Sameer Khan
The AI Shelf: Why Your Product Catalog Needs Machine-Readable Attributes
Short answer: The “AI shelf” is the set of products an assistant can parse, evaluate and recommend with confidence. Getting on it requires product attributes that are explicit, structured and consistent across every source a model might read — typically schema.org/Product markup on your PDPs plus a clean merchant feed. Products that force a model to infer their specifications lose to products that state them outright.
The failure mode is rarely that a model dislikes your product. It is that the model cannot confirm your product meets the shopper's constraint, so it recommends one it can confirm instead.
What machine-readable actually means
It does not mean your page is crawlable — almost every page is. It means a machine can extract discrete, typed facts without guessing. “Breathable and lightweight for all-day comfort” is marketing copy: a human understands it, a model cannot verify it. “Weight: 248g. Upper: engineered mesh. Drop: 8mm.” is machine-readable: it can be compared, filtered and cited.
Both belong on the page. Only one of them gets you recommended when a shopper asks for a running shoe under 250 grams.
The structured data layer
Google's product structured data documentation defines the baseline, and it is the same vocabulary generative engines lean on. At minimum each product should express identity and resolvability through name, description, brand, sku and gtin — GTIN in particular lets a model match your listing to the same product elsewhere and cross-check claims. It should express offers with price, currency and availability, because an assistant will not confidently recommend something it cannot confirm is purchasable. Ratings and reviews should appear only where genuinely present and visible on the page. And category-specific facts that no standard field covers — waterproof rating, thread count, wattage, allergen status — belong in additionalProperty.
One caution worth stating plainly: Google's spam policies treat markup that misrepresents page content as a violation. Structured data must describe what is actually visible. Marking up ratings you do not have is not an optimisation, it is a manual-action risk.
The attributes that decide recommendations
Functional specifications. Dimensions, weight, capacity, materials, power, compatibility — in standard units, with the unit stated. These are what constraints get tested against.
Use-case attributes. The single most neglected category. Assistants receive requests as situations, not specifications: “for a beginner runner with flat feet.” If your PDP never states arch support or experience level, you cannot match that request no matter how good the shoe is.
Constraint attributes. The filters shoppers actually apply — price band, size range, colourway, dietary or allergen status, certifications, shipping window. Missing constraint data is silent disqualification.
Differentiators. What makes this variant different from the next one. Assistants explain their picks, and they need a reason to give.
Negative attributes. What the product is not suited to. Counter-intuitive, but stating limits makes every other claim more credible — and prevents recommendations that end in returns.
Consistency is the multiplier
Models read your PDP, your merchant feed, your marketplace listings and third-party reviews as competing testimony about one object. When those disagree — a weight that differs between feed and page, a variant named differently on a marketplace — the contradiction does not average out. It reduces confidence, and confidence is the deciding factor in how AI assistants choose which products to recommend.
Practical rule: one source of truth, syndicated outward. Never maintained in parallel.
An audit you can run this week
Take your ten best-selling SKUs and write down the five questions a shopper most often asks before buying each one. For each question, check whether the answer exists as an explicit attribute — not implied in prose, not shown only inside an image. Validate the resulting markup with Google's Rich Results Test. Then diff your feed values against the live page for those same SKUs; mismatches are your fastest wins. Finally, ask an assistant directly for a product matching your best SKU's use case, and see whether you appear — and what it says about whoever does.
Why this compounds
Attribute work is unusually durable. Assistants, protocols and feed specifications will keep changing, but the underlying requirement — explicit, accurate, consistent facts about what you sell — is what every one of them consumes. Effort spent here transfers across ChatGPT, Gemini, Google AI Mode and whatever follows, which is not true of tactics built for a single surface.
It compounds in the other direction too. Shoppers increasingly describe situations rather than products, and structured attributes are what let you answer situations at all — the shift underlying conversational commerce replacing keyword search, and the reason so much of this influence never registers as a click in zero-click commerce.
The bottom line
Being on the AI shelf is not a ranking you win. It is a threshold you clear: can a model state, with confidence, that your product meets this shopper's constraint? Clear it and you become recommendable everywhere at once. Miss it and you are invisible regardless of how good the product is.
Want to see how AI assistants currently read your catalog? Book a demo with agentShop for a read on your product data across ChatGPT, Gemini and Google AI Mode.
