If you sell footwear or apparel online, you have probably watched the acronyms pile up over the past eighteen months. MCP. A2A. ACP. AP2. UCP. x402. Most people I speak to in retail have quietly decided this is someone else's problem, or have picked one and hoped it was the right horse.
Both reactions are understandable, and both rest on the same wrong premise. These are not competing standards. They are floors in a building.
The four floors
- Communication How an AI model reaches anything outside itself — a database, an API, a product catalogue. The Model Context Protocol, released by Anthropic in late 2024 and since moved to neutral governance under the Linux Foundation, has become the default. Think of it as the socket in the wall. It is not a commerce protocol at all; it is what commerce protocols travel through.
- Commerce How an agent discovers a product, builds a cart and triggers a purchase at a specific merchant. This is where the Universal Commerce Protocol and the Agentic Commerce Protocol sit, and it is the only floor where two standards genuinely compete for the same ground. UCP arrived at NRF in January 2026 with Google, Shopify, Etsy, Wayfair, Target and Walmart behind it, and a long list of endorsing payment and retail names. That breadth is the story.
- Authorization How anyone proves that a human actually mandated this particular purchase, within these particular limits. AP2 handles it with cryptographically signed mandates. Unglamorous — and the reason a bank will ever let an agent transact at all.
- Settlement How value moves machine to machine. Safe to ignore for now.
Hold that shape in your head and the "which one wins" argument mostly dissolves. A single purchase can touch all four.
The stack as it stands, and the floor that is missing from it
What the merchant now has to hand over
Here is where it gets concrete, and where I think footwear should be paying closer attention than it is.
To make a product eligible for agent purchase on Google's surfaces, a merchant has to supply structured data that until recently lived in a PDF or a link in the footer. Return policies are mandatory — windows, costs, policy links, resolvable down to product level. Shipping and tax must be configured well enough that the total cost is computable instantly. Agents require cost certainty before they act. If the agent cannot compute the final bill, it does not open the checkout.
This is a genuine and overdue improvement. Operational promises become data. Folklore becomes structured logic.
But read the list again.
Every one of those attributes describes what the return will cost. Not one of them describes whether the return will happen.
The variable nobody modelled
Across published industry analyses, somewhere between forty and seventy percent of fashion and footwear returns are driven by size and fit — not damage, not delivery, not a change of heart. Footwear sits at the punishing end of that range. Fit is not one return driver among several. In this category, it is the return driver.
So an agent buying a pair of boots on your behalf can tell you, to the last öre, what it will cost to send them back. It can tell you the return window in days. It can tell you who pays the return shipping.
It cannot tell you whether the boots will fit the person it is buying them for.
It has no representation of that person's body. There is no field for it. There is no schema for it. There is no floor in the building for it.
Why the obvious fixes aren't
The first reflex is: the agent already knows the shopper's size. It's in the order history. They're a 41.
But a size label is not a measurement. It is a claim made by a factory about a last — the three-dimensional form a shoe is built around — and lasts differ between brands, between models inside a brand, and between production runs. This is why you are a 39 in one brand and a 40 in another while your foot has not changed at all. Feeding size labels to an agent does not give it body data. It gives it the same unreliable proxy that produced the returns problem in the first place.
The second reflex is: use purchase history. What did people like this keep? That works, up to a point, and it is what most of the existing size-recommendation industry is built on. But it is correlation. It tells you what happened, not why. It degrades on new products, on new brands, and on any shopper without much history — which is precisely the situation an agent shopping across an open catalogue is usually in.
What the missing layer would have to be
If a body-data layer is going to exist, several of its properties follow from the architecture rather than from anyone's business plan.
- Neutral An agent shops across brands and retailers. A body representation locked inside one retailer's app is invisible to it.
- Portable and consented The profile belongs to the person, not to the store. The agent fetches it with permission. This is also, conveniently, what European data law would prefer.
- Causal Geometry against geometry — the foot against the actual form the shoe was built on — so that the answer holds for products with no purchase history behind them.
- Rail-agnostic It cannot assume UCP wins, or ACP, or whatever arrives in 2027. It has to be callable from any of them.
None of that is exotic. It is a machine-readable body profile, and a fit verdict that any checkout can request.
Where this goes
The protocols are being written right now, in public, in specifications anyone can read. Whoever defines how a body is described to an agent will shape this category for a long time — in the same quiet way that whoever defined how a product is described to a search engine shaped the last one.
At the moment, nobody is writing that part.
Agentic shopping has a body problem. The plumbing is nearly finished. The body is still missing from it.
← All insights