Universal Cart Suggests Alternatives. Did You Write the Logic?

At Google I/O 2026, the Universal Cart demo did something most shopping software has never tried to do in public: it overruled the shopper. A person assembling a custom PC dropped an AMD Ryzen processor and an Intel board into the same basket, and the software stopped them cold. "The CPU and motherboard in your cart are not compatible," it said. "The Ryzen 7 CPU requires a motherboard with an AM5 socket." Then it offered something else.

That is substitution, shipped into a mainstream surface and demoed on stage rather than promised on a roadmap. Gemini chose the replacement. None of the merchants whose items sat in that basket had any say in which one it picked. For years substitution has been treated as a merchandising courtesy, the "customers also bought" rail or the back-in-stock email. Universal Cart turns it into a live surface where an AI model decides, at the moment of purchase, which product stands in for which. The only question left for a brand is whether it authored the logic behind that decision or let the system infer one on its behalf.

---

Substitution shipped, and Gemini makes the call

Google's account of the behavior is exact. The moment an item lands in the basket it "gets to work in the background," and when a shopper builds a custom PC by adding parts "from several retailers," the cart "will proactively flag any product incompatibilities and suggest alternatives." The whole thing runs on Gemini models, so by Google's own framing it "gets even smarter as the models improve."

The system reasons across merchants, identifies a conflict, and proposes a replacement, and the compatibility verdict in the demo belonged to the model rather than to any store. PCMag's reviewer caught the obvious risk and named it plainly: "we wonder if Universal Cart had guardrails to prevent possible errors." A substitution can be wrong, and when it is, the brand swapped into the basket wears the consequence while the session rolls on toward checkout. An agent that recommends a poor alternative damages the merchant it just promoted, in front of a shopper at the exact moment of decision. The day a reasoning model began making that call live, substitution stopped being a quiet cross-sell and became a competitive surface.

---

A PC build is an apparel problem with the labels swapped

The demo used computer parts because compatibility is legible there: a socket either fits or it does not, and "Ryzen 7 requires AM5" is checkable in a single sentence. Apparel has the same structure with softer edges and the same exposure. Being visible to the agent at all is the first hurdle, and what the agent recommends once a brand clears it is the second.

Picture a shopper whose first-choice jacket has sold out, or reads wrong for the occasion they typed in: "something for a summer wedding that also works back at the office." An alternative gets suggested either way. If the merchant wrote the reasoning down ahead of time, the basket can carry it, recommending the cotton-blend in the same cut when the linen blazer is gone because it still reads dressy and breathes in July heat. Without that authored rule, the model improvises from whatever it can see (price band, category label, a similar silhouette) and may swap a linen wedding blazer for the nearest-priced navy wool. Same price. Wrong garment.

The shopper never notices the gap, because what they see is a confident recommendation from a system they have started to trust. The merchant stays in the dark just as long, until a return shows up or the sale simply fails to convert. The whole swap happened inside a layer the brand was never given a field to write into.

---

Availability is a checkbox. Substitution is a judgment.

Bessemer Venture Partners called substitution a "strategic battleground" in the intelligence layer of the agentic-commerce stack, and it is the cleanest illustration of why that tier matters at all. A product feed carries an availability boolean: it reports whether an item is in stock and stops at the edge of that fact. Nothing in it captures the merchant's view of what to put forward when the honest answer is "out of stock" or "wrong for this prompt." The model fills that silence on its own, because it still owes the shopper a recommendation.

A merchant's judgment about which item satisfies a stated constraint draws on knowledge a system can only approximate from surface features. A catalog with four similar jackets is full of opinions. It knows which one leads for "casual weekend," which is the right stand-in for an out-of-stock "gift for dad," and which a petite shopper asking about travel should be steered toward ahead of the oversized cut. A merchandiser makes those calls every season without thinking twice. The feed has nowhere to hold them, so they never reach the basket, and the model writes its own version instead.

This is the difference between being found and being chosen. What AI agents need from your catalog covers the data side of that gap, and substitution is its sharpest edge, the lone moment when the model has to pick a product the shopper never named. Whoever authored the logic in that instant decides the sale. Product feeds were never built to carry that reasoning, and bolting on richer schema fields leaves it exactly as empty.

---

Merchant of record is a billing status

Google goes out of its way to reassure the seller on one point. "No matter which way you buy," the announcement reads, "the brand stays the merchant of record." The sentence does real work, promising that the transaction, the receipt, and the returns ledger still belong to it, and that checkout itself runs through the Universal Commerce Protocol that Google co-developed with Shopify and other retail leaders.

That billing status settles who charges the customer and who handles the return. It says nothing about who decides what gets surfaced and what gets swapped, and that authority now sits with the cart. A seller can hold the billing relationship on a sale it would never have recommended, for a substitute it would never have chosen, and still appear on the receipt as though it made the call. That seam between owning the transaction and owning the reasoning is the whole shape of the agentic shift: the platform concedes the relationship on paper while keeping the surface where the decisions actually get made. Substitution is where the concession gets tested first, because it is the place the model most visibly acts for a brand that is nowhere in the room.

---

Author it once, and it travels

The substitution logic a merchant writes lives above any single basket. It belongs to the authored reasoning that travels through the protocol layer into every agent surface reading the same catalog. Write the rule once, the swap to recommend when an item is gone and the reason it fits, and it holds whether the buyer arrives through Universal Cart, a side conversation with Gemini, or a rival agent entirely. Leave the rule blank and each surface invents its own answer in isolation, with no merchant anywhere in the loop.

How AI shopping agents make decisions gets into why these systems reason rather than rank, and why that changes what a catalog has to supply. The merchandising consequence lands sooner than most strategy decks assume: substitution authoring is a live competitive surface today, already running inside the apparel and beauty pilots that include exactly the kind of company reading this piece. The cart suggests alternatives right now, and the only variable still open is whose reasoning it borrows when it does.

A brand that has authored that reasoning gets recommended on its own terms, while a brand that has left the layer empty gets whatever the model infers in the moment. Both appear as the merchant of record. Only one of them is doing the merchandising.

---

Sumit Jagdale is the founder of Sartorial.