IAB Incrementality Guidelines.
An independent, section-by-section decode of the newest canonical document in commerce media measurement — what it grades, what it recommends, and the seven things it stays silent on.
The Guidelines for Incremental Measurement in Commerce Media are a joint IAB and IAB Europe paper, final November 3, 2025, that defines incrementality as the causal impact of marketing, sorts measurement methods into four families by causal strength, and sets three requirements any method must meet to claim causality.
Last validated July 13, 2026 · Maintained by Evgeny Popov
The most consequential sentence in this paper is a classification, not a rule: platform-reported incrementality is a weak hybrid proxy — filed there by a body whose task force includes the platforms.
Fast read
- What it is
- A joint IAB and IAB Europe guidance paper — Guidelines for Incremental Measurement in Commerce Media, final November 3, 2025 — that defines incrementality, sorts measurement methods into four families graded by causal strength and holistic scope, and sets three requirements a method must satisfy to claim causality.
- What it covers
- The canonical incrementality definition, the four methodology families (experiment-based, model-based counterfactual, econometric, hybrid proxies), a nine-use-case fit matrix, and a technical deep dive on what makes a model causal.
- What it is not
- Not a compliance standard. It self-describes as a paper, carries no conformance clause, no audit or certification mechanism, and no MRC involvement — unlike the January 2024 IAB/MRC Retail Media Measurement Guidelines.
- The buried headline
- The paper files platform-reported incrementality in its weakest tier — a hybrid proxy relying on platform-specific attribution or modeling rather than an independent, true control group.
- Best for
- Brand, agency, retail media network, commerce platform, and measurement leaders deciding what an incrementality claim is worth — and what to write into the contract, since the paper mandates nothing.
- Best next read
- Retail / Commerce Media Measurement for the full five-document corpus this paper sits inside; Research & Measurement Science for the evidence layer.
What this document is — and is not.
Released November 3, 2025 as a final joint publication of IAB (US) and IAB Europe, developed by the IAB Commerce Board, IAB's Task Force on Incrementality, and IAB Europe's Retail Media Committee and Retailer Council, acknowledging prior work by Alliance Digitale (France) and BVDW (Germany). It is a 13-page designed PDF with roughly seven pages of content — an introduction, one method table, a two-page use-case matrix, a technical deep dive, and a conclusion; pages 8–13 are contributor rosters. It self-describes as a paper. No public-comment draft stage is documented for it, and no MRC involvement — two deliberate contrasts with the January 2024 IAB/MRC Retail Media Measurement Guidelines, which had both.
The definition it exists to fix
The paper's boxed definition makes incrementality the causal impact of marketing: the additional business outcomes a campaign directly drove, "compared to what would have occurred in the absence of marketing activity." Its demarcation line: attribution and ROAS show what happened; incrementality asks whether marketing caused it. It also warns that incrementality is not static — competition, consumer behavior, and tactics shift outcomes over time.
One more vital worth stating plainly: zero statistics appear in the document. Adoption and attitude percentages quoted in trade coverage of the release come from separate IAB Europe surveys, not from this paper — which cites no numeric evidence at all.
The four methodology families.
The paper's central move is a taxonomy: every incrementality method belongs to one of four families, each graded on causal strength and holistic scope. The grades below are the paper's; the failure modes are the practical reading of the weaknesses the paper itself names.
| Family | What it is | When it is credible | How networks game it |
|---|---|---|---|
| Experiment-based RCTs · holdouts / ghost ads · matched markets | A designed intervention with a true control group that never sees the campaign. | Graded strong on causality, low on scope — usually one platform unless multi-platform holdouts are run. Credible when randomization is real, contamination is controlled, and the test is adequately powered. Stated weaknesses: costly, contamination-prone, time-intensive. | Non-random or contaminated holdouts sold as experiments; test markets selected after results are known; underpowered tests read as wins. The tell: no disclosure of the randomization unit, contamination controls, or statistical power. |
| Model-based counterfactual synthetic control · ML propensity models | A modeled what-would-have-happened, built from data rather than randomization. | Graded strong to moderate; cross-platform when data is linked, often limited to silos. Credible when identification assumptions are explicit and linkage is clean. Stated weaknesses: omitted-variable bias, selection bias, data quality. | The modeler chooses the specification — and the choice moves the answer. Vendor-sourced evidence below: matching-methodology choices alone flipped 83% of campaigns from positive to negative within a single network. |
| Econometric MMM · time-series regression | Regression across channels and time, including non-media factors. | The only family graded high on scope — all measured channels — at moderate-to-weak causality. Credible for portfolio-level allocation over time. Stated weaknesses: backward-looking, lacks granularity, multicollinearity, not predictive for one-off campaigns. | Priors, adstock curves, and channel granularity steer credit between collinear channels. Commerce media lines are often too small and too correlated with promotions for the model to isolate — a confident coefficient is not proof the model could see the channel. |
| Hybrid proxies new-to-brand % · baseline vs. exposed · platform-reported incrementality · simple MTA | Attribution-adjacent shortcuts that borrow incrementality language. | Graded weak on causality, single-platform or campaign-specific on scope; the paper allows them only for fast tactical reads with limits acknowledged. Stated weaknesses: lack causal rigor, limited actionability. | The most common move in commerce media: relabeling attribution output as incrementality. The paper describes platform-reported lift as platform-specific attribution or modeling rather than an independent, true control group — and its use-case matrix warns that proxies used alone risk falsely confirming what the platform already claimed. |
The buried headline
Platform-reported incrementality — the number most RMN dashboards surface — is classified by the IAB's own paper as a hybrid proxy with weak causal strength. The industry body, with the platforms in the room, filed self-reported network lift in the lowest-rigor bucket. When a network's deck cites "incrementality," the first question is which family it came from.
The three requirements to claim causality.
The paper's only "must" lives here, and it is conditional: for a model to credibly answer the counterfactual and claim causality, it must satisfy three requirements. Methods that do — experiment-based, model-driven, or econometric — can claim causal validity. Others remain directional proxies. Hybrid proxies are conspicuously absent from the can-claim list.
- 01
A credible counterfactual
A valid "what if the campaign had not run" scenario — a randomized control, a holdout or ghost-ads design, matched markets, or a synthetic control group. The paper states that the rigor of this initial design is the primary factor determining the causal strength of the entire model.
- 02
Controlling for bias
The measured effect must reflect the campaign, not hidden factors. Named bias sources: selection bias, omitted-variable bias, multicollinearity in MMM, cross-platform data-linkage quality. Named remedies: explicit identification assumptions, confounder controls, parallel-trends tests, geo-randomization, instrumental variables. The paper is honest that bias is reduced, not eliminated.
- 03
Signal separated from noise
A lift estimate inside the noise floor is not reliable for decision-making. Remedies: confidence intervals that exclude zero, bootstrapping, falsification tests, sensitivity analyses — and repeated measurement over time, because incrementality is an ongoing process, not a one-and-done test.
The use-case matrix applies the same logic across nine business needs. For validating ROI, experiments are the gold standard; for cross-channel comparison, MMMs are best suited while experiments are rarely practical across ecosystem silos and proxies are not suited at all. And one use case is telling in itself: calibrating and validating platform-reported lift estimates — a tacit admission, inside the paper, that platform numbers need external checking.
Requires vs. recommends vs. silent.
Most coverage of this paper reads it as rules. It is not. Triaging every directive in the document into what it requires, what it recommends, and what it never addresses is the fastest way to know what you can actually hold a network to — and what you must write into the contract yourself.
| Tier | What is actually there |
|---|---|
| Requires | One conditional "must," and it is definitional: a model must satisfy the three requirements — credible counterfactual, bias control, signal separated from noise — to claim causality. There is no "members shall," no conformance clause, no audit trigger. Nothing in this document requires anyone to do anything. |
| Recommends |
|
| Silent on |
|
The pattern in the silences: everything that would constrain a network's reporting — formulas, disclosure, verification, certification — is absent. The paper defines the vocabulary of rigor and then leaves enforcement entirely to the buyer.
The methodology-variance problem: 6.5x, labeled.
The guidelines contain no numbers, so the size of the hole their silences leave has to be measured elsewhere. The best available evidence is vendor-sourced, and it should be read that way.
The finding — source: Ovative Group & Albertsons Media Collective, March 2026
In a March 2026 analysis — Retail Media iROAS Demystified, by agency Ovative Group and RMN Albertsons Media Collective in partnership with professors from Northwestern's Kellogg School of Management — varying only matching-methodology choices produced 54 distinct iROAS outcomes across 42 campaigns. Within individual campaigns, the gap between highest and lowest iROAS averaged 6.5x (median 2.5x), and 83% of campaigns flipped from positive to negative depending on methodology.
- Scope caveat one: the analysis varies matching choices within one RMN — propensity-score matching designs, feature sets, filtering rules — not differences across vendors. Cross-vendor variance is a separate, unmeasured question.
- Scope caveat two: it is vendor-sourced. An RMN and its agency partner published it, academically partnered, alongside Albertsons' launch of its own onsite incrementality product. The finding argues for exactly the product being launched.
- Timing caveat: it postdates the IAB guidelines by four months and appears nowhere in them. It is evidence about the problem the guidelines address, not evidence from the guidelines.
- Why it still matters: even discounted for interest, the mechanism is undeniable — the guidelines standardize no formula, no baseline window, and no disclosure, so two defensible methodology choices can put the same campaign on opposite sides of zero. That is the arbitrage the triage above exists to close.
The self-grading question, handled fairly.
The standing critique of commerce media measurement — the structural reason the January 2024 IAB/MRC guidelines exist — is that the network sells the media, owns the transaction data, and reports the lift: the seller grades its own homework. This paper was written partly inside that structure, and both readings of that fact deserve the page.
- The exposure
Who wrote it
The Task Force on Incrementality is heavily vendor-populated — Criteo (five contributors), Instacart, Walmart, Meta, Google, Uber, Epsilon, Koddi, Inmar, and others — and IAB Europe's Retailer Council (REWE, Schwarz Media, Ahold Delhaize, Tesco Media, Bol.com, Uber Ads) sits alongside. The parties whose numbers get graded helped hold the pen.
- The omission
What it never says
The document never names the conflict of interest. There is no disclosure mandate, no independent-verification pathway, no MRC accreditation route. The closest it comes is oblique: a use case for calibrating platform-reported lift against experiments, with a warning that proxies alone risk false confirmation.
- The concession
What cuts the other way
The same body — platforms included — filed platform-reported incrementality in its own lowest-rigor tier. That is a concession against interest, and the most useful thing the industry has collectively put in writing about its own dashboards. Vendor-heavy task forces are also how every IAB document gets made; the roster is disclosure, not scandal.
Net reading: this is a genuinely good taxonomy attached to deliberately toothless governance. Treat the vocabulary as an asset and the enforcement as your job.
What to demand from your network under these guidelines.
The paper mandates nothing, but it defines everything — which makes it usable as contract language. Seven demands, each anchored in the document a network's own trade body signed.
- 01
Name the family
Which of the four families produced this number? If the answer is platform-reported incrementality, the guidelines grade it a hybrid proxy — weak causal strength, no independent control group — whatever the dashboard calls it.
- 02
Show the counterfactual
The paper’s first requirement. Ask what the no-campaign world actually is: a randomized holdout, ghost ads, matched markets, a synthetic control — or a modeled baseline. Design rigor is the primary determinant of causal strength, per the paper itself.
- 03
Show the bias controls
Named identification assumptions, confounder controls, parallel-trends tests, geo-randomization — the paper’s own remedy list. "We control for that" without specifics is not an answer under these guidelines.
- 04
Demand statistical robustness
Confidence intervals that exclude zero, falsification tests, sensitivity analyses. The paper is explicit that a lift estimate inside the noise floor is not reliable for decision-making — ask where the interval sits, not just the point estimate. Two disciplines the paper implies but never names: a minimum detectable effect declared at plan time — an underpowered test cannot fail honestly — and a pre-committed look schedule, because a lift number re-read until it clears significance is not evidence.
- 05
Put disclosure in the contract
The guidelines mandate no methodology disclosure — so make it contractual: matching method, baseline windows, filtering rules, metric versions. The Ovative/Albertsons data below shows matching choices alone can flip the sign.
- 06
Calibrate with true holdouts
The paper’s own use-case matrix positions experiments as ground truth for benchmarking platform claims, and warns that proxies used alone risk false confirmation. Fund periodic holdouts on the spend the platform calls best-performing — that is where the surprises live.
- 07
Match rigor to stakes, then repeat
Strategic budget moves ride the strong-causal families; proxies serve fast tactical calls only, limits acknowledged. And re-measure: the paper notes incrementality is not static — competition, consumer behavior, and tactics shift outcomes.
Frequently asked questions.
What are the IAB Guidelines for Incremental Measurement in Commerce Media?
A joint IAB and IAB Europe guidance paper, final November 3, 2025, developed by the IAB Commerce Board, IAB’s Task Force on Incrementality, and IAB Europe’s Retail Media Committee and Retailer Council. It defines incrementality as the causal impact of marketing, organizes measurement methods into four families graded by causal strength and holistic scope, maps them to nine business use cases, and sets three requirements a method must satisfy to claim causality.
Are the guidelines a binding standard?
No. The document self-describes as a paper; it contains no conformance clause, no audit trigger, no certification mechanism, and no MRC involvement — unlike the January 2024 IAB/MRC Retail Media Measurement Guidelines. Its only "must" is conditional: a model must satisfy three principles to claim causality. Nothing in it requires any network to do anything.
What are the four methodology families?
Experiment-based (RCTs, holdouts and ghost ads, matched markets — strong causality, low scope), model-based counterfactuals (synthetic control, ML propensity models — strong to moderate), econometric (MMM, time-series regression — high scope, moderate to weak causality), and hybrid proxies (new-to-brand share, baseline vs. exposed, platform-reported incrementality, simple multi-touch attribution — weak causality, single-platform scope). Platform-reported incrementality sits in the weakest family.
Do the guidelines define iROAS or a standard lift calculation?
No. iROAS never appears in the document; there is no standard lift formula, baseline window, or attribution-window rule, and no methodology-disclosure requirement. That silence has a measurable cost: a vendor-sourced March 2026 analysis by Ovative Group and Albertsons Media Collective, with Northwestern Kellogg professors, found within-campaign iROAS varying 6.5x on average purely by matching-methodology choices within one network, with 83% of campaigns flipping sign.
How is incrementality different from attribution and ROAS?
The paper’s own demarcation: attribution and ROAS show what happened — credit assigned under a chosen rule — while incrementality estimates whether marketing caused the result, against a counterfactual where the campaign never ran. A high attributed ROAS can coexist with near-zero incremental sales when ads harvest demand that already existed.
Deciding what a lift number is worth?
The guidelines give you the vocabulary and the grading axes — but no enforcement. The operating work is turning that vocabulary into contract language: named families, disclosed methods, calibrated holdouts, and decision rules that match rigor to stakes.