Product data for AI shopping agents: feeds, markup and protocols

Why does product data decide visibility in AI shopping assistants?

AI shopping assistants recommend products by matching a buyer's request to structured attributes, so a product without those attributes cannot be matched. A buyer asks for a 60 cm stainless steel extractor hood under a set budget. The assistant can only shortlist products whose width, material, type and price it can read as data.

The inputs are product feeds and page markup. OpenAI publishes a product feed specification for merchants. Google takes product data from Merchant Center and from schema.org markup on product pages. Both describe the same facts: identifiers, title, description, brand, price, availability and descriptive attributes.

Three properties of the data decide visibility:

• Completeness: the attributes buyers filter on are filled.
• Consistency: the feed, the product page and the markup state the same values.
• Validity: identifiers pass checksums and units are explicit.

This is a data model problem first and a channel problem second. The AI in PIM pillar makes the same point for AI features inside the PIM. A company with attributes locked in PDFs or free-text descriptions has no structured facts to send. A PIM with a defined data model holds them once and feeds every channel.

Claims about AI shopping traffic and conversion circulate widely. PIMvendors found no independent, verifiable figures to cite, so this page sticks to what the published specifications require.

Which protocols and feeds connect products to AI agents?

Two protocol families and two data formats matter today: OpenAI and Stripe's Agentic Commerce Protocol, Google's Universal Commerce Protocol, product feeds and schema.org markup. The protocols handle checkout between agent and merchant. The feeds and markup carry the product data.

At its announcement, ACP powered Instant Checkout in ChatGPT for US users buying from US Etsy sellers, with Shopify merchants listed as coming soon. UCP supports APIs, Agent2Agent and the Model Context Protocol as transports, and works with Google's Agent Payments Protocol. Over 20 partners endorsed UCP, including Adyen, Mastercard, Stripe, Visa and Zalando. Agents acting on the merchant side, inside the PIM, raise separate risks, covered in Agentic PIM: a reality check.

Which attributes do AI shopping feeds require?

OpenAI and Google require the same core fields and reward the same optional ones. A PIM that holds these attributes in validated form can serve both from one source.

Both specifications warn against invented values. OpenAI's reference says not to invent an MPN to replace a missing GTIN, and to leave unknown optional values out. Google says not to guess a GTIN and notes that an incorrect GTIN leads to disapproval.

Optional fields carry the matching power. Width, material, colour, compatibility and capacity are what buyers ask about. Google's product_detail attribute accepts up to 100 entries per product, which suits technical catalogues. Every one of those values should come from an approved PIM attribute, never from generation.

Why must feed, page markup and PIM agree?

Feed, page and markup must agree because systems compare them and penalise conflicts. Google states that it may verify product data before displaying merchant listings. A price or availability on the page that contradicts the feed undermines trust in both.

Schema.org Product markup is the page-side source. Google's merchant listing documentation requires name, image and an Offer with price and currency. It recommends brand, GTIN, MPN, SKU, colour, material, size and other properties. The schema.org Product type also offers additionalProperty for attributes without a dedicated property. The schema.org Product page covers the vocabulary.

Consistency starts in the PIM:

1. One approved value per attribute, used by the feed, the page template and the markup.
2. Stable identifiers per product and variant, never reused or changed with price.
3. Explicit units on every measured value, converted per channel on export.
4. Price and stock from the system of record, usually the ERP, refreshed when they change.

OpenAI's reference adds that feeds should be updated when prices or availability change. Price and stock rarely live in the PIM, so the integration design matters. The product data syndication page explains the patterns.

What should a PIM selection require for AI shopping readiness?

A PIM selection should require validated identifiers, structured descriptive attributes, explicit units and multi-format export. These requirements serve marketplaces and search engines today and AI agents as they mature.

Requirements to add, grouped by checklist domain:

Test the result in a demo. Ask each vendor to produce an OpenAI-format and a Google-format feed for 20 of your products and the matching schema.org markup. Compare the three for consistency. The product data quality page explains how to measure completeness per channel.

What is verified today and what remains open?

The specifications are published and verifiable; adoption, regional availability and commercial impact are still open. Both protocols are young, and decisions should rest on the data fundamentals they share.

Verified as of this page's date:

• ACP is open source under Apache 2.0, announced by Stripe and OpenAI on 29 September 2025.
• UCP is open source, announced by Google on 11 January 2026, with its specification on GitHub.
• OpenAI and Google publish product data specifications with overlapping required fields.

Open questions for buyers to track:

• Availability of agent checkout for European merchants and markets.
• How protocols handle B2B pricing, configurable products and customer-specific assortments.
• Whether further protocols emerge or existing ones converge.

The safe investment is the same under any outcome. Structured, complete and validated attributes in a PIM serve every feed format and protocol. Avoid buying a tool for one protocol before the data model is ready. The AI channel-specific descriptions page covers the text fields these feeds require. In licence talks, check that feed connectors and export volumes are priced for growth; the negotiation guide shows how.