Skip to content
UCP
Menu

Pillar · Data

Product catalogs for AI systems

Turning a PIM feed into an agent-retrievable, policy-aware, semantically addressable product surface — the concrete model.

Updated : April 2026 · Primary query : product catalog for ai agents

An agent-ready product catalog is a catalog that an AI agent can retrieve, parse, reason about and act on — without scraping, guessing or inferring. This page defines the target shape, the minimum viable attribute set, and the validation steps that separate a feed that "works" from a catalog that interoperates.

Working definition

Definition

A product catalog is agent-ready when each SKU has: a canonical identifier, an addressable URL, typed attributes covering the category's query surface, explicit policy fields, and a state (price + availability) that is fresh within a well-defined freshness SLA.

Five properties of an agent-ready catalog

  1. Identifiable. Every SKU resolves to a canonical identifier (GTIN, or MPN+Brand, or an internal SKU that is stable).
  2. Addressable. Every SKU has a permanent URL that returns structured data without login or consent walls.
  3. Typed. Attributes have names, units, enumerated values — not just marketing prose.
  4. Policy-aware. Returns, warranty, shipping, age/geo restrictions are expressed as structured data, not only as legal text.
  5. Fresh. Price and availability have a documented freshness window, and feeds match that window.

Minimum viable attribute set

The exact list depends on category. The following is the baseline any agent will expect for a physical consumer good.

FieldExampleWhy agents need it
idSKU-39281Internal reference, deduplication
gtin5012345678900Cross-merchant identity
brandPatagoniaFilter, trust, match
titleWomen's Torrentshell 3L Rain JacketHuman-facing, query match
description(semantic, attribute-rich)Embedding / retrieval
categoryApparel > Outerwear > Rain JacketsTaxonomy placement
price129.00 EURRanking, filtering
price_effective_until2026-04-30T23:59:59ZPromotion honesty
availabilityin_stock / pre_order / out_of_stockReliability
stock_qty14Scarcity signal
shipping{region, cost, handling_time, delivery_time}Fit to shopper constraints
returns{window, cost, method}Pre-sale confidence
warranty2 years manufacturerConsideration
attributes{material, waterproof_rating, weight_g, …}Query filtering
variants[{size, color, gtin, stock_qty, price}]Resolve the right SKU
images[{url, role, alt}]Visual verification
urlhttps://merchant.com/p/torrentshellHandoff / attribution
localeen-GB, fr-FR, …Regional routing
updated_atISO 8601 timestampFreshness

Anti-patterns

Patterns we see regularly that silently degrade agent performance:

  • Marketing-only titles. "Ultimate Rain Protection for Real Adventurers" — unsearchable. Prefer "Women's 3L Rain Jacket — Torrentshell — Patagonia".
  • Attributes inside description prose. An agent has to extract them. Type them.
  • Region-sensitive PDPs without structured locale. The agent sees a French page priced in EUR for a US shopper.
  • Price mismatch between feed and PDP. Agents cross-check; mismatches kill trust.
  • Stock set to "available" when the back-order window is 6 weeks. Use structured availability states.
  • Returns policy as a single freeform paragraph. Structure it.

The three expressions of the same catalog

An agent-ready catalog should be expressible in three parallel forms. Each form serves a different retrieval channel; all three should reconcile.

  1. Structured data on the PDP — JSON-LD using schema.org Product, Offer, AggregateRating, MerchantReturnPolicy, ShippingRateSettings.
  2. Feed — Google Merchant Center XML/CSV (or equivalent), enriched to cover policy fields.
  3. API / MCP tool — programmatic endpoint for agent orchestrators. Read-only endpoints for catalog, offers, inventory, policy.

The three forms must agree on identifiers and prices. Reconciliation is an operational discipline, not a one-time export.

Validation and telemetry

How you know you are agent-ready:

  • Google Rich Results Test — JSON-LD parses as Product and Offer.
  • Merchant Center diagnostics — no high-severity errors.
  • Feed coverage — >95% of SKUs have brand + GTIN (or MPN+Brand).
  • Price parity — feed and PDP match at >99% of sampled SKUs.
  • Stock parity — availability states match checkout behaviour at >98%.
  • Structured returns / shipping — JSON-LD present on >90% of PDPs.
  • Agent crawl volume — measurable traffic from OAI, Anthropic, Perplexity, Amazon, GoogleBot agents; you track via server logs.

This set is the core of our audit methodology.

Shopify, WooCommerce, Salesforce, custom stacks

Platform-specific notes:

  • Shopify — structured data is partial out-of-the-box; metafields are the leverage. Expose rich attributes via Shop / Shopify Shopping APIs as they evolve agent endpoints.
  • WooCommerce — plugins like RankMath / Yoast cover schema basics; returns/shipping structured data typically requires custom snippets.
  • BigCommerce / Adobe Commerce / Salesforce Commerce Cloud — agent-side APIs appearing in product roadmaps; track vendor release notes.
  • Custom — the opportunity is biggest and the discipline required is highest. Invest in schema, feed and API parity from day one.

Where to go next