The order of operations in apparel discovery has reversed. A shopper states the requirement to an assistant in plain language, and the assistant assembles the candidate set before any retailer’s site opens. Ranking, the function retailers own and monetise, has been demoted to sorting a list that was cut somewhere else. Apparel’s discovery budget is aimed one stage downstream of the decision it was bought to influence.

The traffic already shows the handoff. Adobe Digital Insights put AI-driven visit share in US retail up 138% year over year in May 2026, its highest level since the series began in October 2024. Nothing seasonal was pushing that figure. Adobe’s March consumer survey found 39% of US shoppers had used an AI assistant for shopping. The argument does not need a majority, only a shortlist that forms before the visit.

The value of an assistant-referred visit shows what the assistant did before sending it. Adobe has those shoppers generating 53% more revenue per visit than non-AI traffic in May. A year earlier the comparison ran the other way, and by a wide margin: non-AI visits were then worth 128% more. A sign flips like that when the comparing and the eliminating have already happened offsite, in a conversation the retailer never sees.

The bill for inclusion is now itemised. OpenAI charges Shopify merchants 4% on sales completed through Instant Checkout, a rate first reported by The Information and relayed by PYMNTS. Google’s AI Mode and Gemini, and Microsoft’s Copilot, charge nothing for now. Shopify switched all three channels on in January, with the paid one opt-in. What matters is the shape of the charge: a company holding no inventory takes a share of a completed sale for a place in an answer, while two competitors give the same place away to build the habit.

Retail media is the arrangement being displaced, and it is enormous. eMarketer’s H1 2026 forecast has Amazon’s retail media revenues passing $75bn by 2028, a lead over the second-largest network of more than $65bn. That money buys placement among goods the platform already stocks, billed before anyone converts, to a shopper who has already walked in. When the requirement is spoken before the store is picked, that budget is buying the wrong stage.

Apparel is the category least equipped to compete for inclusion, because inclusion is not bought with money.

Adobe scores retail sites on what it calls AI citation readability, and apparel lands at 51%, behind both cosmetics and electronics. Ingredient lists and spec sheets parse cleanly for an assistant; descriptive brand copy does not. Apparel wrote its product pages for a reader who could see the photograph — the new reader sees only the text. What is rationed here is legibility, and legibility has no rate card.

The assistant has its own reasons to keep the pool small. Microsoft Research’s Magentic Marketplace, an open simulation of a two-sided agent market published in October 2025, found every model tested showed severe first-proposal bias, worth 10 to 30 times more advantage from response speed than from offer quality. Welfare in the simulation also fell as the number of search results rose. A system that degrades on long lists will shorten them. If that behaviour survives into production, the price of inclusion is paid in milliseconds and clean fields, and it is paid to engineers rather than media buyers.

The thesis fails if the candidate pool is simply re-owned one layer up. Google’s Shopping Graph carries over 50 billion product listings, on Sundar Pichai’s account at NRF in January, and the Universal Commerce Protocol he announced there was built with Shopify, Etsy, Wayfair, Target and Walmart. If retrieval runs through four or five indexes held by the same incumbents, intent capture has changed the landlord and left the tenancy intact. The new holders of the index, though, do not hold the stock. Google’s own merchant documentation describes UCP checkout as happening on Google’s surfaces “while keeping you the merchant of record”, which leaves the intermediary earning on completion or not at all. A gatekeeper ranking goods it does not own cannot prefer its own margin over fit, and cannot stop a merchant listing in every index at once.

The one player defending the old arrangement is doing it in court. Amazon won a preliminary injunction in March barring Perplexity’s Comet agent from its logged-in pages, with a stay built into the ruling to let Perplexity pursue the appeal it has since taken to the Ninth Circuit. Litigation is what owning a candidate pool looks like once intent is captured somewhere else.

Which way the money finally runs is still open; which way control has moved is not. Filippos Ventirozos and Matthew Shardlow argued in June that agent-native payment rails change what is scarce: once the buyer can investigate exhaustively, the bottleneck stops being the match between shopper and catalogue and becomes trustworthy information about the goods. In their design, buyer agents spend fractions of a cent to unlock test reports and bills of materials, which points the payment the opposite way from OpenAI’s fee. It is a proposal with no empirical results, and it still lands where the traffic data lands: such a market would reward product quality and yield “truer competition than ranking-based storefronts”.

The choice facing apparel is a budget question. Money spent on ranking buys the right to sort a list the shopper has already been shown. Legibility, latency and verifiable product facts buy something else: entry to the list, at a price no competitor’s media budget can outbid. One of those lines already exists, with a committee behind it and a dashboard attached. The other starts with conceding that the store is now the last thing chosen rather than the first.