Retail Media Meets Agentic Commerce: Paying to Be the Agent’s Default

TL;DR

• Retail media was built on one premise — influence the shopper before the checkout button. An AI agent removes the shopper’s glance entirely, so the placement you used to buy no longer exists as a single thing to buy.

• “Paying to be the agent’s default” is not one purchase. It splits into three separable layers — eligibility (are you in the candidate set), selection (does the agent pick you), and settlement (whose rail completes the sale). The decisive layer and the buyable layer are usually not the same one.

• Use the Agent Decision Stack to mark, per surface, which layer is decisive, which is buyable, and who captures the value if you pay. Retail media’s reflex is to pay at the buyable layer, which is rarely the decisive one.

• Apply the Default-Durability Test: an agent re-decides on every task, so a paid default only holds if it is a property of the product (price, availability, fit, corroboration), not a property of the placement — which evaporates the moment the agent re-runs its choice.

• On retailer-owned assistants the retailer owns selection and can sell sponsored rank. On third-party engines (ChatGPT, Gemini, Claude) answer-independence fences selection off — your money buys eligibility at most, and the durable default is earned, not bought.

The premise retail media was built on just broke

Retail media grew into a roughly hundred-billion-dollar discipline on a single, sturdy premise: you influence the customer before the checkout button is pressed. Sponsored placements, on-site search ads, visual merchandising, the endcap of the digital shelf — all of it is bought attention aimed at a human eye deciding between options. PYMNTS put the tension plainly in May 2026: retail media’s growth “has been built on the bedrock premise of influencing the customer before the checkout button is pressed,” and agentic commerce raises hard questions for that premise, because when software chooses products, compares alternatives and completes the purchase, sponsored placement and visual merchandising lose much of their grip.

The money riding on the answer is not small. Morgan Stanley projects $190–385 billion in US agentic e-commerce spending by 2030; Mordor Intelligence sizes the agentic-AI retail market at $60.43 billion in 2026, growing at a 29.29% compound rate through 2031; Checkout.com estimates agentic commerce could reach a fifth of monthly household spending within five years. DemandSage reports 59% of US consumers already use generative AI for shopping tasks, 65% of them favouring ChatGPT, which fields an estimated 50 million shopping-related queries a day. A brand that treats the agent as a novelty is mis-sizing the channel.

So the strategic question in the title is real, and urgent: can you pay to be the agent’s default the way you paid to be the top of the sponsored carousel? The honest answer is that the question contains a hidden assumption — that “the default” is one thing you can put money against. It is not. To see why, start with what the agent actually removed.

An agent has no attention to sell

The unit of value in retail media was a human glance at a curated consideration set. The retailer decided which twelve products appeared on the shelf, in which order, with which of them boxed in a coloured “sponsored” frame, and you paid to occupy the frame that the glance would land on first. The whole apparatus — impressions, viewability, share of shelf — is a way of pricing attention against a layout.

An agent does not glance and there is no layout. It resolves a selection function over a structured feed: it reads product data, applies the shopper’s stated and inferred criteria, ranks a handful of candidates, and returns — as Criteo notes of conversational interfaces — far fewer products than a grid would, which makes every surviving recommendation worth disproportionately more. There is no carousel to sit at the top of. There is no coloured frame the eye lands on, because there is no eye. The placement, as a purchasable object, dissolves.

Consider what the sponsored frame was actually pricing. On a digital shelf, the retailer manufactured scarcity out of screen real estate: only so many products fit above the fold, ranking was a proxy for the eye’s path, and the auction sold the fraction of a second before the shopper’s attention decayed. Every metric — impression, viewability, share of shelf, position — exists to quantify one scarce human resource: where the glance goes and how long it lingers. Strip out the human and every one of those metrics loses its referent. An agent does not have a fovea, does not fatigue, does not skip past result four because result one had a nicer image. It reads the whole feed at once and applies criteria without a layout mediating the choice. The scarcity retail media monetised — attention against limited screen space — simply is not present, which is why the frame cannot be repriced for agents; there is no glance to put a price on.

What replaces it is not nothing. It is three different things that the old placement used to bundle into one. When a human scanned a sponsored shelf, eligibility (being stocked), selection (being seen and chosen) and settlement (paying at that retailer’s till) all happened in one visual moment, and the sponsored frame influenced all three at once. The agent pulls them apart. You can now be eligible without being selected, selected without controlling settlement, and pay at a rail you do not own for a sale you never saw. “Buying the default” means buying into one of these layers — and the layer you can most easily buy is usually not the one that decides the outcome.

The shift in one line: retail media priced attention against a layout; agentic commerce has no attention and no layout, so the same budget now has to be aimed at a specific layer of the agent’s decision — and aimed wrong, it buys nothing.

The Agent Decision Stack: eligibility, selection, settlement

The instrument for this article is a teardown of the agent’s purchase into three layers, read against three questions each. For any agent surface you sell through, mark whether the layer is decisive (does winning it change the outcome), whether it is buyable (can money move it directly), and who captures the value if you spend there. The four agentic-commerce protocols map straight onto the layers, which is why treating them as one “agent strategy” hides the decision:

• Eligibility — the candidate set. Whether your product is even visible to the agent as an option. Governed by your product feed and by which protocol camps you connect to: Google’s Universal Commerce Protocol (public since January 2026, backed by Shopify, Walmart, Target and Visa and native to Search AI Mode and Gemini) and OpenAI and Stripe’s Agentic Commerce Protocol, with Anthropic’s Model Context Protocol acting as the data layer feeding both. Over a million Shopify merchants are live on ACP through auto-enrolment. Eligibility is largely binary — you are in the set or you are not — and mostly un-auctionable: there is no premium bid that makes you “more eligible.”
• Selection — the pick. Whether, given that you are in the candidate set, the agent chooses you. This is the layer retail media instinctively wants to buy, and it is the layer whose buyability depends entirely on whose agent it is. On a surface the retailer owns, selection can be a sponsored-rank auction. On a third-party engine that has declared answer-independence, it cannot be bought at all.
• Settlement — the rail. Where and how the money moves. Under ACP, OpenAI charges roughly 4% on completed transactions — comparable to a marketplace fee — on top of standard payment processing; one protocol analysis puts ACP’s all-in cost near 7% against UCP’s ~3%. The merchant of record stays intact, but the purchase event completes inside the assistant, off your property. You keep the margin minus the toll; you lose the click, the session and, as covered below, the attribution.

It is worth being precise about the eligibility layer, because it is the one most people underrate and it is where the protocol wars actually land. Four standards emerged across 2025–26, and they compose more than they compete: ACP (OpenAI and Stripe, released under Apache 2.0 in September 2025) handles agent-to-merchant checkout through your existing payment service provider; UCP (Google and Shopify, public since January 2026) covers the fuller discovery-to-returns lifecycle across Google’s surfaces at standard processing rates; Google’s AP2 attaches signed intent, cart and payment mandates as verifiable credentials; and Anthropic’s MCP is the data layer that lets an agent read your catalogue in real time. A full agentic purchase might use MCP to read the catalogue, UCP or ACP to build and confirm the cart, and AP2 to prove the shopper authorised it. For a merchant, the operational point is blunt: eligibility means being wired into the camps your buyers’ agents use, and being wired in cleanly. The rollout has not been smooth — OpenAI’s branded Instant Checkout launched to fanfare, saw only a handful of Shopify merchants ship against it, and was pared back in early 2026 even as the underlying protocol kept expanding to over a million merchants — which tells you the feature names will churn while the eligibility requirement hardens. JD Sports became the first UK retailer live on Stripe’s Agentic Commerce Suite, the same on-ramp any UK brand will use; being early there is an eligibility advantage, not a selection one, and the two should never be confused on a media plan.

Read down the table and the finding is uncomfortable: the layer that is most reliably buyable on any given surface is not the layer that is decisive. Eligibility is decisive-as-a-gate but not an auction. Selection is the whole game but only buyable where you or the retailer own the agent. Settlement is a toll you pay after the decision is already made. A retail-media budget aimed by habit at “buy the placement” lands on whichever slot a sales rep will sell you — and on most surfaces that is not where the default is actually manufactured.

Design rule: never buy “the default” as one line item. Split every agent surface into the three layers first, and only then decide which one your money can actually move on that surface.

Where the default is actually manufactured (three surfaces, three answers)

“The agent’s default” is a different purchase depending on whose selection function is running. There are three surfaces, and conflating them is the most expensive mistake in the channel.

Retailer-owned assistants — selection is buyable.

When Amazon’s Rufus, a Walmart assistant or a grocer’s own chat picks a product, the retailer owns the selection function and can absolutely monetise it. Criteo launched its Agentic Commerce Recommendation Service in February 2026 precisely to power these experiences, claiming up to a 60% lift in recommendation relevancy, and it is productising “sponsored prompts” that shape the questions a shopper is nudged to ask — moving paid influence upstream of the recommendation itself. Here, sponsored rank is real, disclosed and growing. This is the surface where “paying to be the default” behaves most like classic retail media, because the retailer still owns the shelf — it has just turned the shelf into a conversation.

Third-party engines — selection is fenced off.

When ChatGPT, Gemini or Claude names a product in its answer, the platform owns the selection function and, on the engines that have declared answer-independence, has ruled that advertising money cannot buy a place in the recommendation. Ads on those surfaces sit on a separate system from the answer; you can pay to appear beside the recommendation, not to become it. On this surface your retail-media dollar buys eligibility — being in the feed the model reads — and, at most, an adjacent ad slot. It does not buy selection. This is the surface shoppers increasingly start on, and it is exactly where the reflex to “buy the default” fails silently.

Your own site or app assistant — selection is a merchandising decision.

When your own on-site assistant recommends, there is no auction to enter because you control the ranking outright; “buying the default” collapses into a merchandising and margin decision, the same one you have always made, now expressed as assistant logic rather than a planogram.

Read across the middle column and the strategy writes itself: the only cell where selection is genuinely for sale to an outside brand is the retailer-owned surface. Everywhere else you are either earning it or you already own it. A budget that does not distinguish these cells is a budget spent by reflex.

Three surfaces, three different answers to the same sentence. On the first you buy rank; on the second you cannot; on the third there is nothing to buy. A media plan that prices “the agent’s default” as a single cross-surface product is quietly paying retailer-owned-surface prices for third-party-surface outcomes it will never receive.

The Default-Durability Test: property of the product, or property of the placement?

Suppose you do buy selection where it is buyable. A second instrument decides whether the purchase is worth anything: the Default-Durability Test. It turns on a structural difference between a human placement and an agent’s choice.

A sponsored shelf bought a moment that could not be re-run. The shopper glanced once, chose, and left; the placement did its work in that single pass. An agent has no single pass. It re-decides on every task, and often several times within a task as it narrows criteria — “cheaper,” “in stock near me,” “better reviews,” “actually, something that fits a small kitchen.” Each re-decision re-runs the selection function from scratch. So the question is not “did I win the placement” but “does what I bought survive the agent choosing again.”

That splits paid defaults cleanly in two. A default that is a property of the product — the best price, genuine availability, a real fit for the stated need, corroboration the model trusts — survives re-decision, because the criteria the agent tightens are criteria you actually meet. A default that is a property of the placement — a paid rank ungrounded in any selection criterion — is overwritten the instant the agent re-runs with a sharper filter or a rival’s feed improves. You paid for a position the next query erases.

The test, in three questions: (1) If the agent re-decided right now with a stricter criterion, would I still be chosen on the merits? (2) Is the thing I paid for attached to my product, or to a slot? (3) When my spend stops, does the default persist for even one more query? Two or more “no”s means you bought a placement, not a default — and in an agentic channel that is a rental with an unusually short lease.

Make it concrete. A shopper asks an assistant for “a good burr grinder under £200.” Brand X has paid for a “recommended” slot; it surfaces first. The shopper then says “actually, quiet enough for a flat, and available for next-day.” The agent re-runs: it re-reads the feed, filters on noise data and delivery windows, and Brand Y — which never bought a placement but publishes accurate specs and has genuine next-day stock — wins the second pass, because the second pass is decided on criteria attached to the product. Brand X’s paid rank did not carry into the re-decision, because it was never grounded in anything the agent re-checks. Two queries, and the bought default is gone. That is the failure mode the test is built to catch before the money is committed, not after.

This is why the durable defaults in agentic commerce are unglamorous: accurate structured product data, honest availability, competitive pricing, and third-party corroboration the model can lean on. They are boring precisely because they cannot be switched off — which is the only definition of “default” that means anything to a machine that keeps deciding.

Worked example: Kelder’s £300k and the three layers

Kelder is a UK direct-to-consumer coffee-equipment brand — grinders, pour-over kit, a well-reviewed home espresso machine — doing about £8M a year, planning a £300,000 “agentic commerce” budget for 2027. Two plans reach the same December visibility on paper and diverge sharply underneath.

The naive plan treats “be the default” as one purchase. Kelder’s agency books every “AI default” and “sponsored recommendation” placement a rep will sell: a sponsored-rank package on a retailer’s assistant, an ad slot beside ChatGPT’s shopping answers, a contextual buy it is told will make it “the recommended grinder.” The feed gets connected almost as an afterthought. The £300k is spent as if all three layers were the coloured frame on a shelf.

The durable plan spends against the stack. Kelder splits the budget by layer and surface. Roughly £40k goes to eligibility — connecting the product feed cleanly to both protocol camps (UCP via Shopify, ACP via Stripe’s Agentic Commerce Suite, the same route JD Sports took as the first UK retailer on it) and fixing the data hygiene that decides whether the espresso machine is even a candidate. On the third-party engines Kelder spends nothing trying to buy selection it cannot buy; instead it funds the un-buyable default — earned citations and first-party corroboration that make the model land on Kelder on the merits (the same channels the site’s guide to earning editorial backlinks and its competitor backlink analysis workflow are built to produce). It buys sponsored rank only on retailer-owned assistants, where selection is genuinely for sale and measurable. And it instruments settlement before spending a pound on-rails.

The durable ledger is worth seeing in numbers. Of the £300k: about £40k to eligibility (protocol integration and feed hygiene, a one-off that never needs re-buying); roughly £150k to the earned layer — digital PR, first-party test data worth citing, and the corroboration that wins un-buyable selection on the third-party engines, which is the only line that compounds across every surface at once; about £80k to sponsored rank on the two or three retailer-owned assistants where Kelder’s espresso machine actually sells and where the placement is measurable; and £30k held back for settlement instrumentation — geo-holdouts and incrementality reads on the on-rails conversions. The naive ledger inverts this: near-zero on eligibility and earned media, the whole £300k poured into placements, most of it on surfaces where the recommendation was never for sale.

Month one, the two plans look identical: both brands appear when an assistant is asked to recommend a home grinder. By month six they have separated. The naive plan’s paid “defaults” fail the Default-Durability Test — every time an agent re-decides with a tighter criterion, the ungrounded rank is overwritten, and the ChatGPT contextual buy never converted the way the pitch implied because the recommendation was never for sale there in the first place. Worse, Kelder cannot cleanly measure the on-rails ACP conversions it paid roughly 4% on, because the purchase completed inside the assistant, off its property. The durable plan’s earned selection persists across re-decisions and spans every engine at once; the retailer-owned sponsored rank pays for itself where it is measurable; and settlement is read through incrementality rather than a dashboard that cannot see the decisive touch.

Same headline budget, same month-one screenshot. One brand rented a position the next query erased; the other became a choice the agent keeps making.

The strongest objection: retail media is adapting, not dying

The hardest counter to everything above is not that agents don’t matter — it is that retail media is re-forming around them, richer than before, and the “no attention to sell” claim is simply wrong. Criteo’s Agentic Commerce Recommendation Service exists, ships, and claims a 60% relevancy lift. Retailers are building sponsored placements directly into their own conversational assistants. “Sponsored prompts” move paid influence upstream of the recommendation by shaping the questions shoppers ask. Payment platforms are turning purchase data into ad infrastructure, widening what “retail media” even means. On this reading, the agent does not abolish paid influence; it relocates it to a new surface with fewer, higher-value recommendations, and the smart money is already there.

Concede the strong form, because it is true: on retailer-owned agentic surfaces there is abundant monetisable selection, and Criteo is right that conversational interfaces surfacing fewer products make each paid recommendation more valuable, not less. Retail media on those surfaces is not dying; it is arguably about to have its best decade.

But the concession does not rescue “pay to be the agent’s default” as a general strategy — it sharpens exactly where the strategy holds and where it fails. Everything in the objection describes the retailer monetising its own selection layer, on a surface it owns, where it controls the decision. None of it extends to the third-party assistant where a growing share of shoppers start and where the platform, not the retailer, owns selection and has fenced it off from advertising money. So retail media does not die; it bifurcates: buyable and thriving on owned agentic surfaces, un-buyable on the independent ones. The strategic error the objection can lull you into is porting a tactic that genuinely works on your own assistant onto someone else’s — paying, in effect, for a default that the other platform has already told you is not for sale. The 60% relevancy lift is a superb reason to connect your feed and to monetise your own assistant. It is not evidence that you can buy the recommendation on an engine whose owner has ruled that out.

Falsifier: if a genuinely third-party assistant — not retailer-owned — opens a disclosed, paid auction for position in the agent’s actual product selection (not a separate ad slot beside it), then eligibility and selection collapse back into one buyable layer and the core claim here — that the decisive layer and the buyable layer come apart — is refuted on that surface. Every current move points the other way: the platforms are walling selection off, not auctioning it.

The link builder’s play: become the default you cannot buy

The whole analysis converges on one operational point for anyone who works in earned media. The un-buyable layer — selection on the third-party engines — is precisely the layer earned authority moves. You cannot pay ChatGPT or Gemini to recommend a product, but you can become the option their selection function keeps landing on, because the corroboration they read is earned, not bought. That is the same mechanism the rest of this site treats as the fundamentals of link building, now pointed at an agent’s decision rather than a ranking algorithm.

Retail media, seen through the stack, buys you two of the three layers and only two: eligibility everywhere, and selection on surfaces you or a retailer own. It cannot buy the durable, cross-engine default — the one that survives re-decision and appears on every assistant at once — because that default is a property of your product and its reputation, not of any placement. The link builder’s job is to manufacture that property: earned citations, first-party data worth citing, and the kind of independent corroboration a model treats as trustworthy. The site’s 15 link building strategies hub, its sponsorship link building and niche edit playbooks, and the case for hiring a dedicated link building specialist all resolve, in agentic terms, to the same thing: owning more of the recommendation’s independent inputs than any competitor can.

Two adjacent disciplines get more important, not less, once selection is un-buyable. The first is protecting the corroboration you have earned: a model that reads a poisoned or manipulated citation graph can be steered against you, which makes negative SEO defence and a clean backlink profile audit and disavow discipline part of agent-visibility hygiene, not just search hygiene. The second is recovering from suppression fast — the manual action recovery muscle applies directly when an entity signal that fed your selection gets pulled. And because agent visibility is now global by default across international link building and specifically across link building in India and South Asia, the earned default compounds in every market at once rather than being re-bought market by market the way paid placements are.

Settlement is the last piece, and it is a measurement problem before it is a media one. When the purchase completes on-rails under ACP — the ~4% toll, the sale finishing inside the assistant — the decisive touch (the model naming you) leaves no click you can instrument, so the dashboard credits whatever it can see and under-credits the earned recommendation that actually caused the sale. Reading that layer through incrementality rather than last-click is the only way to avoid funding the buyable layer at the expense of the decisive one. The broader data on where citations come from keeps pointing the same way: the earned layer is the main event, which is why the 2026 link building statistics and the tooling in the best link building tools round-up are worth more to agent visibility than another sponsored-rank package on a surface that was never selling the recommendation. If any of this is new, the primer on what backlinks are and why they still matter is the place to start.

The arithmetic underneath all of this is what makes the earned layer the anchor rather than a nice-to-have. Muck Rack’s May 2026 analysis of more than 25 million AI citations found roughly 84% of them earned and about 0.3% paid — the recommendation surface is overwhelmingly won by things money cannot directly buy. Set that against a retail-media reflex to pour the budget into placements, and the misallocation is stark: the channel spends where the buying is easy and starves the layer that produces five-sixths of the citations. Funding earned selection first is not idealism about “good content”; it is where the citations demonstrably come from, and it is the one investment that clears on every engine — buyable-selection and un-buyable-selection alike — with a single spend. The paid layers then do what they are actually good at: eligibility everywhere, and rank on the surfaces that genuinely sell it.

What to do Monday

Spend against the stack, not against the word “default.” Five moves, in order:

1. Map the three layers for every agent surface you sell through. For each, mark decisive / buyable / who-captures. Do not approve a single “be the default” line item until it is attributed to a specific layer on a specific surface.
2. Connect the feed to the protocol camps first. Eligibility is binary and un-skippable — UCP via Shopify, ACP via Stripe’s suite, with your data clean enough to be a candidate. You cannot be selected if you are not in the set, and no bid fixes that.
3. On third-party engines, stop trying to buy selection. Redirect that money to earned citation and first-party corroboration — the only route to the recommendation on surfaces that have fenced advertising off from the answer.
4. Buy sponsored rank only where selection is genuinely for sale. Retailer-owned assistants (Amazon, Walmart, big-grocer chat) and your own assistant. Elsewhere the placement fails the Default-Durability Test the moment the agent re-decides.
5. Instrument settlement before you spend on-rails. If the purchase completes under ACP’s ~4% toll, set up geo-holdouts or incrementality reads now, because the click attribution has moved off your property and the dashboard will over-credit the buyable layer by default.

“Paying to be the agent’s default” is a coherent goal and an incoherent purchase. Split it into eligibility, selection and settlement, and it becomes three decisions — one you must make, one you can only sometimes buy, and one you must measure rather than spend. The default worth having is the one no cheque can switch off: the choice the agent keeps making because, on the merits it re-checks every time, you are still the right answer.