The size chart is the one part of a product page a human can skim past and an agent cannot. A person orders two sizes and keeps one, filing the wrong guess under the cost of shopping online; an agent spending its owner’s money to a budget must get the size right the first time. As shopping agents move from suggesting purchases to making them, machine-readable size-and-fit data becomes the variable that decides both conversion and returns, and the brands that never built fit intelligence are the ones agents will quietly pass over.

The agent that buys on your behalf is no longer a slide in a keynote. OpenAI and Stripe shipped the Agentic Commerce Protocol with Instant Checkout after an earlier checkout feature that failed to stick; the protocol, not the feature, is what set the race in motion. Google co-developed a Universal Commerce Protocol in January and announced a shared cart at Google I/O in May that follows the shopper across Search, Gemini and Gmail. Amazon has pushed its Alexa for Shopping assistant (formerly Rufus) past summarising products and is now handing the same agent technology to other retailers. Each runs on the same input: a structured product feed carrying current price, stock, images, descriptions and shipping data. It holds no standard field for whether a size 10 fits like a size 8 — the one attribute that decides whether the box comes back.

Returns are the cost these agents exist to cut, and fit is what drives them. The average US online apparel return rate reached 23.4% in 2025, according to a Coresight Research report — the accessible summary was published by FashionUnited under sponsorship from Alvanon, a sizing-technology vendor — and its cause is unusually concentrated. Nearly 70% of the shoppers behind those returns named size and fit, which makes measurement, not taste, the largest recurring cost in online clothing. An agent working to a budget and graded on whether the box stays shut does not indulge a label whose sizes come back at that rate; it learns to route around it.

The same gap shows up on the other side of the ledger, in sales that never close. Coresight puts the share of AI-aware US consumers who already use or plan to use AI tools to shop at 58%, a demand curve that did not exist three years ago. The same population walks away easily: roughly 40% of US shoppers have abandoned an apparel order over confusing or missing product information, and nothing goes missing more reliably than a size a machine can act on. For the agent, an unreadable size chart is not a smaller sale; it is an unmade one, since it will not buy what it cannot justify.

The skip is a data problem before it is a commercial one. Most apparel catalogs were built for keyword search, and sizing was, in Bold Metrics’ own words, “added as an afterthought: a static size chart linked from a product page, formatted for humans to read, not for machines to reason about.” An agent that cannot parse that chart cannot price the risk of recommending you. Coresight is blunt about where that leads: inaccurate sizing yields lower recommendation confidence and, once returns climb, outright brand deprioritization. A human who can’t find your size chart might still buy on hope; an agent that can’t read it moves to the SKU it can.

An agent that skips you files no complaint, abandons no cart, and surfaces in your analytics as weather.


The brands that see this are already turning fit from a document into a service. In March, Gap Inc. began testing Bold Metrics’ Agentic Sizing Protocol™ and said it would place checkout inside Google’s Gemini, which CNBC called an AI first for a major fashion company. The protocol is small on purpose: an agent calls it, puts a few questions to the shopper, and gets back a size and a note on how the garment will sit. True Fit has done the same with more than a decade of purchase-and-return data, repackaging it as an intelligence layer that other AI systems query rather than a chart a shopper squints at. That Gap, a mass-market incumbent with no special claim to fit science, is paying to pass this gate is the tell: the gate is real, and standing at it costs money.

The strongest objection is that the fit intelligence need not live in the brand at all. If an agent can learn from millions of transactions that a label runs a half-size large, it can correct the size itself, and a brand with a chaotic size chart pays no penalty. This is not hypothetical: True Fit’s corpus spans tens of thousands of brands, and Amazon grades sizes against its own mountain of returns. But inference needs volume, and volume is what the new, the niche and the independently hosted lack; the agent that has never sold your jacket has nothing to infer from. Bodies are moving, too, and Coresight notes that 70% of US GLP-1 users have dropped at least one clothing size, which stales the historical prior at the moment it is consulted. The dependency moves; it does not disappear, and it lands hardest on the brands least able to buy their way past it.

None of this is a model problem, which is the part the industry keeps getting wrong. The rendering engines can already put a coat on your body, as we argued when Taobao’s try-on system shipped. The shopping agents pass in simulation and slip when real shoppers grade them, as we reported, and fit is precisely the variable a simulation cannot feel. In both, the intelligence is ready and starved of the one input it needs, a measurement it can act on. A brand can keep its fit knowledge as a PDF a person skims once, or as a number an agent can trust before it spends, and the sorting between those two states is already running. The agents will not tell the losers they lost; they will keep buying elsewhere, quarter after quarter, and the brands they skip will file it under soft demand.