The AI-Ready Product Catalog: The Foundation of Agentic Commerce (2026)

A catalog is not complete because the product-system dashboard is green. It is ready when search, recommendations, and shopping agents can read the same sellable record — parent and child, a unique barcode, a checkout-true price, fresh stock with a timestamp — without each system inventing its own product.

Adobe Analytics (July 2026, 1T+ visits): 39% of an expanded U.S. retail homepage cohort were not machine-readable; apparel 76% LLM visibility. That is page readability. Agentic commerce fails one layer down: SKU graph quality.

The job. Publish one contract every consumer reads — search, recs, your copilot, external agents, GEO.

This week. Copy ai-ready-catalog-contract.md. Pick one category. Count SKUs missing required attrs, parent IDs, GTIN uniqueness, or fresh asOf.

A person still signs. Draft and validate in code; a merchandiser publishes. Auto-publish is how wrong care instructions go live.

Skip it when completeness % is high but variants have no parent IDs, inventory is a different SKU than the storefront sells, or the only catalog is a rendered PDP.

This is part 35 of eCommerce AI Agents. The PIM workflow is post 9. GEO packaging is post 34.

Our take: do not let search, recs, GEO, and agents each invent a product model. One contract, many consumers. Merchandising cannot keep a “web-only” title that fights the spec. You stop recommending SKUs you cannot fulfill.

Copy the contract — ai-ready-catalog-contract.md. Production gates: catalog-validation-rules.md. Series index: ecommerce-ai-agents-series/.

FactualMinds is an AWS Select Tier Services Partner. We do not invent completeness or conversion KPIs.

Baymard 70.22% average cart abandonment is still about confidence and totals. A catalog that cannot tell the truth about stock and price hands that problem to every agent that reads it.

From supplier sheet to sellable record

flowchart LR
  raw[Raw]
  clean[Clean]
  norm[Normalize]
  enrich[Enrich]
  val[Validate]
  struct[Structure]
  ready[AI-ready catalog]
  cons[Search / recs / agents / GEO]
  raw --> clean --> norm --> enrich --> val --> struct --> ready --> cons

FactualMinds catalog automation (part 9) is how large catalogs get through validate. This post assumes that loop exists or is being built. If it does not, do not skip to JSON-LD.

There is no native Shopify AgentCore connector. Shopify Admin GraphQL, Akeneo, Salsify, or a custom PIM sit behind your OpenAPI. AgentCore does not ship an AI-ready catalog app.

Defects that break agents (not just SEO)

Current: on-site search and recs already punish these defects. Emerging: external shopping agents and AI search as a channel retrieve the same holes at larger blast radius. GEO (post 34) is packaging of this record — not a second product, and not a ranking guarantee.

Your copilot needs this contract on tools. External buyer agents need it on feeds/APIs/JSON-LD. Same fields.

Context: illustrative JSON — align names to your PIM:

{
  "sku": "SHOE-TR-W-10",
  "gtin": "00012345678912",
  "parentId": "SHOE-TR",
  "attributes": { "size": "10", "width": "wide", "waterproof": true },
  "offer": {
    "price": { "amount": 140.0, "currency": "USD", "taxIncluded": false },
    "availability": { "status": "in_stock", "quantity": 22, "asOf": "2026-08-27T09:00:00Z" }
  }
}

If asOf is older than the SLA, consumers must see unknown, not in_stock.

What broke

What broke — Recs, on-site search, and a prototype shopping prompt each used a different export: recs on parent IDs, search on URL slugs, the agent on marketplace GTIN. A wide-size trail shoe was “in stock” on the parent and 0 on the child. Detection: golden compare questions recommended a size the WMS could not pick. Fix: one contract object; all consumers read parentId + child availability; recs cannot emit a parent as a buyable SKU. Lesson: AI-ready is a shared schema, not three CSVs.

A second failure: calling the catalog AI-ready because enrichment filled descriptions while care, GTIN, and variant links still failed post 9 gates.

What to use instead

• Three shadow exports. List consumers (search, recs, copilot, feed, JSON-LD). If any uses a shadow export, that is the ticket.
• Marketplace CSV without validation. Duplicate GTINs and stale stock get retrieved as facts.
• You need the PIM loop. Go to post 9 — do not re-implement extract here.

If you only do one thing

Pick one category. Count SKUs that would fail the contract JSON. If you cannot name parent IDs, GTIN uniqueness, and asOf on offer fields, you are not AI-ready. You are ready to fix PIM gates.

For your technical lead

On June 17, 2026, AgentCore Harness reached general availability (What’s New). Use Harness when you draft against a short tool list. The foundation of agentic commerce is still whether the published record is a contract.

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. Catalog draft agents should use Bedrock AgentCore. Full matrix: lifecycle roundup. The AI-ready catalog is a data product, not a Classic action group.

First-party signals we reuse (not eCommerce outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). PIM and inventory p95 will dominate agent latency. Model remaining sessions on the AgentCore pricing calculator.

Gateway ~95 ms is not your inventory service. Batch validation should not hold a chat session per SKU. Draft agents: Harness (GA June 17, 2026), createDraft only. Not Classic after July 30, 2026.

What to do this week

1. Copy ai-ready-catalog-contract.md. Map each field to PIM.
2. Pick one category. Count SKUs missing required attrs, parent IDs, GTIN uniqueness, or fresh asOf. That count is yours.
3. List consumers (search, recs, copilot, feed, JSON-LD). If any uses a shadow export, that is the ticket.
4. If the fail mix is extraction and publish, go to post 9 — do not re-implement extract here.
5. Packaging only: GEO checklist. Copilot APIs: post 10.
6. Implementation help: contact us, Generative AI on AWS, retail / eCommerce.

What this post doesn’t cover

• Extract / OCR / HITL publish — post 9
• GEO JSON-LD and FAQ packaging — post 34
• Recommendation ranking algorithms (related: ai-product-recommendation-agent-ecommerce-2026 when it ships)
• Measured FactualMinds catalog-completeness or conversion lifts — we are not inventing them
• Marketplace-specific feed specs
• Auto-merge of near-duplicates

It does not invent completeness or conversion KPIs. 39%, 76%, 70.22%, ~180→95 ms, and ~$791/mo are the published figures we reuse.

Primary next step: contact us or see Generative AI on AWS and retail / eCommerce.

FAQ

When should you NOT call the catalog AI-ready?

Skip the label when completeness % is high but variants have no parent IDs, inventory is a different SKU than the storefront sells, enrichment ran before validation passed, or the only catalog is a rendered PDP. Pretty descriptions are not a contract. Search, recs, and agents will fill gaps with language.

What could go wrong if you feed agents a marketplace CSV and skip the contract?

Duplicate GTINs, orphaned colorways, prices that exclude the checkout fee, and six-hour in-stock flags get retrieved as facts. External agents and your copilot both fail closed or, worse, fail open. Baymard still measures 70.22% cart abandonment — stale offer facts amplify it. Publish a record with asOf, or return unknown. Baymard 70.22%.

How is this different from AI product catalog management (post 9)?

Post 9 is how you produce quality inside PIM: extract, deterministic validate, merchandiser publish. This post is the output contract search, recommendations, GEO, and shopping agents consume. Do not duplicate the extract/publish workflow here. If the contract fails, go back to post 9’s gates — do not prompt the storefront harder.

What could go wrong if recommendations use a different product graph than the agent?

The rec widget shows a parent; the agent compares a child; inventory sits on a third identifier. Shoppers get a size that is not in stock. One contract, many consumers. Recs must read the same parent/child and availability fields. Do not let merchandising maintain a shadow catalog for AI.

Does Adobe homepage visibility mean our catalog is AI-ready?

No. Adobe’s expanded July 2026 cohort found 39% of homepages not machine-readable; apparel 76% LLM visibility. Homepages are not SKU graphs. AI-ready means required attributes, unique GTINs, variant relationships, and checkout-true price/stock on the product record. GEO packaging comes after that contract. Adobe / Digital Commerce 360.

Should a catalog agent publish directly so the AI-ready feed stays fresh?

No. Freshness is a pipeline SLA, not an unbounded publish tool. Draft and validate in code; a merchandiser publishes. Harness (GA June 17, 2026) fits a short draft loop. Skip Agents Classic after July 30, 2026. Auto-publish is how wrong care instructions go live.

Need one product contract that search, recs, and agents can share? Contact FactualMinds or see Generative AI on AWS and retail / eCommerce.