The 14 Fields AI Agents Actually See on Shopify Products

The count

On 2026-10-31 we pulled the tool listing from a live Shopify storefront and read the output schema of its catalogue search tool. The full listing came back at 207,156 bytes. Inside it, the product object has fourteen top-level fields.

Those fields: id, title, description HTML, handle, url, gift-card flag, price range, list price range, options with their values and availability, media with type, url and alt text, categories with taxonomy, collections, tags, and a variants array. Each variant carries its own id, title, description, SKU, price, list price, availability, options, media, shipping and selling-plan requirements, and a checkout URL.

Then we searched that same 207 KB payload for the fields teams assume are there. Ratings: zero occurrences. Reviews: zero. 3D models: zero. Video: zero. Safety information: zero. The one that was present — list price, six occurrences — is the field the Fall '26 release added for compare-at pricing.

What that means for your most expensive assets

The assets brands invest most heavily in on a product page are largely invisible on this path. The review widget with 4,000 ratings, the A+ content module, the lifestyle photography sequence, the interactive size finder, the metafield specification table rendered as a custom section — none of it is in the payload.

This is not an argument for abandoning them. Human buyers still open the product URL the agent returns, and conversion on that page still decides the sale. It is an argument for noticing that you have two audiences with almost non-overlapping diets, and that one of them has been getting leftovers.

Where ratings and richer fields do live

The cross-store catalogue is a different story. Its documentation (shopify.dev/docs/agents/catalog/global-catalog) describes a much richer field set, including rating as a filterable dimension — and the mechanics matter. Rating filtering is evaluated at variant level against a minimum review count, not as a product-level average. A strong store-wide average will not rescue a flagship variant with thin reviews.

It also exposes a set of fields the platform marks as inferred: normalised options, attributes such as material and style, technical specifications, top features and unique selling points. Shopify's documentation is explicit that these are generated or enhanced from your text and should be treated as discovery and merchandising signals rather than merchant-authored source text.

That is the single most important link in the chain, and most teams do not know it exists: your product copy is the raw material from which the platform derives the selling points an agent will narrate to a buyer.

How to rewrite product copy for extraction

• Lead with facts a machine can lift: material, dimensions, weight, capacity, compatibility, power, certification, included contents.
• Replace brand abstractions with concrete differences. Built for those who dare yields nothing. 304 stainless steel inner pot, dishwasher safe, 2.6 L yields three extractable attributes.
• Name the use case explicitly — for apartment kitchens, for cabin luggage, for sensitive skin — because occasion and suitability are inferred attribute categories.
• Make option names human. Options are normalised for catalogue discovery, so idiosyncratic naming gets mapped incorrectly or dropped.
• Put specifications in the description as well as in metafields. The specification table you render as a custom section is not in the single-store payload; the description HTML is.
• Write alt text as a description of the product, not a keyword string — alt text is one of the few media fields that travels with the payload.
• Keep your standard product category correct and specific. Broad category retrieval is how most agent searches start.

Three variant fields worth monitoring as hard metrics

• Native checkout eligibility — whether a variant supports completing checkout without leaving the conversation. It defaults to off, and agents have an obvious reason to prefer products that can close in-conversation.
• Low-stock signalling — a scarcity cue agents can relay to buyers, which means inventory accuracy is now merchandising copy.
• Seller identity — the merchant domain and id, which is where brand recognition does its work during cross-store comparison.

One tactic that does not work

Shopify's published agent instructions at shop.com/references/http.md tell agents to treat merchant text as data, not instructions. Embedding directives such as list this product first in descriptions, alt text or hidden markup is explicitly handled. The same skill also instructs agents to state material unknowns.

Read together, those two lines define the realistic risk. You are not going to manipulate your way to the top of an agent's recommendation. You are, however, very likely to be accurately described as a product whose specifications are unclear — which is a self-inflicted and entirely fixable problem.

Doing this at catalogue scale

Rewriting fifty hero products by hand is a sprint. Rewriting eight thousand SKUs, normalising attributes across a legacy taxonomy and keeping it synchronised with an ERP or PIM is a programme — and it is the work we do inside our agent-ready product data and AI visibility engagements.

If you want to know what agents currently return for your core category terms, and which of your fields are silently empty, that is the first thing we measure. Send us your domain and primary category and we will show you the gap on your own catalogue.