Agentic Commerce Product Data: Why Structure Wins

AI shopping agents consistently chose the more expensive product - because it had clean structured data. That's the finding from a controlled experiment published by O'Reilly last week. Even when the cheaper alternative led with polished marketing copy, agents ignored it. With ChatGPT visual shopping now live for 900 million weekly users as of March 25, 2026, your products are being evaluated on the same terms, every day, at scale.

Agentic commerce product data optimization is the practice of enriching product catalogs with machine-readable structured attributes - materials, dimensions, certifications, compatibility - that AI shopping agents use to match, compare, and recommend products. It replaces reliance on marketing language with parseable data fields. This article breaks down the O'Reilly experiment and what your ecommerce team should do differently.

TL;DR: O'Reilly's structured data experiment confirms that AI agents make decisions based on machine-readable product attributes, not marketing language. With ChatGPT visual shopping live for 900M users and Google's AI Mode receiving its biggest-ever shopping upgrade this week, catalog enrichment is the most concrete action ecommerce teams can take right now. Teams that structure their data first will capture AI recommendation share; those that don't will watch competitors take it.

What the O'Reilly Experiment Actually Found

The O'Reilly Radar piece "Engineering Storefronts for Agentic Commerce" (April 6, 2026) is worth reading in full if you work in ecommerce. The short version: researchers gave AI shopping agents two products to compare. One cost more and had structured JSON specs. The other was cheaper and described through persuasive marketing prose. The agents picked the expensive one. Every time.

The reason isn't complicated. AI agents are parsers, not readers. They resolve queries against machine-readable fields: weight, dimensions, materials, compatibility, certifications, ingredients. "Premium quality craftsmanship" passes right through them. "Material: 316L surgical stainless steel" gives them something to work with. When those structured fields exist, the agent can reason about fit. When they don't, the product is functionally invisible for constraint-based queries.

The article also describes what they call the "Sandwich Architecture" for AI storefronts: an LLM translator on top, deterministic validation in the middle, LLM decision at the bottom. Your product data is read at the deterministic layer. Marketing copy never reaches it.

Key Stat: In O'Reilly's experiment, AI agents consistently chose the more expensive product with clean structured data over the cheaper product with only marketing copy. Price was not the deciding factor. Data quality was.

Why the Timing Makes This Urgent

The O'Reilly finding would matter in the abstract. It matters more because two of the largest shopping surfaces in the world just expanded simultaneously.

OpenAI rolled out visual shopping in ChatGPT to all users, including the free tier, on March 25, 2026. That is 900 million weekly users who can now browse products with images, prices, and specs directly in their chat window. At Shoptalk 2026, OpenAI's product partnerships lead said more than half of searches on ChatGPT are discovery-based, and 70% of those include constraints. Constraints like "waterproof," "under $100," and "compatible with iPhone 16" are resolved against your structured product data. If a field doesn't exist in your catalog, the constraint fails, and the agent moves on.

Then on April 7, Google announced what it called the biggest-ever upgrade to shopping in AI Mode, backed by a Shopping Graph containing 50 billion products updated at two billion per hour. Google's system ranks products by matching structured attributes to query intent, not by keyword density.

Two of the dominant shopping platforms in the world are simultaneously running the same O'Reilly experiment, on your catalog, in real time.

"AI agents are a new storefront," said the Novi CEO at Shoptalk Women in AI. "If your brand is not discoverable there, it risks becoming invisible."

What Should Your Ecommerce Team Do Differently?

The shift from "we sell online" to "our products are optimized for agentic commerce" comes down to structured attributes. Most product pages carry 5 to 8 machine-readable fields. AI agents need 30 or more to handle the constraint-based queries that now account for 70% of AI shopping traffic.

Here's where to start:

Audit your top revenue SKUs first. Export the 20% of products that drive the most sales. Count the structured attributes per SKU. Fewer than 15 machine-readable fields means you're likely failing constraint queries in your category. This is the fastest way to size the gap before you start filling it.

Prioritize elimination fields. AI agents use attributes to rule products out before they rule them in. Material, dimensions, certifications (organic, vegan, waterproof, UL-listed), compatibility, and care instructions are the fields that resolve constraint queries. If those fields are missing or buried in prose descriptions, your product fails the filter before it ever reaches comparison.

Enrich your feed, not just your product pages. Manual enrichment doesn't scale past a few hundred SKUs - AI-powered catalog enrichment can generate structured attributes from existing product data automatically. Many brands have decent PDPs but stripped-down catalog feeds. AI agents read your feed. The enrichment needs to happen at the data layer, not the front end.

Track how you appear, not just whether you appear. There is a real difference between a product card (image, price, link) and a text mention. Adobe Analytics data shows that AI-referred visitors convert 38% better than visitors from traditional search. But that conversion requires a product card. If your products show up as text mentions while competitors appear as cards, you are generating awareness that your competitor collects.

Want to see your current score before you start? Run your free AI Readiness Report at paz.ai/ai-readiness. It takes 30 seconds and shows exactly which attributes are missing from your current catalog.

Key Stat: AI-referred visitors convert 38% better than visitors from traditional search, according to Adobe Analytics. Getting to a product card, not just a text mention, is what drives that conversion lift.

Who Is Already Getting This Right

The competitive gap is visible in the data. ChatGPT now accounts for 20% of Walmart's referral traffic and 15% of Target's, according to Accenture research published in April 2026. Neither figure is accidental: both retailers have invested heavily in structured product data, feed quality, and AI channel distribution.

Sephora has added a second layer. Their ChatGPT app integration uses loyalty data to deliver personalized product recommendations in-chat, meaning they're not just structured but contextually matched to individual shoppers. That's a level of optimization a thin, unstructured catalog cannot reach.

The table below shows what the O'Reilly experiment looks like in practice for common shopping queries:

According to NRF and Stripe's January 2026 survey, 75% of retailers say they're implementing or planning agentic commerce strategies. The brands that move from "planning" to "executing on data quality" in the next two quarters will set the category baseline. Closing that gap later is harder than opening it now.

"AI agents are a new storefront: if your brand is not discoverable there, it risks becoming invisible."
Novi CEO, Shoptalk Women in AI (April 2026)

Frequently Asked Questions

What is agentic commerce product data optimization?

Agentic commerce product data optimization is the process of enriching a product catalog with machine-readable structured attributes so AI shopping agents can accurately match, compare, and recommend products to consumers. It focuses on fields like materials, dimensions, certifications, and compatibility rather than marketing language. AI agents parse structured data; they cannot reliably interpret prose descriptions for constraint resolution.

Why do AI agents ignore marketing copy?

AI shopping agents use structured fields to evaluate whether a product matches a shopper's query. Marketing copy exists in unstructured text that agents cannot reliably parse for specific attributes. When an agent needs to verify whether a jacket is waterproof, it checks a structured "water resistance" field. If that field is empty or missing, the product fails the query regardless of how well the description is written.

How many structured attributes does a product need for AI shopping?

Most product pages today carry 5 to 8 structured attributes. AI agents need 30 or more to handle the range of constraint-based queries shoppers use, particularly on ChatGPT, where over 70% of discovery searches include specific constraints like material type, size compatibility, or product certifications.

How can I find out how my products appear to AI shopping agents right now?

Paz.ai's free AI Readiness Report scores any product URL on a 0-100 scale across four dimensions: product mapping, structured attributes, attribute context, and product context. It runs in about 30 seconds and shows exactly what AI agents see and what structured fields are missing from your current catalog.

Does structured data optimization help with Google AI Mode as well as ChatGPT?

Yes. Google's AI Mode shopping runs on its Shopping Graph, which contains 50 billion products and processes 2 billion updates per hour, as of Google's April 7, 2026 announcement. Structured product data is equally critical for AI Mode visibility. The same catalog enrichment that improves ChatGPT recommendation share also improves Google AI Mode results.

What's the most common structured data gap in product catalogs?

The most common gap is missing constraint-resolution fields: certifications (organic, vegan, waterproof, BPA-free), compatibility specifications, precise dimensions and weight, and material composition. These are the exact fields AI agents use to narrow results when shoppers include constraints in their queries. They're frequently excluded from product feeds even when the information exists in brand documentation.

Start With a Score

The O'Reilly experiment stripped away everything except the one variable that actually matters to AI agents: structured data quality. Your catalog is facing that same evaluation every day, on platforms where a combined billion-plus users shop.

Get your free AI Readiness Report at paz.ai/ai-readiness. You'll see your score in 30 seconds, the exact attributes that are missing, and where to focus enrichment first. That's the gap between the product the agent recommends and the one it passes over.