How a shopping agent actually decides. The matching step, explained.
Checkout protocols from Shopify, OpenAI, and Google solve the transaction. None of them solve selection. Here's what happens between a shopper's question and an agent's pick, and what your catalog has to supply for that step to go your way.
There has been a lot of announcement traffic about agentic checkout. Shopify’s agentic commerce tooling, OpenAI’s commerce protocol work with payment partners, Google’s agent payments efforts. All of it matters, and all of it answers the same question: once an agent has decided what to buy, how does the money move?
That is the easy half. The hard half comes before it, and nobody’s protocol touches it: how does the agent decide what to buy in the first place? This post is about that step, in plain terms, because it is the step your catalog either wins or loses.
Start with a real question
“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. Which should I get?”
Before an agent can rank a single product, it has to take that sentence apart. Watch what it becomes.
About the person: combination skin, warm undertone, attending a wedding, outdoors, 12 hours, August, Austin (so heat and humidity).
About the product: foundation, medium coverage, natural finish.
Hard constraints: under $60, available to purchase now.
Seven facts about her. Three about the product. Two constraints. That ratio is typical, and it’s the first thing to notice: most of the information the agent has to act on isn’t about any product at all.
What the agent needs from the catalog
Now the agent goes looking for products it can match against those seven facts. For each one, it needs a product attribute to compare to. Here is what it needs and what a typical feed actually has.
- Warm undertone. Needs: which shades carry warm pigment. Has: shade names like “Sand” and “Honey,” which carry no undertone information a model can rely on.
- Combination skin. Needs: which formulas suit combination skin. Has: marketing copy that says “for all skin types,” which answers nothing.
- 12 hours, heat, humidity. Needs: wear performance under those conditions. Has: “24 hour wear,” tested in a lab, with no climate qualifier.
- Natural finish. Needs: a finish classification. Has: this one is usually present. One out of four.
- Under $60. Needs: price. Has: price. Constraints are the easy part.
The pattern is the same across fashion. “Flattering on a pear shape” needs a fit attribute no size chart carries. “Works for a fall wedding in Chicago” needs occasion and climate suitability that no feed field holds. The agent is being asked to match a person to a product using data that only ever described the product.
What happens when the data isn’t there
The agent doesn’t give up. It fills the gap with whatever it can find: reviews, editorial mentions, a competitor’s page that happens to say “great for warm undertones,” its own general knowledge about the brand. Then it produces a confident shortlist of two or three, with reasons.
If your product wasn’t in that shortlist, it’s usually not because the agent judged it and found it wanting. It’s because the agent couldn’t judge it at all. There was nothing in your data to match “warm undertone” against, so the product never entered the comparison for that shopper. You weren’t rejected. You were unanswerable.
What answerable looks like
The fix isn’t more adjectives. It’s a layer of structured, shopper-facing data that sits alongside the product data and answers the person half of the question. For that foundation, it looks like this: each shade tagged with its undertone from pigment analysis, not from the name. The formula tagged with the skin types and climates it performs in. Wear time qualified by condition. Finish, coverage, and the occasions each combination suits.
For a dress: the body shapes the cut flatters, the undertones the color works with, the occasions and climates it fits, the pieces it pairs with. Extracted from the product itself by computer vision, because that is where the information lives, and written back into the catalog in a form a model can reason over.
That is the matching layer. It’s what turns a product from legible into answerable, and it’s the layer we build at Veristyle for exactly this step.
Why the protocols don’t help you here
Agentic checkout is plumbing. It makes it possible for an agent to complete a purchase it has already decided on. If the decision went to a competitor, the plumbing carries the competitor’s order. Integrating with every checkout protocol on the market puts you in position to transact. It does nothing to put you in position to be selected.
The sequence is: decompose the question, match against the catalog, rank, then transact. The protocols own the last step. Your data has to win the second one.
Find out where you stand
Take your best seller. Write the most personal question a real shopper might ask about it. Give it to an AI assistant and see whether it can answer specifically or has to guess. A free AI visibility audit runs that test across your catalog and shows which products are answerable and which are only legible. Or book a demo and we’ll walk through the matching step on your own data.