The next moat in commerce isn't the catalog. It's the model of the customer.
Every brand's product data is about to be clean, structured, and readable by AI. That makes it worthless as a differentiator. The asset that compounds is on the other side of the query.
Here is a prediction you can hold me to. Within eighteen months, the product data of every serious beauty and fashion brand in the US will be clean, structured, and legible to a language model. Not because brands will have done something clever, but because the tools to do it are cheap, the inputs are the same for everyone, and the platforms will do most of it for you. Shopify will. Google will. So will we.
When that happens, a very large amount of money will have been spent making product data a non-factor. A model asked to recommend a long-wear foundation will have a hundred candidates that all say the same thing, in the same schema, with the same completeness. The catalog will have stopped being a way to win.
So what will the model pick on?
The half of the query nobody is optimizing
Look at how people actually ask. “I have combination skin with a warm undertone. I need a foundation that holds up through a 12 hour outdoor wedding in Austin in August. Medium coverage, natural finish, under $60.” Six words about the product. Thirty about the person.
Nearly all of the effort in AI search optimization goes into the six words. Descriptions, attributes, schema, llms.txt. That is the product half, and it is necessary. It is also the half that converges, for the reason above.
The thirty words are a different problem. To act on them, a model needs a structured understanding of a person: skin type, undertone, body shape, style sensibility, the climate she is in, the occasion she is dressing for, the budget she set, the thing she tried last time and returned. None of that is on a product page. None of it can be, because it is not about the product.
That understanding, held in a form a model can reason over, is what I mean by a model of the customer. It is not a persona or a segment. It is a representation of one specific person, assembled from what she said and what she did, that can be matched against a product’s actual traits to produce a shortlist with reasons.
Why it compounds and catalogs don’t
A catalog is a description of inventory. It is complete the moment the last SKU is tagged, and it is identical in value to the catalog next door once both are tagged. A customer model is different in two ways.
First, it is built from interactions, so it gets better every time it is used. Every query, every click, every return teaches it something about how people describe themselves and what that maps to in a product. The brand with more matched interactions has a better model, and a better model produces more matched interactions.
Second, it is only useful in combination with product understanding deep enough to match against. “Warm undertone” means nothing unless the model also knows which of your 50 shades carry warm pigment, and that is not in the shade name. It comes from computer vision run on the product itself, extracting the 200 plus physical and style traits a photo contains and a feed does not. The two halves need each other, and building both is hard, which is the whole point. Hard is where moats are.
Why this gets built in fashion and beauty first
Electronics has specifications. A laptop either has 16 gigabytes of memory or it doesn’t, and the match is a lookup. Fashion and beauty have nothing like that. Fit depends on body shape and the garment’s cut together. Foundation shade depends on undertone, lighting, and season. Whether a dress works for a fall wedding depends on the wedding, the weather, and the woman.
That is why generic recommenders bolted onto apparel have always disappointed, and why return rates in the category sit where they do. It is also why a shopper model that works here will work anywhere. If you can match a person to a foundation, matching a person to a mattress is a smaller problem. The hardest category is the right place to build the general capability.
What this means for the next two years
Agentic commerce is early. Today an assistant recommends and a shopper clicks. Within a few product cycles, the assistant will compare, decide, and complete the purchase, and the brand’s site will never load in a browser. At that point the entire funnel collapses into one step: did the agent’s matching function select your product for this person.
Brands that invested only in the product half will be indexed. They will be in the candidate set, alongside everyone else who did the same work. Brands with a customer model will be chosen. Indexed is table stakes. Chosen is revenue, and the gap between them is the model of the customer.
That is what we are building at Veristyle, and it is why we describe ourselves as a Product-Customer Intelligence Layer rather than a GEO tool. GEO is one of the places the layer shows up. It is not the layer.
See the layer on your own catalog
Ask an AI assistant a question that describes a person in detail and your product barely at all. If it names you, read why. If it doesn’t, read why it named someone else. Either way, the reasons are a specification for what you are missing. A free AI visibility audit runs that check across your whole catalog, or book a demo and we will show you what the layer sees in your products.