Products Need Two Descriptions Now

A product can be obvious to a person and ambiguous to software. A photograph makes the shape of a sofa clear. A “weather ready” badge implies something about a jacket. A size guide hidden behind a tab resolves fit. A promotion banner can be understood from context even when the eligibility rules live elsewhere.

Humans are skilled at assembling these fragments into a judgment. Agents are much less forgiving.

That creates two legitimate descriptions of the same product. The human description should persuade: imagery, narrative, positioning, proof and the details that help someone imagine ownership. The machine description should disambiguate: identifiers, attributes, variants, dimensions, compatibility, price, stock, restrictions and policy relationships.

These are not competing content strategies. They serve different readers.

From page content to product representation

The existing web already contains part of this second layer. Schema.org and Google’s merchant structured data allow retailers to expose price, availability, shipping and return information explicitly. Merchant Center feeds provide a normalized product surface. The shift toward agents pushes the idea further.

Shopify’s 2026 Catalog API turns products from millions of merchants into structured, queryable data for AI systems. The company says searches powered by its normalized catalogue convert at twice the rate of searches using scraped data. That figure comes from Shopify and should be treated as platform evidence rather than a universal benchmark, but the direction is plausible: clean product identity and current commercial data reduce errors.

Visa is moving in the same direction from the payments side. Intelligent Commerce Connect includes a service that creates AI-ready catalogue representations and connects them to agent platforms. This is striking because a payment network now treats product discoverability as part of the infrastructure required for agentic transactions.

Machine readability becomes operational

Static metadata is only the first layer. A product description may say “in stock” while the relevant store has just sold the last unit. A loyalty member may see a different price. A promotion may apply only above a threshold or only to selected variants. Certain items may be unavailable for a delivery method.

An agent that merely recommends can tolerate some uncertainty. An agent that acts cannot.

Retailers therefore need to distinguish published facts from live commercial state. Public structured data can explain what the product is. Merchant APIs establish whether it can be bought now, by this customer, under these conditions.

This is also why protocols such as UCP are more consequential than another schema field. They move from representation toward capability: what the merchant allows an external system to do.

A merchandising problem disguised as data plumbing

The shift has consequences for brand teams as well as engineers. When an AI constructs a comparison, vague differentiation becomes fragile. “Travel friendly” is marketing; exact dimensions determine whether a bag fits an airline rule. “Suitable for sensitive skin” is positioning; ingredients, allergens, certifications and test conditions provide stronger evidence.

The answer is not to turn every product page into a spreadsheet. Human persuasion remains valuable. The answer is to ensure that the persuasive layer is backed by an explicit product model that software can interrogate.

Retailers have historically treated product data as back-office hygiene and the product page as the storefront. In an agent-mediated market, the product model itself becomes part of the storefront. Missing data does not merely create a bad page. It can remove the product from a machine-generated comparison before a person sees it.

A product story persuades a person. A product representation must survive a query.

“Catalog API turns Shopify’s global product catalog into structured, queryable infrastructure for AI agents.” — Shopify