Get the latest insights right in your inbox
Amazon Incrementality: How to Prove Your Ads Are Actually Working
Published: March 21, 2026
Last updated: March 21, 2026
Marissa Incitti
Marissa Incitti leads research and content at Feedvisor focused on Amazon, Walmart, and the broader e-commerce marketplace ecosystem. Her work covers retail media performance, pricing strategy, and how AI-driven discovery is reshaping how brands compete across marketplaces. Prior to Feedvisor, she worked in content leadership roles at a Fortune Global 500 omnichannel commerce technology company.
Table of Contents
- Your Best-Looking Campaigns May Be Your Least Effective
- iROAS vs. ROAS: The Number That Changes Everything
- Why Amazon’s Own Attribution Overstates Your Results
- How to Run an Incrementality Test on Amazon
- Which Ad Types Deliver Real Incremental Value
- Amazon Marketing Cloud: The Access Barrier Just Disappeared
- Three Mistakes That Kill Your Incrementality Test
- FAQ
Your Best-Looking Campaigns May Be Your Least Effective
That branded Sponsored Products campaign showing 12x ROAS? A large portion of those sales were going to happen anyway. The retargeting campaign with stellar click-through rates? Research indicates only about 40% of retargeting conversions are truly incremental - the other 60% were already heading to checkout.
This is the problem incrementality testing solves: separating ads that create sales from ads that simply claim credit for them.
Most Amazon advertisers optimize toward ROAS - and they are making budget decisions based on a number that systematically lies to them. Campaigns targeting existing customers always look more efficient because those customers convert at higher rates regardless. The math rewards you for spending money on people who were already buying.
Incrementality measures the additional revenue your ads generate that would not have occurred without them. It isolates causal impact by comparing a group that saw your ads (test) against a matched group that didn’t (control). When brands first run these tests, they routinely discover that 30-60% of their “ad-driven” sales were organic purchases that would have happened regardless.
iROAS vs. ROAS: The Number That Changes Everything
Standard ROAS divides total attributed revenue by ad spend. iROAS (incremental Return on Ad Spend) divides only the incremental revenue - sales that wouldn’t have happened organically - by ad spend.
The formulas are simple. The gap between them is not.
| Metric | Formula | What It Captures |
|---|---|---|
| ROAS | Total Attributed Revenue / Ad Spend | Everything the platform claims, including organic cannibalization |
| iROAS | Incremental Revenue / Ad Spend | Only the sales your ads actually caused |
Consider a seller spending $10,000/month on branded Sponsored Products. The dashboard shows $100,000 in attributed revenue - a 10x ROAS that looks excellent. But incrementality testing reveals 70% of those buyers were searching your brand name and would have purchased regardless. True incremental revenue: $30,000. True iROAS: 3x.
That 3x might still justify the spend. But the budget decision at 3x looks completely different from the decision at 10x.
Industry benchmarks show iROAS ranging from roughly 2.5x to 16x across Amazon advertisers. Mature brands with strong organic presence see the largest ROAS-to-iROAS gap - sometimes 50-80% - because their organic baseline captures much of the demand their ads claim credit for. Newer brands see a smaller gap, often 10-30%.
Below 1.0x iROAS, you are paying for sales you would have gotten for free. This is the only advertising cost metric that tells you whether your ads are creating value or just consuming it.
Why Amazon’s Own Attribution Overstates Your Results
Until January 2026, Amazon used a basic last-touch model with a 14-day lookback window. If a shopper saw your DSP ad and bought within two weeks, DSP took credit - even if the ad had zero influence on the decision. Amazon addressed this with a “shopping-signal enhanced” model that uses machine learning to evaluate whether view-through impressions actually contributed to conversions. A step forward - but it still cannot tell you what revenue disappears when you turn the ad off.
Not all ad types overstate equally:
| Ad Type | Overstatement Risk | Why |
|---|---|---|
| Sponsored Products (non-branded) | Moderate | Click-based; targets shoppers with genuine purchase intent |
| Sponsored Products (branded) | High | Captures shoppers already looking for you |
| Sponsored Brands | High | View-through component inflates credited conversions |
| Sponsored Display (retargeting) | High | Targets shoppers already deep in the funnel |
| DSP (prospecting) | Low-Moderate | Reaches genuinely new audiences |
| DSP (retargeting) | Very High | Historically relied on view-through attribution with 14-day windows |
The pattern is counterintuitive: your highest-ROAS campaigns are frequently your lowest-incrementality campaigns. Retargeting looks efficient because it targets people who are about to convert anyway. Prospecting looks expensive because it reaches people who need convincing - but those are the ads actually creating new demand.
Amazon’s January 2026 update applies ML to view-through conversions for Sponsored Brands, Sponsored Display, and on-platform DSP. Offsite DSP still uses the traditional 14-day window. Attribution tells you who saw an ad before buying - not whether the ad caused the purchase.
How to Run an Incrementality Test on Amazon
Three methodologies work, ranked by reliability:
1. Geo-Holdout Testing (Gold Standard)
Split your markets geographically. Run ads in 80% of regions (treatment) while suppressing them in 20% (control). Compare sales between groups after 4-6 weeks.
This works well for DSP and Sponsored Display because both support geographic targeting. Sponsored Products is harder - geo-targeting options are limited. Verify that your ad platforms respect geographic boundaries before launching. Programmatic delivery can bleed across borders.
2. Audience-Based Holdout
Create matched user cohorts within Amazon DSP. Expose one group to ads, suppress the other. More granular than geo-testing but carries higher contamination risk - users share devices and households, leaking ad exposure into your control group.
3. Pre/Post Analysis
Compare performance before and after a campaign change. Simplest to run, least reliable. Use as supplementary evidence only.
For your first incrementality test, start with a geo-holdout on your largest DSP campaign. Run for minimum four weeks. Pre-commit to a primary KPI before you start - changing the goal mid-test is how you get false positives.
Which Ad Types Deliver Real Incremental Value
The reallocation playbook that emerges from incrementality data is consistent across brands: shift away from retargeting-heavy and branded search strategies (strong ROAS, weak incrementality) and toward prospecting and upper-funnel campaigns (weaker ROAS, strong incrementality).
One beauty brand tested incrementality across Amazon, Ulta, and Sephora. Amazon showed 2.8x iROAS - strong enough that the brand shifted 30% of its Sephora budget to Amazon. Result: 22% more incremental sales overall. A DTC apparel brand that scaled prospecting based on incrementality data saw a 25% increase in incremental revenue over eight weeks.
On the other end, a SaaS company discovered its display retargeting produced only a 1.2% increase in subscriptions - with a real ROAS 62% lower than what standard attribution reported. That budget found a better home elsewhere.
Test each campaign type, rank by iROAS instead of ROAS, and reallocate from the bottom to the top. Accept that total attributed ROAS will likely decrease - but total revenue, organic plus paid, should increase. Brands that run this process regularly report 10-20% improvements in marketing efficiency. If you are tracking TACoS alongside iROAS, you will see the full picture: ad-attributed efficiency may dip while total advertising cost of sale improves.
One caveat - this breaks down if your catalog is seasonal or if you are a new brand with minimal organic presence. Newer brands have less organic cannibalization, so ROAS and iROAS are closer. Your priority is capturing share, not optimizing between incremental and non-incremental channels.
Amazon Marketing Cloud: The Access Barrier Just Disappeared
The biggest change for incrementality measurement in 2026 is not a new methodology - it is access. In September 2025, Amazon made AMC (Amazon Marketing Cloud) free for all Sponsored Ads advertisers. Before that, clean room access required DSP usage and significant ad spend, limiting incrementality analysis to large brands with dedicated data teams.
AMC is a privacy-safe clean room where you can query your advertising and sales data across all Amazon ad types. For incrementality, the capabilities that matter:
- Customer path analysis across Sponsored Products, Sponsored Brands, Sponsored Display, and DSP - showing the actual journey, not just the last touch
- New-to-brand metrics that separate genuinely new customer acquisitions from repeat purchases your ads claimed credit for
- Five-year historical data (expanded from 13 months at CES 2025), enabling multi-year cohort analysis and stronger incrementality baselines
- AWS Clean Rooms integration for combining your first-party data with Amazon signals without either party sharing raw data
One catch - AMC requires SQL knowledge. Amazon launched an AI-powered “Ads Agent” that accepts natural language queries, which lowers the bar. But for serious incrementality work, you still need someone comfortable writing queries or a third-party tool like Intentwise that provides an AMC interface.
For brands spending over $50,000/month, dedicated platforms like Measured ($250K+/year) or Haus offer more sophisticated experiment design. Below that threshold, AMC with the AI Agent is the cost-effective starting point. Either way, the advertising fundamentals matter before you layer on incrementality analysis.
Three Mistakes That Kill Your Incrementality Test
Contaminated control groups. Your control group must have zero exposure to the ads you are testing. Harder than it sounds - programmatic delivery bleeds across geographic boundaries, users switch devices, household members share accounts. Audit every channel to confirm your holdout markets are truly dark. One leak and your test measures nothing.
Stopping early because the data looks good. Pre-commit to a test duration (four to six weeks minimum) and do not peek. Stopping when results look favorable produces false positives and overestimated effects. Significance at week two is not significance at your pre-committed endpoint.
Testing what you already know. Running an incrementality test on a pure prospecting campaign or a branded search campaign just confirms the obvious - you sacrifice holdout revenue to learn nothing. The highest-value tests target campaigns where you genuinely don’t know the answer. That mid-funnel Sponsored Display campaign targeting competitor ASINs? That is where incrementality gets interesting.
FAQ
How is iROAS different from regular ROAS?
ROAS counts all revenue the platform attributes to your ads, including sales that would have happened organically. iROAS counts only the revenue your ads actually caused, measured through controlled experiments. It is always equal to or lower than ROAS.
Do I need Amazon Marketing Cloud to measure incrementality?
No. You can run geo-holdout tests by pausing campaigns in specific markets and comparing sales. AMC gives you deeper path-to-conversion data, but the basic experiment works with standard reporting. AMC became free for Sponsored Ads advertisers in September 2025.
How much ad spend do I need before incrementality testing is worthwhile?
Around $15,000-$20,000/month in the campaign type you want to test. Below that, you won’t generate enough conversions in your holdout market to reach statistical significance in a reasonable timeframe.
Which Amazon ad type has the highest incrementality?
DSP prospecting campaigns targeting audiences outside your existing customer base typically show the highest incremental lift. Non-branded Sponsored Products on category keywords also perform well. Branded search and retargeting consistently show the lowest incrementality.
How long should an incrementality test run?
Minimum four weeks, ideally six. Avoid running tests during major sales events unless that is specifically what you want to measure - seasonal spikes make it harder to isolate ad impact.
Your ad spend is only as good as the revenue it actually creates. Feedvisor’s AI-driven advertising optimization helps brands identify true incremental value across their Amazon campaigns - moving budget from claimed conversions to caused conversions. See how it works
Your Ads Are Taking Credit for Sales They Didn't Create