Product Feed and Catalog Data Quality for AI Agents: Attributes, GTINs, and Structured Specs

AI shopping agents select products by matching structured attributes, GTINs, and policy data against a shopper's stated constraints. They do not read your marketing copy first. If your feed lacks a valid GTIN, a normalized size value, or a machine-readable return policy, agentic commerce systems exclude your product before a human sees it. Fixing catalog data quality is the single highest-leverage preparation step for agentic checkout.

Why Catalog Data Decides Visibility in Agentic Commerce

A human shopper tolerates vague titles and scans photos. An AI shopping agent receives a request such as "waterproof trail running shoes, women's size 8, under $140, delivered by Friday" and translates it into filters. Every filter maps to a field. A missing field is a failed match, not a weak one.

Three data layers determine whether an agent can recommend and buy your product:

• Identity: GTIN, MPN, and brand, which let the agent confirm it is the same product across merchants and compare price.
• Attributes: color, size, material, dimensions, compatibility, and certifications, which let the agent filter and rank.
• Commerce terms: price, availability, shipping speed, and return policy, which let the agent complete agentic checkout without guessing.

Most merchants invest in the first layer partially, the second inconsistently, and the third almost never in structured form. That gap is where competitors with cleaner data win.

GTINs: The Identity Layer

A GTIN (Global Trade Item Number) is the barcode-level identifier, written as GTIN-8, GTIN-12 (UPC), GTIN-13 (EAN), or GTIN-14. Agents use it to deduplicate listings, pull reviews and specs from other sources, and benchmark your price. Without it, your product is an orphan record.

Concrete guidance:

• Source GTINs from the manufacturer or GS1, never generate them. Invented GTINs fail checksum validation or collide with another brand's product, which triggers disapprovals in Google Merchant Center.
• Validate the check digit programmatically before export. A 12-digit UPC with a wrong final digit is the most common silent error.
• Assign GTINs at the variant level, not the parent. A shirt in 5 sizes and 4 colors has 20 GTINs, not one.
• For custom or handmade goods without GTINs, set identifier_exists to false, and compensate with a strong brand plus MPN pairing and richer attributes.
• Bundles need their own GTIN or a clear multipack attribute, or agents misread the unit price.

On Shopify, the barcode field on each variant feeds the GTIN. On WooCommerce, GTIN support depends on your theme or plugin, and many stores store it in a custom field that never reaches the feed. On BigCommerce, the GTIN field exists per variant but is frequently left blank during bulk imports. Audit all three for blanks.

Attributes: The Filtering Layer

Agents rank on attributes that match user constraints. The quality of those attributes depends on completeness, normalization, and granularity.

Normalize Values

"Navy," "Dark Blue," and "Midnight" are three colors to a filter. Maintain a controlled vocabulary: a base color field (navy) plus an optional marketing name (Midnight Harbor). Do the same for sizes by using one standard (US women's 8) and including size_system and size_type attributes where relevant.

Use Units and Numbers, Not Prose

"Fits most laptops" is unusable. "Fits laptops up to 15.6 inches" is filterable. Store dimensions, weight, capacity, wattage, and battery life as numeric values with explicit units, not embedded in descriptions.

Prioritize by Category

Do not try to fill every field. Fill the ones shoppers constrain on:

• Apparel: size, color, gender, age group, material, fit.
• Electronics: compatibility, connector type, storage, voltage, model number.
• Home and furniture: dimensions, material, assembly required, weight capacity.
• Beauty and supplements: ingredients, skin type, count, certifications.

Structured Specs on the Product Page

Feeds are one channel. Agents also crawl your storefront and parse schema.org markup. Your product page must expose the same data as JSON-LD, and the two sources must agree. A feed that says in stock while the page says backordered gets your listing demoted for unreliable data.

At minimum, mark up the following:

• Product with name, image, description, brand, sku, gtin, and mpn.
• Offer with price, priceCurrency, availability, itemCondition, and priceValidUntil.
• AggregateRating and Review where you hold genuine reviews.
• OfferShippingDetails with handling time, transit time, and destination.
• MerchantReturnPolicy with return window in days, fees, and method.
• additionalProperty using PropertyValue for category-specific specs such as thread count or IP rating.

Render specs as an HTML table or definition list, not as an image or a collapsed accordion populated by late JavaScript. Agents that do not execute scripts see nothing. Server-side render the spec block.

Feed Comparison: Where Each Channel Fits

Maintain one source of truth, typically your PIM or the platform's product database, and generate every channel from it. Editing each feed separately guarantees drift.

Preparing for Agentic Checkout

Agentic checkout lets an agent complete a purchase inside a conversation using protocols such as OpenAI and Stripe's Agentic Commerce Protocol or Google's Agent Payments Protocol. The agent needs to retrieve a final price, shipping options, and tax before payment, with no human clicking through your cart.

Operational requirements:

1. Real-time inventory. Sync stock at least every 15 minutes for fast-moving SKUs. Overselling through an agent produces cancellations that damage your standing on those surfaces.
2. Deterministic pricing. The price in the feed must equal the price at checkout. Dynamic discounts that apply only after login break the agent's quote.
3. Machine-readable shipping. Publish rates by zone and service level with delivery estimates in days, not "ships soon."
4. Guest checkout and tokenized payment. Required account creation blocks delegated purchase.
5. Clear returns and support contact. Agents surface policy details to shoppers before purchase.
6. Order status endpoints. Agents report fulfillment back to the shopper, so webhooks and tracking numbers must be reliable.

The trade-off: opening checkout to agents means ceding some control over upsells and cross-sells. Compensate with bundles defined as products, loyalty data passed through approved fields, and strong post-purchase email.

A Data Quality Audit You Can Run This Week

1. Export your full catalog and count variants with a blank or invalid GTIN. Target under 2% for branded goods.
2. Measure attribute fill rate for your top 5 category-critical fields. Target 95% or higher on top-revenue SKUs.
3. Compare 50 random SKUs across feed, JSON-LD, and visible page for price and availability mismatches. Target zero.
4. Check Merchant Center diagnostics for disapprovals and warnings, and resolve the top three by volume.
5. Test a product page with JavaScript disabled to confirm specs, price, and shipping render.
6. Ask ChatGPT, Gemini, and Perplexity to recommend products in your category using your customers' constraints, and record whether your SKUs appear and what data they cite.

Repeat monthly. Data quality decays with every new product import, supplier change, and seasonal promotion.

Agentic Commerce Data FAQ

Do AI shopping agents require GTINs?

Not universally, but GTINs are the strongest identity signal. Products with valid GTINs are matched, compared, and enriched more reliably, and Google Merchant Center requires them for most branded goods.

What is agentic checkout?

Agentic checkout is a purchase completed by an AI agent on a shopper's behalf, inside a chat or assistant interface, using a payment token and your real-time price, shipping, and tax data.

Is schema markup still necessary if I submit a product feed?

Yes. Feeds supply data to specific platforms, while schema markup lets any crawler or agent verify that data on your site. Mismatches between the two reduce trust and can suppress listings.

Which attributes matter most for AI shopping agents?

The ones shoppers use as constraints: size, color, material, dimensions, compatibility, price, availability, delivery time, and return window. Prioritize these per category before filling optional fields.

How often should I refresh my product feed?

Refresh price and availability at least daily, and every 15 minutes or by webhook for high-velocity or limited-stock items. Refresh descriptive attributes whenever the source record changes.