Agentic commerce.

An agent does not browse. It queries, ranks, and picks. When OpenAI rolled checkout into ChatGPT in 2025 via the Agentic Commerce Protocol with Stripe, and Shopify exposed a Storefront MCP server that lets any model read a catalog and place an order, the shape of the buyer changed [1][2]. The next high-intent shopper is software acting on behalf of a person. The storefront it sees is not the one a human sees.

The new stack

Four layers matter, and most brands have none of them production-ready.

• An agent-readable feed. Structured product data — schema.org Product, Offer, AggregateRating, plus the new agentic-commerce extensions — is what an agent reads first. If the feed is thin, the agent picks a competitor whose feed is thick [3].
• Spinnable, inspectable assets. Agents that are doing a comparison surface 3D, AR, and multi-angle stills back to the user. Flat hero shots lose. Shopify's native 3D viewer and Apple's USDZ pipeline are already the floor [4].
• Machine-trustable signals. Verified reviews with structured markup, returns and shipping data exposed as fields rather than buried in PDFs, and a public sustainability disclosure that an agent can parse. Trust is now a field, not a feeling [3].
• A checkout an agent can complete. Agentic Commerce Protocol, Shopify MCP, and Stripe's delegated payments mean a model can place an order without a human in the browser. Brands that gate checkout behind interstitials, captchas, or human-only flows will not be picked [1][2].

Why the timing is now

Gartner projects that by 2028, agentic AI will autonomously make 15% of day-to-day enterprise work decisions, up from effectively zero in 2024 [5]. Consumer-side, the Adobe Digital Economy Index already shows generative AI traffic to US retail sites grew over 1,200% year-on-year in late 2024, and those visits convert at competitive rates to traditional search referrals [6]. The volume is real, the infrastructure is real, and the protocol is published. What is missing is the brand-side readiness.

What an agent rewards

An agent does not have brand loyalty. It has a ranked list, a budget, and a return policy. The brand the agent picks is the one that made itself easiest to pick.

We see four patterns in the feeds agents prefer:

1. Single canonical product entity with stable IDs across web, marketplace, and ad surfaces. Re-SKUing for every channel is a tax the agent makes you pay.
2. Rich media bundled with the product node — 3D models, dimensional data, materials, regulatory data — not stored in a separate "assets" silo the agent never sees.
3. Plain-language attribute coverage: "machine washable," "dishwasher safe," "contains tree nuts." The model reasons over claims, not category trees.
4. A returns and warranty policy expressed as structured data, because agents weigh post-purchase risk explicitly in the ranking step [3].

What we build

We build the asset layer (spinnable 3D via HYP3D), the entity layer (structured product feeds and schema.org coverage), and the agent-readiness layer (Storefront MCP integration, ACP-compliant checkout endpoints). The output is a catalog an agent can read, trust, and transact against in one pass.

The takeaway

The next category default will be set by the agent's first answer. Build for that answer, in the agent's grammar, before the category is decided without you.