ChatGPT Shopping, Perplexity Shopping, and Google AI Mode have a fundamentally different evaluation model than keyword search. They are not matching your product titles to search queries. They are trying to answer a question. And most product catalogs, even well-maintained ones, were built to answer the wrong kind of question.
The Evaluation Model Has Changed
For about fifteen years, the dominant product discovery model was keyword matching. A shopper typed a query. An algorithm matched terms in the query to terms in product titles, descriptions, and category labels. Products with higher relevance scores and appropriate bids appeared near the top. Catalog optimization for this model meant: get the right keywords in the right places, maintain a clean feed, and keep your bids competitive.
AI shopping agents work through a different process. A shopper asks a natural language question: "What is the best waterproof jacket for hiking in the Pacific Northwest in fall?" The agent does not match that query to product titles. It interprets the intent, identifies the product attributes that are relevant (waterproofing standard, packability, fit for layering, hood adjustability for rain use), and then evaluates available product listings to see which ones can confirm a match on those attributes.
If a product listing cannot confirm the match because the relevant attributes are absent, imprecise, or buried in marketing copy, the agent does not recommend it. Not because the product is bad, but because the agent cannot verify that it is good enough.
What Your Catalog Is Missing
When we run the audit on a new retailer's catalog, the most common gaps look like this. A jacket listing with the title "Women's Waterproof Rain Jacket Hooded Outdoor Windbreaker," a description that leads with "Stay dry in style," and no structured attribute fields. The product probably does have meaningful specifications: the waterproofing might be 15,000mm hydrostatic head rated with 15,000g/m2 breathability. The seams might be fully taped. The hood might have three-point adjustment. All of that information exists somewhere, possibly in the manufacturer's product data, but it has not made it into the listing in a form an agent can use.
An AI agent evaluating this jacket for a query about waterproof hiking jackets has nothing to verify with. The title says "waterproof" but does not say to what standard. The description says "stay dry" which is a benefit claim, not a specification. There is no schema additionalProperty data providing the waterproof rating, breathability, or seam construction. The agent's response to this information state is to either omit the product or hedge heavily: "This jacket claims to be waterproof but I could not find specific performance ratings."
Across our early pilot catalogs, an average of about 70 percent of listings lacked the structured attribute coverage AI agents need for confident recommendations. This is not a niche problem. It is the standard state of most retail catalogs today.
The Three Failure Modes
Attribute gaps. Category-relevant specifications that exist in the real world but are absent from the listing in machine-parseable form. The weight is not listed, or it is listed only in the description paragraph in inconsistent units. The compatibility is mentioned in passing but not as a structured field. The temperature rating is in the manufacturer's PDF but not on the product page.
Description written for skimmers, not agents. Human shoppers skim product descriptions looking for emotional confirmation: does this look right, does it feel like the right choice? Marketing copy designed for skimmers front-loads emotional language and benefit claims, which are exactly what AI agents cannot use. An agent evaluating a sleeping bag needs to know: what is the temperature rating, how is it measured, what is the fill power, what is the packed size? A description that opens with "stay warm on your next adventure" answers none of those questions.
Schema markup that meets the Google rich results bar but nothing more. Google's rich results requirements are minimal: a Product schema object with name and at least one of a short list of optional fields. Many catalogs meet this bar and stop. But the schema fields that matter for AI agent comprehension go beyond what Google rich results require: additionalProperty with category-specific specifications, variant-level availability, shipping details in machine-readable format. A product with a valid Google rich results schema and no additionalProperty fields is invisible to agents that rely on structured data for attribute comparison.
What to Do About It
The path out of this is not mysterious, but it requires a systematic approach. You need to know which of your listings are failing which criteria, prioritize by traffic and margin, and fix the highest-impact gaps first.
For the jacket example above, the fix involves: adding a structured attribute section with waterproofing rating, breathability rating, seam construction, weight, and hood configuration; rewriting the description's first two sentences to anchor the use case specifically ("designed for active hikers in sustained Pacific Northwest rain who need breathability during high-output climbs"); and extending the schema to include additionalProperty fields that expose these specifications in machine-readable form.
That work takes time, and doing it manually at catalog scale is not realistic. The problem is structural enough that most catalogs have the same gaps across hundreds or thousands of SKUs in the same category. A systematic audit identifies the pattern; a systematic rewriting engine fixes it at scale.
The Channel Is Already Here
We want to be clear about timing. This is not about preparing for an AI shopping future that may or may not arrive. ChatGPT Shopping integrations launched in 2024. Perplexity Shopping is actively indexing product catalogs now. Google AI Mode is surfacing product recommendations in a significant share of shopping queries today. The retailers who are waiting to see whether AI shopping agents become important are watching from the wrong side of the window.
The good news: the gap between where most catalogs are today and where they need to be for AI agent visibility is largely a data problem, not a product problem. Your products are probably good. The agents just cannot find them yet.
Is your catalog ready for AI shopping agents?
ReFiBuy audits your product feed and rewrites the listings AI agents cannot evaluate. See how it works.
See plans
