The Maze: Retailers have spent years making product pages persuasive. AI shopping agents need them to be computable. Acom audited 235 product pages and found a median readiness score of 66 out of 100; even the best page stopped at 88. More strikingly, none exposed size, specifications, or materials as data an agent could filter. The information often existed in prose. That is good enough for a patient reader. It is weak infrastructure for a machine comparing waterproof size-11 shoes under $150.
The readiness gap is broad, not marginal. The Acom audit puts the median page at 66 and the strongest at 88. A 22-point gap separates an ordinary page from the leader, yet another 12 points separate the leader from a fully ready score. Nobody cleared the bar. That matters because AI discovery does not reward the prettiest page in isolation. It rewards the merchant whose catalog can answer a constrained request without guessing. A retailer can rank above peers and still be structurally incomplete.
Zero of 235 turns a content problem into a distribution problem. Every audited page failed to publish size, specifications, or materials in a form the agent could filter. The distinction is mechanical. “Waterproof” buried in a description is a sentence to interpret; `material`, `size`, and other product attributes are fields to test. The first creates ambiguity. The second lets software eliminate mismatches before it recommends anything. When two stores sell comparable products, cleaner data becomes shelf space: the agent can safely shortlist what it can verify.
Variants need identities, not just dropdowns. Google’s product-variant guidance uses `ProductGroup` and `Product` markup to connect sizes, colors, materials, prices, availability, and unique identifiers. It recommends a directly selectable URL for each variant and `Product` data in the initial HTML. That turns “green coat, size small” from a human interface state into a crawlable commercial object. A slick selector without a machine-readable variant model can look complete to a shopper while remaining opaque to the system building the shortlist.
Filtering is the bridge between a prompt and a product. Google Cloud’s recommendation system evaluates filters against category, color, availability, and custom product attributes. A primary item can qualify when one of its variants matches the constraint. OpenAI’s shopping research similarly asks about price, size, features, and trade-offs, then uses merchant product data and public retail information to compare options. The interface may feel conversational. Underneath, the useful work is still catalog logic: identify the item, test the attributes, confirm stock, and explain why it fits.
Why it matters: The next ecommerce shelf is not only a page rank or marketplace placement. It is the set of products an agent can confidently query, compare, and buy. Checkout rails cannot rescue an uncertain match. Give every variant a stable identifier and URL. Publish price, stock, size, color, material, shipping, and returns as structured fields. Keep the prose—it still sells the story—but stop asking prose to do the database’s job. In agentic commerce, invisible attributes create invisible products.
