Every dollar of the retail media premium rests on one promise the walled gardens have repeated for a decade: we saw the exposure, and we recorded the purchase, so we can tell you exactly what your money did. That is the closed loop. It is the reason a retailer can charge a margin on media that a pure publisher never could, because the retailer owns both ends of the shopper journey and can staple an impression to a basket. US retail media spend is on pace to reach $107.6 billion in 2026, nearly triple where it stood five years ago, according to NIQ. That number is a bet on the closed loop holding.
It will not hold. The loop is already leaking at the measurement layer, where brands cannot prove that the sales being credited to media would not have happened anyway. And a second, more fundamental break is arriving in the form of agentic shopping, where an AI agent now stands between the shopper and the shelf. When an agent mediates the basket, the moment of influence and the moment of purchase stop happening in the same place, on the same screen, under the same retailer’s telemetry. Retailers who keep selling attributed return on ad spend into that environment are selling a receipt for a transaction they no longer fully witnessed. The trust that funds the margin will go first, and the margin will follow.
The Loop Was Always Weaker Than the Invoice Suggested
Start with what the industry already knows and rarely says out loud. Return on ad spend, the metric that underwrites most retail media invoices, measures efficiency rather than causality. It tells a brand where the budget went. It does not tell the brand what the budget created. NIQ calls treating ROAS as an incrementality answer a category error, and an expensive one, because retail media rarely supports the controlled testing available on platforms like Meta or Google, so ROAS and last-touch data became the default optimization signal almost everywhere even though neither was built to answer the incrementality question.
The gap is structural. A shopper sees a product on social media, gets curious, searches for it directly on a retail site, and buys it there. The attribution model credits the search. The actual driver of the sale happened somewhere else entirely, and the retailer booked media revenue against a conversion it merely intercepted. Multiply that across a brand’s full commerce footprint and you get a number that gets reported and does not get trusted. The unreliability runs deeper still: no two retailers calculate ROAS the same way, the formulas are frequently opaque, and NIQ cites independent academic research finding that switching methodologies alone can shift the resulting number significantly.
The market’s own posture confirms the fracture. NIQ and Unlimitail found in June 2026 that 67% of CMOs plan to increase retail media investment in 2026, while only 53% believe their retail media networks provide adequate measurement and attribution to support reliable incrementality measurement. Read those two figures together. A majority of senior marketers are increasing spend into a channel that a near-coin-flip of them does not believe they can measure. Incrementality and cross-channel measurement sit at the top of brands’ list of measurement pain points even as budgets climb toward $108 billion. That is a market buying on faith, and faith is a fragile thing to build a margin on when a new technology is about to test it.
Agentic Shopping Severs Influence From Purchase
Here is what changes the physics. For the closed loop to mean anything, the retailer has to observe both the exposure and the purchase inside one environment it controls. Agentic shopping dismantles that premise by inserting an autonomous layer that shops on the consumer’s behalf.
Consider what Rezolve AI described on October 9, 2026, when it unveiled RezolvePay, infrastructure that lets a consumer’s chosen AI agent interact with a merchant-side agent and complete a purchase. The company’s own framing of the opportunity is the tell: a global retail industry of roughly $30 trillion in annual sales, with McKinsey estimating that AI agents could mediate $3 trillion to $5 trillion of global consumer commerce by 2030. The demonstration featured Meta’s Muse placing a coffee order for in-store collection, with a receipt that identifies Muse as the agent that placed the order. Rezolve’s CEO was explicit that the opportunity is every AI agent and merchant, and that consumers will choose different assistants to shop for them.
Sit with that. When an agent selects the product, compares the options, and executes the transaction, the retailer’s role contracts to fulfillment and payment at the end of the chain. The decision that retail media exists to influence happened upstream, inside the agent’s reasoning, across whatever signals the agent weighed. The retailer can record that a purchase occurred and, with infrastructure like RezolvePay, can even record which agent placed it. The retailer cannot see why the agent chose that product, which means it cannot honestly claim its sponsored placement caused the basket. The loop is physically open. Influence lives with the agent. Purchase lives with the merchant. They are no longer the same event.
The agent layer is advancing on the sell side at the same time, which deepens the problem rather than solving it. At unBoxed 2026, Amazon Ads introduced Amazon Ads Agent, an agentic AI platform backed by what Amazon describes as trillions of shopping, browsing, and streaming signals. Its new Full-Funnel Campaigns, now available to all advertisers in the United States, let the advertiser provide budget, products, and creative while Amazon’s AI handles planning, execution, and continuous optimization, connecting discovery to purchase and tuning toward long-term growth. Amazon says DVA+ will begin rolling out to advertisers in late October 2026. Amazon reports that advertisers using natural language to apply campaign targeting recommendations saw more than 25% additional unique customers while lowering cost-per-impression costs by more than 10% on average.
Those are real efficiency gains, and they sharpen the point. On one side, the retailer’s AI is optimizing the buy autonomously. On the other side, the consumer’s AI is making the purchase autonomously. Two agents transact, and the human who was supposed to be influenced by the ad may never have seen it in a form the old attribution model understood. NIQ warns that as agentic commerce and automated activation take on more day-to-day decisions, AI systems will act on whatever measurement they are given, good or bad, at a speed no human reviewer can match. Feed a self-optimizing system a flawed incrementality signal and it will scale the flaw faster than any team can catch it.
What the Open Loop Actually Costs
The cost arrives as the slow withdrawal of advertiser belief, and the margin erosion that follows belief.
Retail media margins are high precisely because the closed loop let retailers sell proof, not just placement. Strip the proof and what remains is undifferentiated inventory competing on price with every other channel. The nearly half who already doubt their networks’ measurement are the leading indicator. When an agent mediates the basket and the retailer still sends an invoice crediting its sponsored placement for the sale, the sophisticated brand will run the mental experiment every CFO eventually demands: would this purchase have happened without the media dollar? If the honest answer is unknowable, the dollar becomes discretionary, and discretionary dollars move to wherever causality can be demonstrated.
There is a sequencing trap here too. Brands that treat measurement as a problem to solve once a perfect model exists will find that agentic commerce does not wait for the model. The decisions move into the agent layer while the brand is still debating methodology. NIQ’s point that 91% of consumers now buy both in-store and online, switching the moment a preferred product is unavailable where they first looked, already strains single-channel measurement. Agents make that switching instantaneous and invisible, executing in milliseconds across retailers the brand’s measurement treats as separate scoreboards. A measurement regime built one walled garden at a time will underestimate what is actually happening and misallocate the budget it is supposed to protect.
The deepest cost is strategic. A retailer that keeps selling attributed ROAS into an agent-mediated world is training its own advertisers to stop trusting its numbers at the exact moment those advertisers are deciding how to budget for the agentic era. That is a trust debt that compounds. Once a brand concludes that a network’s reporting is theater, it routes around it.
The Operating Model That Replaces the Scoreboard
The fix is a different operating model, and it rests on three foundations that reinforce one another.
The first foundation is validated incrementality rather than reported ROAS. Incrementality has a simple, widely agreed definition: the sales that would not have happened without a specific dollar of spend. Proving it requires visibility into the full conditions of commerce around every purchase, including price, promotion, organic and paid rank, ratings and reviews, buy-box ownership, the entire digital shelf, alongside retail distribution data. NIQ’s own example is the brand that does not know a product moved from 2,000 doors to 3,800 in a single reset and therefore cannot separate that effect from media. Because retailers rarely support randomized controlled trials or geo-lift tests, credible measurement substitutes an ensemble: econometric modeling similar to marketing mix modeling for strategic allocation two or three times a year, designed experiments where a retail partner allows them, and passive experimentation that treats naturally occurring single-variable situations as proxies for controlled tests. Critically, this has to run at the daily, campaign level to match the cadence at which purchases and now agents actually move. Knowing that Amazon media returned 1.7x does not tell a brand which keywords to adjust tomorrow morning, and it tells an autonomous optimizer even less.
The second foundation is agent-readable product data. This is the one most retailers and brands are underinvesting in, and it is about to become decisive. When a consumer’s AI agent evaluates products, it reads structured signals: price, availability, specifications, ratings, fulfillment terms, the digital shelf rendered as machine-legible fact. The brand whose product data is clean, complete, and structured for agent consumption wins consideration inside the agent’s reasoning. The brand whose data is thin gets passed over by a system the brand cannot see or appeal to. Rezolve’s architecture is explicitly designed so merchants retain control over their products, commercial terms, and fulfillment while serving demand that arrives through any agent. That control is only worth something if the underlying product data is good enough for an agent to act on. Treat the digital shelf as a required input to both discovery and measurement, not as a separate reporting exercise.
The third foundation is measurement instrumented across networks rather than a scoreboard per walled garden. No single retailer can solve incrementality for a brand operating across dozens of them, and no single retailer can see the agent-mediated journey that spans its competitors. NIQ’s guidance to retailers is direct: make performance data available to brands, their agencies, and their chosen measurement partners rather than holding it back, invest in in-market experimentation infrastructure, and collaborate on standards and data-sharing because that strengthens the case for the channel for everyone. The retailer that opens its loop to independent, cross-network measurement will look like it is surrendering an advantage. It is actually trading a proof it can no longer deliver for a credibility it can.
Underneath all three sits the discipline NIQ names as the most overlooked: the organizational muscle to convert a trustworthy number into a clear next move. Most brands suffer from an inability to act on their data. In an agentic world that gap becomes fatal, because the systems acting on the data move faster than the committees debating it.
The Window Is the Build, Not the Rollout
The honest read as of October 2026 is that the infrastructure for agent-mediated commerce is shipping now. Amazon Ads Agent is live for US advertisers, with DVA+ scheduled to begin rolling out in late October 2026. RezolvePay is demonstrating cross-agent transactions and courting the merchant ecosystem for a market McKinsey sizes at $3 trillion to $5 trillion by 2030. The closed loop is going to be routed around by agents that transact whether or not the measurement model has caught up.
Retail and CPG leaders have a choice that will look obvious in hindsight and feels counterintuitive today. Keep defending the closed loop and the attributed ROAS it produces, and watch the nearly half who already doubt the numbers grow into a majority that stops paying the premium. Or build the operating model that survives the open loop: validated incrementality running at campaign cadence, product data structured for the agents that now mediate the shelf, and measurement instrumented across networks by partners the whole market trusts. The retailers and brands that start that build in 2026, while the agents are still learning the aisles, will be the ones setting the terms when agent-mediated commerce becomes the default. The ones waiting for a perfect model will discover the agents did not wait for them.