The line at the top of the product row reads “Functional hydration, lower sugar,” set in the weight and position a merchandiser’s headline used to occupy, and no merchandiser wrote it. It is a sample output from Instacart’s cascaded merchandising framework, which generates a module’s theme first and then the constrained keywords that retrieve products into it. Every product image beneath that line is therefore downstream of a sentence a model wrote about you. Instacart reports the arrangement lifting cart adds per page view by an estimated 2.7% over a strong baseline, tested across hundreds of thousands of users against a catalogue of tens of millions of items.

The guess underneath is now itself a research subject. A pipeline posted to arXiv on 11 August by five JPMorgan Chase researchers opens by rejecting the standard reading of click data: observed actions get treated as reliable proxies for preference when they often record exploration or comparison instead. Their Inverse Theory of Mind method runs the inference backwards, rebuilding what a shopper saw and what they passed over, then generating natural-language statements about the beliefs that would explain the choice. The authors are direct about the timing. Interfaces are moving from static layouts toward generative UI and extended reality, and those environments must decide not only what to present but “where, when, how prominently, and most importantly why a user acts.”

On the OPeRA dataset of recorded Amazon sessions, browser traces and post-session interviews, the persona the model invented predicted a shopper’s next action more often than the profile OPeRA built from that shopper’s own interview, demographics and rationale: 25.69% against 17.06%, with Claude Opus 4.6 doing the reasoning. The invented persona was wrong about three times in four; the interview-built one, closer to five times in six.

What generation takes away is adjacency, and adjacency is where fashion has always done its work. Alexey Brodovitch art-directed Harper’s Bazaar from 1934 to 1958 on the principle that the unit of design was the spread rather than the page, so that two facing images argued with each other and the sequence across an issue built a rhythm the reader felt without naming. A shop rail runs on the same logic, and so does a window, and so does the six-image grid at the top of a category page. Someone decided what sits beside what. A page assembled per shopper from generated themes has no spread and no author, because no two people are shown the same adjacency.

Instacart’s paper argues that the old stack of fixed themes and pointwise rankers limits “semantic cohesion across the page,” and that generation restores it. Cohesion is the danger. A jumbled grid reads as a shop and the eye discounts it; coherence reads as a point of view, and a point of view assembled from a misread of your afternoon is one you did not ask for and cannot argue with. The failure has changed shape, from a single bad recommendation among decent ones to a whole page styled around a wrong idea of who you are. The correction the shopper types next is the signal we looked at this morning, read from the agent’s side.

None of this waits on a roadmap. Google put generative interfaces into the Gemini app and AI Mode in November 2025, where the model writes the layout and generates the imagery that fills it. A generative UI venture between WPP and Google is already running, listed by Glossy among the tools agencies sell against zero-click search. Instacart’s safeguard is the tell: a content-evaluation and quality-filtering layer screens the output before it ships, and deterministic fallback catches whatever gets through, with the old static store still sitting underneath, waiting to take over when generation fails or scores badly.

Handle the psychology claim carefully. The JPMorgan pipeline was validated against 46 user profiles, drawn from a dataset holding 692 sessions and 51 people in total, which is a small room in which to conclude that anything understands why we shop. The generation is running at scale regardless. Within a year a shopper will open a category page where the headline, the products under it and the crop of the lead image were all set by a model’s reading of their mood that afternoon, and the store will look composed. Nobody will have composed it.