enterprise Consumer Brand

Moving beyond ROAS

Challenge

The business relied heavily on last-click ROAS despite investing across multiple retail media channels.

Our approach

  • Analysed attribution models
  • Reviewed Brand Halo performance
  • Evaluated New-to-Brand acquisition
  • Designed incrementality measurement frameworks

Business impact

The client adopted broader business performance metrics instead of relying solely on attributed ROAS.

Case studies are presented by industry rather than by client name. Figures are drawn from live account analysis. Engagements marked prior agency engagement were delivered by Ana Perez Ibarz in a previous agency role; the work and results are hers, the client relationships were the agency's. TNOMADS does not identify clients or publish client performance data without written consent.

Frequently asked questions

Common questions, answered directly.

The share of sales caused by advertising that would not have happened otherwise. It is measured by comparing an exposed group against a held-out control group, not by reading platform-attributed conversions.
Because attribution credits sales that would have happened anyway. If a shopper was already going to buy your product and clicked an ad on the way, the platform records that as advertising-driven revenue. Attribution measures presence, not causation.
Incremental return on ad spend — incremental revenue divided by ad spend. It is nearly always lower than reported ROAS, and it is the figure that actually answers whether the spend was worth making.
Through controlled experiments. The three main designs are geo tests splitting markets into treatment and control, audience holdouts excluding a randomised group from exposure, and PSA tests serving a placebo ad to the control group.
Splitting geographies into treatment markets that receive the advertising and control markets that do not, then comparing sales outcomes. It is the most practical design in retail media because media can usually be controlled by geography.
A randomly selected group deliberately excluded from exposure, used as the comparison baseline. Without a holdout you have correlation, not causation — there is no way to know what would have happened otherwise.
Typically four to eight weeks of active testing, plus design time beforehand and analysis afterwards. Shorter tests rarely accumulate enough conversion volume to detect an effect with confidence.
The direct cost is the revenue given up by holding out a control group, plus analyst time for design and analysis. That opportunity cost is real and should be budgeted for explicitly rather than discovered afterwards.
It depends on the size of the decision. If you are deciding whether to continue a six- or seven-figure annual investment, the cost of a test is small against the cost of getting that decision wrong. For small budgets the test may cost more than the decision is worth.
Power is the test's ability to detect an effect of the size you care about. Underpowered tests routinely return no significant lift for campaigns that were genuinely working, and brands then cut effective channels on the strength of a badly designed experiment.
Enough conversions in both groups for the difference to be detectable above normal variation. That depends on your baseline conversion volume and the size of the effect you expect. Calculate required sample size before running the test, not after.
Consider whether the test was powered to detect the effect, whether the control group was genuinely comparable, and whether the measurement window captured delayed effects. A null result from an underpowered test is not evidence of no effect.
Typically branded search, retargeting of recent site visitors and high-frequency exposure to existing loyal customers — all of which reach people already likely to buy. They still have defensive value; they simply should not be credited with creating demand.
Usually upper-funnel activity reaching audiences who have not encountered the brand — streaming TV, prospecting display, new-audience social. These also report the worst attributed ROAS, which is why they get cut first in accounts that measure only attribution.
MMM statistically models aggregate marketing inputs against business outcomes over time. It covers all channels including offline, but at low granularity. Experiments give sharper answers on narrower questions; MMM gives broader answers on blunter ones.
Yes, and mature programmes do exactly that. Experiment results are used to calibrate and validate the model, and the model then extends those findings to channels and periods where experiments were not run.
Increasingly, yes — a growing number offer lift studies or MMM access. Bear in mind that a network measuring its own incrementality has an obvious interest in the result, which is why independent design carries more weight in a boardroom.
AMC provides the event-level data to analyse a properly designed test, but it does not create the experimental design. Analysis without a holdout is still correlation. See Amazon Marketing Cloud.
No, and conflating the two is common. NTB tells you a purchaser was new to your brand; it does not tell you whether the advertising caused them to buy. It is a useful directional proxy and nothing more.
Annually for major channels, and additionally whenever media mix changes materially, when a large budget shift is being considered, or when reported performance and business results have diverged and nobody can explain why.
Be explicit that your reported ROAS overstates true return, track new-to-brand as a directional signal, watch total business metrics such as TACoS and total revenue growth alongside platform metrics, and avoid making irreversible decisions on attributed figures alone.
State the design, the confidence interval and the limitations alongside the headline number. Results presented as a single point estimate with no error bars invite exactly the challenge that undermines them. See retail media strategy for how findings translate into allocation.

Why TNOMADS

Why measure incrementality with TNOMADS?

13+years of measurement work across retail media platforms
Honestlimitations stated, not buried — including when not to test
AMCclean room analysis written in-house to support test design
1:1the person designing the test also runs the media

About the author

Who wrote this page?

Ana Perez Ibarz

Senior Retail Media Consultant

Amazon AdsAmazon DSPAmazon Marketing Cloud Walmart ConnectInstacartRetail Media Strategy

Last reviewed 16 August 2026

Ana Perez Ibarz is the founder of TNOMADS Consulting and has spent 13+ years in retail media and performance marketing. She manages live campaign operations across Amazon Ads (Sponsored Products, Sponsored Brands, Sponsored Display, DSP and Amazon Marketing Cloud), Walmart Connect, Instacart, Kroger Precision Marketing, Target Roundel, Loblaws Advance, DoorDash, Criteo, Google Ads and Meta.

On measurement, Ana builds the reporting architecture that makes incrementality testing possible — separating same-SKU from halo and off-catalogue revenue, establishing geo-level analysis, and writing the clean room queries behind it. Where a brand lacks the volume for a properly powered test, this page says so rather than selling a test that cannot answer the question.

She works as a subcontracted specialist for agencies as well as directly with brands, and writes on retail media measurement, incrementality and commerce media architecture. Based in Granada, Spain; operating across North America and Europe.

More about Ana and TNOMADS →

Sources

Where do these figures come from?

A note on these sourcesPublished iROAS results vary by an order of magnitude across advertisers and studies, and much of the most-quoted incrementality research comes from vendors with a commercial interest in the finding. Treat single-source figures — including favourable ones — as directional. The only incrementality number that should drive your budget is one from a test designed against your own account.

Related services

What should you read next?

Amazon Marketing Cloud

The event-level data layer that supports test analysis on Amazon.

Retail Media Strategy

How incrementality findings translate into budget allocation.

Amazon DSP

The channel most often cut on attributed ROAS and vindicated on incrementality.

Retail Media Audit

Where measurement gaps are usually first identified.

Amazon PPC

Branded search is typically the least incremental line in the account.

Walmart Connect

Omnichannel attribution raises its own incrementality questions.

Ready to measure what actually works?

Stop optimizing for ROAS. Start optimizing for growth.

Book a Discovery Call