On this page
iROAS (Incremental Return on Ad Spend)
iROAS (Incremental Return on Ad Spend)
iROAS measures only the revenue that advertising actually caused — sales that would not have happened without the ads — divided by ad spend. Unlike ROAS, which counts all attributed revenue and assumes every attributed conversion was caused by the ad, iROAS relies on a controlled experiment or causal model to isolate genuine uplift. The two metrics answer fundamentally different questions: ROAS reports profitability against attributed revenue; iROAS reports whether advertising grew the business.
How it works
Formula:
iROAS = Incremental Revenue ÷ Ad Spend
Incremental Revenue = Exposed Group Revenue − Control Group Revenue
(scaled to comparable audience size)
A concrete example: $50,000 in ad spend with $150,000 in platform-attributed revenue gives a ROAS of 3.0. If total sales only rose $70,000 versus the holdout period, iROAS is 1.4 — the remaining $80,000 would have occurred organically. (INCRMNTAL, updated 2025-08-26)
Another example: a Facebook ecommerce campaign with ROAS of 4.0 based on $200,000 attributed revenue had an iROAS of only 1.0, because $150,000 of that revenue would have occurred without ads. (Haus, ©2026)
Measurement methodology
iROAS measurement requires three inputs: exposure data, outcome data, and control-group data (Haus, ©2026). The main approaches, ranked by causal strength per Amazon Ads guidance (as-of 2026-08-20):
| Method | Causal strength | Notes |
|---|---|---|
| Geo Holdout Experiments | Strong | Matched markets, geo-level assignment; avoids user-level tracking; needs volume; markets may diverge due to seasonality/promotions |
| Audience holdouts | Strong | Randomly withheld user group from ads |
| Switchback tests | Moderate | Vulnerable to seasonality |
| Media Mix Modeling (MMM) | Model-dependent | Requires large historical data; good for strategy |
| Amazon Attribution / AMC | Descriptive only | Not causal — correlations, not lift |
| Brand survey (Brand Lift) | Brand only | Not sales lift |
Holdout tests typically need 4–8 weeks to yield reliable results (Haus, ©2026). Privacy changes — iOS restrictions, cookie deprecation — make user-level holdout construction harder, accelerating a shift toward geo-level experiments and MMM (agilebrandguide.com).
The IAB's 2025 guidelines recommend experiments, model-based counterfactuals, and econometrics (MMM) as the three primary measurement paths (Skai, updated 2026-07-20).
IAB reference is to 2025 guidelines; 2026 update status unknown.
iROAS vs ROAS — when to use each
Epsilon (2026-06-23) frames the decision:
- ROAS: appropriate for day-to-day reporting and quick profitability reads; biased toward bottom-of-funnel channels that intercept buyers already close to purchase
- iROAS: appropriate for strategic questions — whether spend grows the business or captures sales that would have occurred anyway
Display and online video may look weak on ROAS while creating new demand that iROAS would better capture (Epsilon, 2026-06-23). SalesDuo (updated 2026-08-20) provides a 2×2 decision framework:
| High iROAS | Low iROAS | |
|---|---|---|
| High ROAS | Scale cautiously | Caution — campaign captures existing demand |
| Low ROAS | Investigate — demand creation missed by attribution | Reduce or redesign |
Channel-level Incrementality Factors (Q1 2026)
Common Thread Collective (CTC) Q1 2026 benchmark across 299 DTC brands uses an Incrementality Factor (IF) — the ratio of iROAS to platform-reported ROAS. IF > 1.0 means the platform understates a channel's value; IF < 1.0 means it overstates it (as-of Q1 2026, volatile):
| Channel | IF (Q1 2026) | Platform ROAS | Implied iROAS |
|---|---|---|---|
| Google Brand Search | 0.30 | 19.07× | 5.72× |
| Meta Acquisition | 1.13 | varies | higher than platform |
| YouTube | ~1.1–1.3 | varies | higher than platform |
| AppLovin | 1.47 | 1.17× | 1.73× |
Brand search and retargeting consistently show high platform ROAS but lower iROAS — these placements often reach people already intending to buy (Skai, updated 2026-07-20; Pacvue, date unknown). Upper-funnel channels frequently show the reverse: lower ROAS, higher iROAS.
Methodology fragility — the Ovative/Albertsons finding
A March 2026 study by Ovative Group, Albertsons Media Collective, and Northwestern Kellogg found that a single retail media campaign can produce 54 different iROAS outputs depending solely on measurement design choices. Across 42 Albertsons Media Collective onsite display campaigns (as-of 2026-03, volatile):
- iROAS outputs varied by an average of 6.5× (median 2.5×)
- 83% of campaigns could flip from positive to negative iROAS based on methodology alone
Key design levers driving divergence: pre-filtering of test/control audiences, matching technique (clustering vs. propensity score), feature selection for matching, and whether observed or modeled revenue was used.
Vendor platforms (Pacvue, Skai, Haus) present iROAS as a tractable optimization KPI; the Ovative/Albertsons/Kellogg study (2026-03) shows that without methodology standardization, the same campaign can produce 54 different results, with 83% of campaigns able to flip sign. iROAS numbers from different vendors are not comparable without knowing their measurement design.
miROAS — the marginal variant
Sellforte distinguishes between iROAS and miROAS (marginal incremental ROAS): iROAS measures historical average channel effectiveness; miROAS measures the return on the next dollar spent, derived from the channel's response curve in an MMM model. Sellforte recommends using iROAS to assess past performance and miROAS as the primary optimization KPI for spend allocation decisions (date unknown).
A subscription software brand example: Google Ads ROAS held steady at 3.0 as monthly spend rose from $10k to $100k, but iROAS fell from 2.5 to 1.2, indicating diminishing returns and cannibalized organic sign-ups. This highlights why miROAS — which captures the response curve shape — is more useful for scaling decisions (Haus, ©2026).
Benchmarks (as-of 2026)
No universal iROAS benchmark exists. Targets depend on margin, COGS, fees, promotions, inventory, customer lifetime value, and campaign objectives (SalesDuo, 2026-08-20).
LiftLab FY-2025 category iROAS changes vs. FY-2024 (directional signals from anonymized U.S. DTC client data, not absolute figures):
| Category | iROAS change YoY |
|---|---|
| CPG / Durables | +20% |
| Footwear | +8% |
| Lifestyle | +8–10% |
| Apparel | +6% |
| Tech Services | −4% |
| Home | −7% |
LiftLab data is from FY-2025; 2026 figures not yet published. Vendor-reported; treat as directional.
LiftLab shows CPG/Durables iROAS rising +20% and Home falling −7% in the same FY-2025 retail media period — no single "retail media iROAS" benchmark applies across product categories.
Common measurement mistakes
SalesDuo (updated 2026-08-20) identifies recurring errors:
- Treating attribution as incrementality
- Treating AMC correlations or new-to-brand rates as causal proof
- Using weak controls or insufficient data
- Stopping tests early before statistical significance
- Ignoring inventory or promotion changes that affect control group behavior
Limitations
Haus (©2026) identifies when iROAS measurement may not be worth the effort:
- Smaller advertisers or niche markets lacking audience scale for statistically valid tests
- Businesses with sales cycles longer than 6–12 months
- Teams needing real-time optimization (results take 4–8 weeks)
- Teams without statistical expertise to design valid experiments
Key terms
| Term | Meaning |
|---|---|
| iROAS | Incremental Return on Ad Spend — revenue caused by advertising ÷ ad spend |
| ROAS | Return on Ad Spend — attributed revenue ÷ ad spend (may include non-incremental sales) |
| miROAS | Marginal incremental ROAS — return on next dollar of spend, from MMM response curve |
| Incrementality Factor (IF) | CTC term: ratio of iROAS to platform-reported ROAS; IF > 1 = platform understates value; IF < 1 = overstates |
| Holdout group | Control audience not served ads; used to estimate counterfactual revenue |
| Counterfactual | Estimated revenue that would have occurred without advertising |