How to Make Your Product Feed and Inventory Data Readable by AI Agents

November 18, 2026 - Reading time: 4 minutes

AI shopping assistants now answer questions like "waterproof hiking boots under $150, in stock near me." They choose products from structured data. If your catalog is messy or out of date, the agent has little to work with.

Why this is happening now

Three efforts are pushing catalog data toward a common shape:

  • Agentic Commerce Protocol (ACP) from OpenAI and Stripe, released in September 2025. OpenAI's product feed spec describes the file you push to ChatGPT.
  • Universal Commerce Protocol (UCP) from Google, announced in January 2026 with retailers including Shopify, Etsy, Wayfair, Target and Walmart. It builds on Google Merchant Center data.
  • schema.org Product markup on your own pages, which most crawlers and agents already read.

The details differ by protocol, and some checkout features are limited to approved partners. The data work underneath is the same for all of them.

Start with the required fields

OpenAI's spec asks for nine fields on every item or variant: item_id, title, description, url, brand, seller_name, image_url, availability and price. Optional fields cover variants, attributes, shipping and returns. Treat the nine as your minimum for every feed, whichever platform you target.

Example feed row (JSON for readability):

{
  "item_id": "BOOT-TRAIL-42-BLK",
  "title": "TrailMaster Waterproof Hiking Boot, Black, EU 42",
  "description": "Leather hiking boot with a waterproof membrane and Vibram sole.",
  "url": "https://example.com/products/trailmaster-black-42",
  "brand": "TrailMaster",
  "seller_name": "Example Outfitters",
  "image_url": "https://example.com/img/trailmaster-black.jpg",
  "availability": "in_stock",
  "price": "129.00 USD"
}

Check the current spec for exact value formats before you publish. Accepted values for fields such as availability and price can change.

Make every variant its own item

Agents compare concrete items. A boot sold in five sizes and three colors is fifteen items, each with its own ID, price and stock count. Put the color and size in the title and in structured attributes. A single listing with a "choose a size" dropdown tells an agent nothing about what is in stock.

Write clean values

Format errors are easy to make:

  • Price is a plain number with a decimal point and a separate ISO 4217 currency code (490.00 and EUR). Strings like "490,00 € all included" break parsers.
  • Availability uses a canonical value. In schema.org that is a URI such as https://schema.org/InStock.
  • Identifiers: include a GTIN where one exists, plus SKU, MPN and brand. Identifiers let an agent match your item against others.
  • Titles and descriptions are plain text. Skip promotional phrases and all-caps. State the material, size, compatibility and other specs the shopper would filter on.

Mirror the feed in schema.org JSON-LD

Agents and crawlers also read your product pages. Your page markup has to agree with your feed, because mismatches get flagged.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "TrailMaster Waterproof Hiking Boot, Black, EU 42",
  "sku": "BOOT-TRAIL-42-BLK",
  "gtin13": "0123456789012",
  "brand": { "@type": "Brand", "name": "TrailMaster" },
  "image": "https://example.com/img/trailmaster-black.jpg",
  "offers": {
    "@type": "Offer",
    "price": "129.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock",
    "url": "https://example.com/products/trailmaster-black-42"
  }
}
</script>

Generate this from the same source as the feed. Two hand-maintained copies will drift apart.

Keep inventory fresh

Stale stock data sends shoppers to products they can't buy. The platform receives the feed you push, so you control how current it is.

  • Update availability and price on a short schedule, or on change events. Don't rely on a nightly export.
  • Fire an update when stock hits zero, not when someone next opens the page.
  • Make sure the page, the JSON-LD and the feed show the same stock state.
  • Handle failed uploads with retries and an alert, so a broken job doesn't go unnoticed for days.

Validate before you publish

  1. Run your pages through Google's Rich Results Test and check Merchant Center diagnostics for warnings.
  2. Write a small script that checks each feed row for the required fields, numeric prices and valid currency codes.
  3. Spot-check by comparing ten random items across the feed, the page and the JSON-LD.
  4. Track rejected-item counts after each feed run.

Where to go next

Being in a feed makes you eligible for discovery. Checkout inside an assistant is a separate integration, and for some platforms it is limited to approved partners. Start with clean data, then apply to the checkout programs that fit your stack.

Specs in this area are changing quickly. Check the official documentation from OpenAI and Google for current field names and eligibility before you implement.

About

Hi there! My name is Kevin Earl Gardner. I am a Development Manager and Web Developer and (very) Amateur Content Creator based out of Birmingham, Alabama. I like to blog about coding and streaming.