Generative Engine Optimization for eCommerce: How to Make Products Discoverable by AI (2026)

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. If you draft catalog copy with an agent, use Bedrock AgentCore. Full matrix: lifecycle roundup. GEO does not require Classic.

On June 17, 2026, AgentCore Harness reached general availability (What’s New). That is the paved road for a draft loop. It is not a GEO ranking engine.

Generative engine optimization (GEO) for eCommerce is high-quality, machine-readable product information: attributes, structured data, descriptions that match specs, reviews, FAQs grounded in policy, availability, pricing, relationships, and trust signals. It is not a guaranteed ranking technique. Completing the checklist does not buy a citation in ChatGPT, Perplexity, Google AI Overviews, or a shopping agent.

This post is not an anonymized client engagement. We do not invent organic-traffic lifts.

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). Drafting copy is a session cost; it is not a ranking KPI. Model your mix on the AgentCore pricing calculator.

Reproduce this — Copy geo-product-information-checklist.md. Score one category of PDPs, not the homepage slogan. PIM draft/validate/publish stays in catalog-validation-rules.md — post 9.

Opinionated take: spend the GEO budget on the same record search, recs, and agents already need — not on a parallel “AI SEO” product. Trade-off: you will disappoint anyone who wanted a rank tracker for ChatGPT. You will not ship a second, keyword-stuffed SKU that contradicts checkout.

What Adobe’s visibility numbers are

Adobe Analytics via Digital Commerce 360 (19 Aug 2026), 1T+ U.S. retail visits:

• Expanded July 2026 cohort: 39% of homepages not machine-readable (61% LLM-visible). Do not mix this with Adobe’s April figure (~25% not optimized) on a narrower set.
• Category LLM visibility: apparel 76%, electronics 70%, cosmetics 68%, sporting goods 67%, furniture and home 64%, general merchandise 63%, grocery 59%.
• Same article: AI-referral traffic +62% YoY, +1,219% vs October 2024, converting 60% higher, 53% more revenue per visit than non-AI. That is channel measurement, not a GEO rank.

Do not cite Presenc AI or SEO blogs for these figures. Do not treat 76% apparel visibility as a score you will beat this quarter.

Baymard 70.22% cart abandonment is still about trust and totals. GEO copy that misstates shipping or stock feeds that leak.

What to make readable (the actual work)

The source of those fields is the AI-ready catalog contract. GEO is packaging. Shopping-agent readiness still owns your copilot APIs.

{
  "@type": "Product",
  "sku": "TEE-BLU-M",
  "gtin": "00012345678905",
  "name": "Organic-cotton tee, blue, M",
  "offers": {
    "@type": "Offer",
    "price": "48.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock"
  }
}

Context: illustrative JSON-LD only — not a published schema. If availability is older than your SLA, do not emit InStock. Unknown is honest. Stale in-stock is a chargeback.

There is no native Shopify AgentCore connector. Shopify, Magento, and custom storefronts still need SSR facts or a feed. AgentCore does not install as a GEO app.

Current vs emerging

Current: crawlers and assistants retrieve pages and feeds; Adobe can already score homepage machine-readability; AI-referral sessions exist.

Emerging: richer agent-to-merchant protocols, more of the comparison happening off your PLP.

Not guaranteed: that JSON-LD or a FAQ block will rank you in any generative engine.

If you draft with an agent: Harness for a short tool list; publish stays human; validation is code. Skip Agents Classic after 30 July 2026.

What broke

What broke — “GEO sprint” generated FAQPage JSON-LD for every PDP: keyword-stuffed questions, a 15% coupon that did not exist, and “organic” on SKUs without a certified flag. Detection: golden questions quoted the FAQ, not the policy CMS; legal flagged claims. Fix: delete generated FAQ that is not a pointer to versioned policy; claims gate in PIM (post 9); JSON-LD must match the catalog contract. Lesson: schema is not a place to hide merchandising fiction.

A second failure: optimizing only the homepage because Adobe published a homepage visibility number. Size charts do not live there.

What to Do This Week

1. Open geo-product-information-checklist.md. Score one category of PDPs.
2. Confirm JSON-LD (or equivalent) is on SSR HTML, not only after hydration.
3. Fail any SKU whose description contradicts attributes or whose offer does not match checkout.
4. Ground FAQs in versioned shipping/returns — or omit FAQs.
5. Do not buy a GEO rank package. Do not pause Google Shopping.
6. If attributes are the gap, run catalog validation before packaging.
7. Draft-only agents: Harness (GA 17 June 2026), no publishProduct. Help: contact us, Generative AI on AWS, retail / eCommerce.

What This Post Doesn’t Cover

• AI-search channel instrumentation — post 33
• Catalog as agentic-commerce contract — post 35
• PIM extract/validate/HITL publish workflow — post 9
• How to rank #1 in any assistant
• Measured FactualMinds organic or AI-referral lifts — we are not inventing them
• Marketplace feed specs beyond “GTIN and required attrs still apply”

FAQ

When should you NOT run a GEO program?

Skip it when required attributes are missing, variants have no parent IDs, or price/stock only exist after JavaScript. Packaging a hollow catalog produces confident wrong answers. Finish catalog management first.

What could go wrong if you treat GEO as a guaranteed ranking technique?

You will staff against Adobe’s apparel 76% as an OKR and ship stuffed FAQs. The checklist makes you representable. It does not purchase a citation. Instrument referrals in post 33.

What could go wrong if GEO copy invents claims the SKU is not certified for?

Assistants will quote them. That is legal and returns risk. Claims gates belong in PIM. The model drafts; it is not the system of record.

Is homepage LLM visibility enough?

No. Adobe’s expanded July 2026 cohort: 39% of homepages not machine-readable; apparel 76%. Size, stock, and price live on PDPs.

Does JSON-LD replace a catalog API for shopping agents?

No. JSON-LD helps some retrieval. Comparison and transaction still need APIs or feeds with a freshness SLA. Copilot checks.

Should we use AgentCore to generate GEO content?

Draft only, with validation in code and a merchandiser publish button. Harness GA 17 June 2026. Not Agents Classic after 30 July 2026. Generated meta is not a required-attribute substitute.

Need the catalog contract packaged so assistants can represent you without inventing stock and claims? Contact FactualMinds or see Generative AI on AWS and retail / eCommerce.