Retail media measurement shifting from ROAS to incrementality

Full analysis

Retail media is about to grade its own homework in front of the class. Jeffrey Cohen, on the Signal & Noise podcast and in AdExchanger, says the industry is moving off ROAS (revenue divided by ad spend, a simple ratio) and last-touch attribution (which credits only the last ad someone saw before buying) toward incrementality testing (did the ad actually cause a sale that wouldn't have happened anyway). His own example: early Amazon campaigns posted 40 to 50 percent ROAS because last-touch was the only model and there was nothing else bidding. Now brands want incremental ROAS, or iROAS.

What's being decided: whether retail media networks keep pricing inventory on a number that flatters them, or accept a measure that will make a lot of that inventory look ordinary. This is hard to undo once agencies write it into upfront terms. What sets the deadline: the next round of annual budget commitments, when investment teams decide what proof they'll demand.

The Market Analyst This moves pricing power. Away from retail media networks, toward whoever controls the measurement and the big advertisers with their own data science teams. In plain terms: the retailers built a toll booth on a number they set themselves, and buyers just found a way to check the receipt. CPMs on inventory that can't survive a holdout test will compress, worst in categories where people were going to buy anyway (staples, consumables). The networks split in two. Amazon, Walmart Connect, maybe Kroger own the transaction data to run real holdouts and keep premium pricing. Everyone else becomes display inventory with a retail media sticker on it. Causal-inference and clean-room vendors finally get a wedge with real budget behind it.

The Skeptic The industry has announced the death of last-touch for a decade. It keeps not dying. For iROAS to actually win, three things have to happen at once: brands swallow smaller headline numbers, agencies retrain planners, and networks publish methodology that deflates their own decks. None of those incentives point the right way. And here's the catch nobody says out loud: iROAS needs a control group, and the network decides who gets into the control group. So the gaming didn't disappear. It moved one layer down, somewhere harder for a brand manager to audit. Amazon will define incrementality in ways that flatter Amazon. Smaller networks will shop measurement vendors until the numbers cooperate.

The Operator Tuesday morning, the measurement team has a re-credentialing problem. The ROAS that closed Q4 looks fake sitting next to a holdout-tested iROAS, and the brand managers know it. What breaks first is the self-serve dashboard that still defaults to last-touch. What breaks second is the managed-service team that has been selling on those inflated numbers and now has to explain them. Inside 90 days, agency investment teams start asking for incrementality methodology as a condition of the commitment. Networks without clean holdout infrastructure, which is most of them, lose renewal conversations they thought were locked.

The Customer / End User The advertiser is the one actually pushing this, which is why it's different this time. A brand running across eight retail media networks can't reconcile eight ROAS numbers that each claim credit for the same sale. Incrementality is the only way to stop paying twice. But the buyer should be careful what they demand. A network-run holdout is the network marking its own test. The advertisers who win are the ones with enough scale to run their own experiments or force a neutral third party in. Everyone else gets whatever control group the seller hands them.

The CFO The cost here isn't the measurement tooling. It's the budget you thought was working and wasn't. If iROAS reveals that a third of retail media spend was reaching people who'd have bought anyway, that money doesn't vanish, it moves. Some goes to upper-funnel, some leaves retail media entirely. For a network, the real exposure is renewal revenue built on inflated proof. For a brand, the payback is real but the transition is ugly: worse-looking numbers for a few quarters while you rebuild the plan around a smaller, truer base.

Where the council splits The Market Analyst thinks incrementality redistributes power to buyers and measurement vendors. The Skeptic thinks it just relocates the gaming, because the network still controls the control group. Both are right, and that tension is the whole story. Incrementality is a better measure and a capturable one. The second fight is Amazon and Walmart versus everyone else: the two players with closed-loop transaction data can run credible holdouts, which turns a measurement standard into a moat only they and a couple of others can afford to sit behind.

What it hinges on Two things. First, whether agencies write incrementality into the terms of the commitment. Buyers renew on familiar ROAS long after they've publicly sworn off it, so a press release commitment means little. Second, who controls the holdout. If networks run their own, this is a marketing upgrade. If neutral third parties or the advertiser's own data science runs it, it's a real repricing.

Prediction: By the end of the 2027 upfront/annual retail media commitment cycle (roughly Q1 2027), no top-tier retail media network outside Amazon and Walmart Connect will submit its incrementality measurement to a mandatory neutral third-party holdout as a condition of renewal, and most iROAS reporting will still run on network-controlled control groups.

Confidence: Medium. The incentive to keep grading your own homework is strong, but a single large advertiser could force the issue.

Why: Cohen's own framing shows the networks benefit from setting the measure, the same way last-touch handed Amazon 40 to 50 percent ROAS with nobody competing for the credit. Incrementality needs a control group, and whoever picks who sits in that group controls the answer, so a network volunteering a neutral referee is volunteering to deflate its own pitch. That runs against the revenue it's protecting. The opposite outcome, networks rushing to independent holdouts, only happens if buyers refuse to spend without it, and buyers have a decade-long habit of renewing on familiar numbers while talking a bigger game. Amazon and Walmart are the exceptions because their closed-loop transaction data lets them run credible in-house holdouts and call it proof.

Revisit by 2027-03-31: We're right if, after the 2027 commitment cycle, the major retail media networks (Kroger Precision Marketing, Target Roundel, Albertsons, and similar) are still reporting iROAS off holdouts they design and control, with no mandatory neutral-party requirement in standard agency terms. We're wrong if a top holding company (Publicis, Omnicom, GroupM) makes independent third-party incrementality verification a written condition of renewal across the mid-tier networks.