The pixel buffer

Between 14 and 22 September, fashion announced its relationship with artificial intelligence four separate times.

The Estée Lauder Companies signed a global partnership with Profound, an AI marketing platform, to find out how its brands are being described inside generative AI systems and to change what those systems say. Daydream, the shopping app, shipped two features that read the outfit photos in your iPhone's camera roll and match them against roughly three million products from more than 325 retailers. ALICE + OLIVIA put Glance's virtual try-on at its Spring 2027 presentation during New York Fashion Week, so guests could watch themselves wear the collection as it walked. And Google built two bespoke tools in Flow for Jane Wade and Sergio Hudson — one to style looks on digital models before anything is cut, one to stage a runway before anything is built.

Four announcements, four different technologies, four different parts of the business. Line them up and three of them are the same announcement.

Estée Lauder is buying visibility at the point of discovery. Daydream is inserting itself at the point of discovery. Glance is converting an audience into shoppers at the point of discovery. Only Google's is pointed anywhere else, and it is pointed earlier — at the cost of making the thing in the first place.

Not one of the four is pointed at a garment after it is sold.

That is not a gotcha. It is what the category is for right now, and the companies are candid about it. Daydream's co-founder and CEO Julie Bornstein described the long-term ambition as a shopping agent that understands someone's style well enough to "work across every surface in their life." Read that sentence as an engineer rather than a customer: the surfaces are the point. Photos is a surface. Siri is a surface. Your lock screen — Glance's original home, through its Samsung deal — is a surface. Fashion AI in September 2026 is a land grab for the places a purchase decision can be intercepted.

Which brings us to the camera roll, because the camera roll is the tell.

Daydream's premise is that your saved photos are a wishlist. Some of them are. But a camera roll is mostly not a catalogue of things you want; it is a catalogue of things that happened, and a large share of the clothes in it are clothes you already own, on you, on a day you liked how you looked. The product reads that photo and returns a place to buy something adjacent. The most reliable inventory of a person's actual wardrobe that exists anywhere on earth is being mined, competently, for the purpose of selling them a fifth version of the coat in the picture.

Quick disclosure before we go on: I'm building a wardrobe cataloguing app, so this argument runs directly past my own front door. Read the rest knowing that.

I don't think this is cynical on Daydream's part. I think it is what the funding is for. Nobody in this wave is being paid to build the boring version — the one that looks at the same photo and says you own this.

Now the thing all four announcements leave out.

Apple's developer guidance for integrating visual intelligence is specific about the mechanics: when a user sends an image to a third-party app, the system hands that app a `SemanticContentDescriptor` containing a pixel buffer. The app receives the picture. Apple's session on this walks through doing the matching locally with the Vision framework — but that is a demonstration of what is possible, not a constraint on what any given app does, and matching a garment against three million SKUs is not obviously a local operation. None of the coverage I could find says where Daydream's matching happens or what is retained. It may be entirely on-device and admirable. The announcement simply does not address it.

Glance is the sharper case. Its platform generates photorealistic images of a specific human being wearing specific clothes, from a single selfie, using Google's Imagen 4, with the rendering tuned to body type, gender, ethnicity, skin tone and season. Glance describes the platform as fully opt-in with privacy built into its design. At a fashion week presentation, guests walked up and used it. The coverage of that activation — and I read several versions — describes the experience, quotes Stacey Bendet on imagination and possibility, quotes InMobi and Glance founder Naveen Tewari on the audience participating rather than watching. It says nothing about consent language, retention, or what happens to a generated likeness afterwards.

Again: not an accusation. An absence. The likeness question is the single most contested thing in fashion technology, and four AI announcements in eight days produced approximately zero sentences about it. That silence is a choice somebody made in a comms meeting.

Estée Lauder's deal is the most honest of the four, and the least discussed, because it is the only one that concedes something.

Aude Gandon, the group's global chief digital and marketing officer, said the way people discover beauty "is being rewritten in real time." The partnership exists so the company can see how its brands appear across AI answer engines and improve that representation — optimising product pages, blogs and video so language models surface them accurately. Profound's co-founder James Cadwallader frames it as treating generative engine optimisation the way good marketers treat any platform shift.

Strip the vocabulary and what Estée Lauder has publicly acknowledged is that it no longer controls the surface where its products are described. It is now doing search engine optimisation for a search engine with no results page, no rank, and no audit trail. Everybody in fashion is about to be in this position. Estée Lauder is just first, and said so.

Google's is the outlier and the most concretely useful. Wade used the styling suite to balance looks and spot gaps before cutting material; Hudson used the runway tool to simulate a set before paying to build it. Google's own account notes the styling work compressed what had been three full days of in-person casting and fitting. Read that either way you like — a genuine reduction in waste, or three days of somebody's labour — but note that nobody presented it as a sustainability story, which is what a tool that stops you cutting fabric actually is.

So: four announcements, and the shared assumption underneath all of them is that the interesting clothes are the ones not yet bought. Every technical component required to work the other way is already shipping — on-device image understanding, garment-level recognition, three-million-item taxonomies, generation good enough to put a real body in a real garment. It is all aimed one direction, and the direction is not a technical fact. It is where the money is.

Which is fine, and worth saying out loud, because in about eighteen months somebody is going to describe the other direction as a breakthrough.