AI Shopping Agents Picked the Pricier Jacket. Its Product Data Was Cleaner

"Our copy is great. Why does the assistant keep recommending a competitor?" It is one of the most common questions we hear from ecommerce leaders, and a small experiment published by O'Reilly in April 2026 gives the clearest answer so far. AI shopping agents chose the more expensive product every time, because its product information was structured and the cheaper rival's was not.

This post walks through what the experiment found, why it matters for the way shoppers now ask assistants for products, which fields decide the outcome in common categories, and where a lean team should start.

What the O'Reilly experiment found

In Engineering Storefronts for Agentic Commerce, published on O'Reilly Radar on April 6, 2026, the author built a shopping agent and gave it one instruction: find the cheapest waterproof hiking jacket suitable for the Scottish Highlands. Two merchants were available. Merchant A sold a $90 jacket with polished copy ("Ultra-breathable all-weather shell. Conquers stormy seas!") and no structured specs. Merchant B sold a $95 jacket with no copy at all, only a structured field stating 20,000 mm of water resistance.

The author ran the test ten times. The agent bought the $95 jacket from Merchant B every time and never considered the cheaper one. The agent had translated "waterproof" into a minimum water resistance figure, and Merchant A's listing gave it no number to check. "Conquers stormy seas" is not a specification, so the product failed the filter before price was ever compared.

The article explains this with what it calls the Sandwich Architecture. An LLM at the top turns the shopper's vague request into a structured query with explicit values. A deterministic middle layer checks each product against those values, literally and without interpretation. An LLM at the bottom chooses among the products that survive. Marketing copy never reaches the middle layer. Your product facts do, if they are stated as facts.

Why constraints are the whole game

This would be a curiosity if shoppers asked assistants for "a jacket." They do not. At Shoptalk 2026, OpenAI's product partnerships lead for search and commerce said that more than half of searches on ChatGPT are discovery-based, and 70% of those include constraints, as reported in Stripe's recap of the event. Constraints are phrases like "waterproof," "under $100," "fits a 13-inch laptop," or "safe for sensitive skin."

Every constraint is a filter the assistant applies against the product information it can read. If the relevant field exists on your page or in your feed, the product can pass. If the fact is missing, or buried in a paragraph of adjectives, the product is dropped from the comparison regardless of how good it is. The O'Reilly result is simply this filter observed under controlled conditions.

Google is moving the same way. In April 2026 Google rolled out a major shopping upgrade to AI Mode and the Gemini app, combining its Gemini models with a Shopping Graph of about 50 billion product listings, as Business Today reported. That graph is built from Merchant Center feeds and product pages, so the attributes you publish there are the attributes AI Mode can match.

The fields that resolve constraints

The specific fields vary by category, but the pattern is the same: a measurable fact that a shopper's question can be checked against. Some paired examples of what wins and what loses:

• "Waterproof hiking boot." Wins: a stated water resistance rating or waterproof membrane. Loses: "ready for any adventure."
• "Vegan moisturizer under $40." Wins: a vegan attribute set to yes and a current price. Loses: "ethically inspired formula."
• "BPA-free water bottle, 32 ounces." Wins: material and capacity fields. Loses: "hydration you can trust."
• "Frame that fits a Samsung Frame TV." Wins: a compatibility field listing the models. Loses: "looks great in any living room."
• "Crib mattress for a standard US crib." Wins: exact dimensions and a certification field. Loses: "designed with your baby in mind."

Notice what these have in common. Material, dimensions, weight, capacity, certifications, compatibility, care instructions and price are elimination fields. Assistants use them to rule products out before they ever rule products in. They are also the fields most often left blank in feeds even when the information exists in a spec sheet or a supplier document.

Where to start

You do not need to restructure the whole catalog to benefit. A sensible sequence for a lean team:

1. Start with the products that carry revenue. Take the top sellers and the products you plan to promote this season. For most catalogs that is a few hundred items. Everything below is applied to that list first.
2. Read each product the way a shopper's question would. List the constraints a buyer in that category actually states, then check whether each one is answered as a plain fact on the page and in the feed. Fill the gaps from your own sources: specifications, supplier data, support tickets and reviews.
3. Fix the feed, not only the product page. Many brands have decent product pages and a stripped-down Merchant Center feed. Assistants that draw on Google's shopping data read the feed. Our product feeds service works on exactly those fields.
4. Track how you appear, not only whether you appear. A product card with image, price and link is a different outcome from a text mention. Run a fixed set of shopper questions before and after the fixes and record which one you get. A free AI readiness report shows how a single product page reads to the assistants right away.

Structured data markup belongs in that work too. Schema.org product markup and a complete feed state the same facts in a machine-readable form, so the assistant's middle layer has something to check instead of guessing from prose. The AI visibility service ties the shopper questions, the product content fixes and the tracking together.

Frequently asked questions

Why do AI shopping assistants ignore marketing copy?

They do not ignore it so much as fail to use it. When an assistant needs to verify that a jacket is waterproof, it looks for a value it can compare against the shopper's constraint. Persuasive copy rarely contains one. A stated rating does. The copy still matters once a product is in the comparison and the shopper reads the page, but it does not get the product into the comparison.

Does this apply to Google AI Mode as well as ChatGPT?

Yes. Google's AI Mode shopping draws on the Shopping Graph, which is built from Merchant Center feeds and product pages, and the April 2026 upgrade described above leans on the attributes in that graph. The same product information work that improves how you appear in ChatGPT improves how you appear in AI Mode, with the feed carrying more of the weight on Google.

Should we fix product data before spending on AI visibility?

Treat them as one piece of work. The shopper questions you would use to measure AI visibility are the fastest way to find which product facts are missing, and the fixes are what move the measurement. Starting with the top revenue products keeps the scope small enough to finish.

This is the work Paz does for lean ecommerce teams. A Paz expert reads your top products against the questions shoppers actually ask the assistants, finds the gaps in your product pages and feed, and proposes the fixes with the evidence behind each one. You approve, and the platform publishes the changes and tracks how your products appear before and after. Talk to us.