Structured product data means every product fact lives in a typed attribute with a unit, a stable id…
Shopping Agents Read Your Feed: What Agentic Commerce Changes
Agentic commerce adds a new reader to your product data: shopping agents in ChatGPT, Google AI Mode, Gemini and Perplexity that parse feeds instead of browsing pages. They need facts rather than marketing copy: availability, variants, shipping, return conditions and clean attributes, kept fresh and consistent with your product page and checkout. For most teams, that is not a new system but one more output profile at the point where product data already leaves the platform.
From channels to agents: who reads your product data now
Product data distribution used to serve marketplaces, price comparison sites and retailers, with a human at the end who clicked, compared and bought.
Shopping agents change that last step. An agent ignores your hero image. It reads a structured record, compares it with others and answers a request such as "waterproof hiking boots in size 10, delivered by Friday, free returns". If a field is missing, the product drops out of the answer. If a field is wrong, the agent repeats the error with confidence.
We have covered the storefront side already: structured product feeds and agent endpoints inside the frontend, where agents query your site. This post looks at the other direction: the data you send out to agent platforms and AI shopping surfaces that may recommend you without ever visiting your storefront.
What the platforms have published so far
The standards are moving fast, so separate what is documented from what is expected.
- OpenAI and Stripe. The Agentic Commerce Protocol was developed by OpenAI and Stripe and published under the Apache 2.0 license. Its product feed spec includes fields such as availability and availability_date, variant attributes like color, size and offer_id, fulfillment fields like shipping and delivery_estimate, and seller policy links that become required when checkout is enabled. The spec states that updates are accepted every 15 minutes. In March 2026, OpenAI said Instant Checkout is moving to apps, with purchases inside connected merchant services, while ChatGPT focuses on discovery. The feed stays relevant, the checkout path changed.
- Google. Google announced the Universal Commerce Protocol in January 2026, developed with Shopify, Etsy, Wayfair, Target and Walmart, with Merchant Center as the entry point. Google added conversational attributes such as question_and_answer, related_product, variant_option and popularity_rank to help products surface in AI Mode, plus a native_commerce attribute that opts eligible listings into a buy flow in Gemini and AI Mode.
- Perplexity. Perplexity opened its shopping features, with checkout powered by PayPal, to free users in the US in late 2025 and runs a merchant program for retailers.
What is solid: agents rely on structured data supplied by merchants and reward completeness and freshness. What is still a bet: which checkout protocol wins and how many specs you will maintain. Plan for several.
The fields agents need that a shopping ads feed rarely carries
A classic shopping feed was built to show an ad: title, image, price, link. An agent needs enough context to decide without opening the page.
- Availability with meaning. Not just "in stock", but a restock date and availability per variant.
- Variants as a group. Size, color and material as separate values, linked to one parent product.
- Shipping and delivery. Cost, regions and an estimated delivery date. "Delivered by Friday" is a filter, not a nice to have.
- Return conditions. Return window and cost as data. Agents compare policies across shops.
- Attributes instead of marketing text. "Waterproof: yes, weight: 540 g" beats "your perfect companion for every adventure".
- Relationships. Accessories, compatible parts and substitutes when an item is out of stock.
- Identifiers and policies. GTIN, brand, seller terms.
Most of this data already exists in your PIM, ERP or order management system. The gap is usually the output that shapes it for each platform.
Freshness and consistency: feed, PDP and checkout must agree
Agents punish contradictions. If the feed says available and checkout says sold out, you lose the order, and unreliable data tends to cost visibility. The same applies to prices, delivery promises and return terms.
Three rules help:
- One source per fact. Price and stock come from the commercial system, attributes from product data. Feed, product page and checkout read the same values.
- Update on change, not on a nightly batch. When price or availability changes, the export should run with it.
- Validate before you publish. Check mandatory fields, allowed values, GTINs and price formats per channel, and stop an export when the item count suddenly drops.
As our post on a unified data layer for shopping agents argues, agent errors usually come from seams between systems, not from the model.
One more output profile, not one more system
Laioutr, the Frontend Management Platform (FMP), already connects your backends through Orchestr so the storefront can render product, price and stock. The same reconciled data can leave the platform.
Distributr is the data distribution add-on for Laioutr. It sends product data from Orchestr and content from the Laioutr CMS to marketplaces, price comparison sites, retailer listings and ERP systems, including Google Merchant Center, Meta and Amazon. The approach: 0 imports, the data is already there, and 1 direction, out and never in. Outputs come as CSV, XML, JSON, XLSX or PDF, delivered via pull URL or pushed via HTTPS, SFTP, S3 or webhook, on a schedule or on every data change. Validation runs before export, and a diff guard blocks the export when the item count deviates beyond your threshold. Channel definitions are configured in code and versioned in Git.
To be clear about the gap: agent platforms are not listed as a ready-made Distributr channel today. Our point is architectural. When a spec settles, supporting it means a new channel definition at the same output point, not another tool with its own copy of your catalog. Concretely: Larry AI drafts the channel definition from the published field list, and a developer on your team reviews and merges it. Distributr is not a PIM and not a system of record.
Structured data on your own pages is a separate layer: AI Search & LLM Ready is included in Laioutr and covers Schema.org markup, llms.txt and Markdown delivery of every page. AI Search & Discovery is a separate add-on for semantic on-site search. Neither replaces a feed.
FAQ
Do I need a separate feed for every AI platform?
Today, mostly yes, because OpenAI and Google publish different specs. Many fields overlap, so one canonical product with one output profile per platform keeps it manageable.
Is Schema.org markup on my product pages enough?
It helps agents that visit your pages, and your SEO and GEO setup should cover it. Platforms that recommend without visiting rely on submitted feeds, so you need both.
Does Distributr support the Agentic Commerce Protocol?
Not as a ready-made channel today. Distributr already outputs JSON, CSV and XML, formats these specs use, but a format alone is not a channel. New channel definitions are added as code.
How fresh does an agent feed need to be?
As fresh as your stock and prices. The ACP feed spec accepts updates every 15 minutes, so export on every data change is the safer default.
Next steps
Start with an audit: take ten products and check whether availability, variants, shipping, returns and attributes match across feed, product page and checkout. To see how the frontend layer keeps them aligned, book a demo with the Laioutr team or explore the Agentic Frontend Management Platform.