AI Shopping Agents Chose the More Expensive Product — Because It Had Structured Data

Marketing copy is invisible to your next customer.

That is the conclusion from a controlled experiment published by O'Reilly in which researchers gave AI shopping agents two products to compare. One was expensive with structured JSON specifications — materials, dimensions, certifications, compatibility data, all in machine-readable fields. The other was cheaper and described through persuasive marketing prose — the kind of benefit-driven copy that converts human shoppers.

The agents consistently chose the more expensive product.

Not because it was better. Because it was legible. The cheaper product, wrapped in polished marketing language, was functionally invisible to the agent's decision process. The agent could not parse its attributes, could not match its specifications against the buyer's constraints, and could not verify its claims against structured fields. So it selected the product it could actually read.

This is not an edge case. It is the new default.

How AI Agents Actually Shop

When a human shops online, they read headlines, scan bullet points, look at photos, and form an impression. Persuasive copy works because humans process language holistically — we infer meaning from context, tone, and framing.

AI shopping agents do not read copy. They parse data.

When a shopping agent receives a query like "find a waterproof laptop backpack under $150 with a padded 15-inch sleeve," it does not browse product pages. It resolves that query into a set of constraints — waterproof: true, price: ≤150, sleeve_size: 15, sleeve_padding: true — and then filters product catalogs for items that match.

Seventy percent of AI shopping searches include constraints like these: specific materials, certifications, compatibility requirements, size limits, or feature specifications. Each constraint becomes a filter that the agent applies against structured product attributes.

If your product data lives in marketing prose — "our beautifully crafted backpack features premium water-resistant materials and a generously padded laptop compartment" — the agent cannot extract the structured attributes it needs. It does not know whether "water-resistant" means IPX4 or just "we sprayed it with scotchguard." It cannot determine whether "generously padded" meets the padding threshold. It cannot parse "laptop compartment" into a specific dimension.

The product fails the constraint query. The agent moves on.

The Numbers Behind the Shift

The scale of this transition is no longer speculative.

ChatGPT now has 900 million weekly active users with access to visual shopping features. Adobe Analytics reported that traffic to U.S. retail sites from generative AI sources jumped 1,200 percent in under a year. During the 2025 holiday season, AI-referred retail traffic grew 693 percent year over year, and those shoppers converted 31 percent higher than visitors from traditional search.

Consumer adoption of agentic shopping — where AI agents handle discovery, comparison, and checkout — is expected to jump from 19 percent to 46 percent by the end of 2026. Gartner projects that AI agents will intermediate over $15 trillion in global B2B spending by 2028. The average number of AI agents deployed per organization nearly tripled from five in early 2025 to thirteen by April 2026.

Meanwhile, 51 percent of B2B buyers now start their research with AI chatbots, up from 29 percent. And when AI summaries appear in search results, click-through rates for traditional results drop to 8 percent — half the rate without AI summaries.

The buyer is changing. The way the buyer reads your brand is changing faster.

Why Marketing Copy Fails the Machine Test

The O'Reilly experiment reveals a structural problem, not a content quality problem. Even excellent marketing copy fails when the buyer is an agent because the two systems process information in fundamentally different ways.

Humans read for meaning. "Our industry-leading solution delivers enterprise-grade reliability" tells a human reader that the product is mature, trustworthy, and built for scale. The reader fills in the gaps with assumptions, experience, and brand associations.

Agents parse for attributes. The same sentence tells an agent nothing. There is no structured field for "industry-leading." There is no machine-readable value for "enterprise-grade." The agent needs uptime_sla: 99.95, compliance: [SOC2, ISO27001], supported_users: unlimited — concrete, queryable, verifiable attributes.

This gap between human-readable and machine-readable information is where most brands are losing the agentic commerce race. Their product pages are optimized for human persuasion. The growing share of their buyers cannot be persuaded — only informed.

The experiment's conclusion is stark: AI agents are parsers, not readers. They resolve queries against machine-readable fields. Price was irrelevant. Data structure determined the outcome.

What Agents See When They Evaluate Your Brand

Beyond individual product attributes, AI agents evaluate brands holistically through data signals that most companies have never considered as competitive factors.

Schema.org markup — Organization, Product, Service, and LocalBusiness schema tells agents what you are, what you sell, where you operate, and how to reach you. Analysis of 73 websites found that those with properly implemented structured data schema were cited in AI responses 3.2 times more often than those without. Sixty-five percent of pages cited by Google AI Mode and 71 percent of pages cited by ChatGPT include structured data.

Cross-platform consistency — Whether your name, logo, description, contact information, and business details match across your website, Google Business Profile, marketplace listings, social profiles, and directory entries. Agents cross-reference sources. Every discrepancy is a trust penalty. Unlike a human who might overlook a slightly different phone number, an agent flags every inconsistency as a data quality issue.

Verification signals — Whether your brand data is internally consistent, whether external sources corroborate your claims, and whether your structured data has been recently updated. An agent with spending authority from its principal will preference brands it can verify. It calculates trust from data consistency, not design quality.

API accessibility — Whether your product, pricing, and inventory data is available through machine-readable feeds and endpoints. An agent cannot buy what it cannot parse. Products without structured feeds are excluded from agentic search results entirely.

The Maersk Lesson: Agents Negotiate Harder Than Humans

The O'Reilly article includes a second finding that should concern every brand: when Maersk deployed AI agents to negotiate freight contracts autonomously, the agents achieved a 96 percent agreement rate with carriers — at rates 22 percent lower than human negotiators achieved on identical shipping lanes.

This is the other side of the structured data equation. Agents do not just buy differently. They negotiate differently. An agent with access to structured market data, historical pricing, and competitive offers applies consistent downward pressure across every negotiation. There is no relationship capital, no fatigue, no anchoring bias. Just data-driven optimization.

For brands, this means the structured data you expose is not just how agents find you — it is how they evaluate your pricing against the competition. If your competitor's data is more structured, the agent has better information to negotiate with. You are not just invisible; you are at a disadvantage even when you are visible.

The Brand Data Layer Most Companies Have Not Built

Agentic commerce requires a layer that most companies have never invested in: a machine-readable brand identity.

This is different from a website. A website is a human-readable presentation of your brand. A machine-readable brand identity is a structured data set that resolves to the same information — logos, colors, typography, contact data, social profiles, business descriptions — but in a format that agents can query, verify, and cross-reference.

Most brands have their identity locked inside visual assets: a logo file on a designer's laptop, brand colors in a Figma file, contact information buried in a website footer, social profiles linked from an icon bar. None of this is machine-readable. None of it is queryable. None of it helps an agent determine whether your brand is legitimate, consistent, and trustworthy.

The brands that are already winning the agentic commerce shift have built this layer. Their product data has 15 or more structured attributes per SKU. Their brand identity resolves to clean JSON. Their schema markup covers Organization, Product, Service, and LocalBusiness. Their cross-platform data is synchronized and verifiable.

The 15-Attribute Threshold

Here is a practical benchmark from the research: export your top 20 percent revenue-generating SKUs and count the structured attributes per product. If you have fewer than 15 machine-readable fields, you are failing constraint queries.

Fifteen is not an arbitrary number. It represents the minimum attribute density at which products reliably match the constraint patterns in typical AI shopping queries. Below that threshold, products lack the specificity to survive elimination filtering. Above it, products have enough structured attributes to match against multi-constraint queries that include materials, certifications, dimensions, compatibility, and care instructions.

The priority fields are what researchers call "elimination attributes" — the specifications that agents use to filter products in or out before comparing them. Certifications (organic, vegan, waterproof, FSC-certified), material composition, precise dimensions, compatibility with other products, and care instructions. These are the fields that marketing copy typically omits or describes vaguely but that agents require as structured, queryable values.

What Comes Next

The window between "this matters" and "this is table stakes" is closing fast.

OpenAI already pivoted from Instant Checkout to a discovery-plus-redirect model — agents research and compare within ChatGPT, then send buyers to the merchant's checkout. Seventy-five percent of retailers report implementing or planning agentic commerce strategies. Mastercard, Visa, Cloudflare, and Google are all shipping competing agent identity and payment protocols.

The infrastructure for agents to find, evaluate, and purchase products is live. The only variable is whether your product data and brand identity are structured enough for those agents to work with.

The O'Reilly experiment made this concrete: when the buyer is an agent, the product with structured data wins. Not the cheapest product. Not the product with the best marketing copy. The product the agent can read.

Your marketing copy is not being read by your next customer. Your structured data is.

Make your brand agent-ready. Extract your brand fingerprint and find out what AI agents see when they look at your brand.