Category Architecture for Agent Discovery: Fan-out Queries, Taxonomy and the Category Page

What is category architecture for agent discovery?

Category architecture for agent discovery is the work of shaping an e-commerce site's category tree, category pages and product-to-category mapping around the way AI agents split a request and find products. Classic categorisation answered two questions: can the shopper find the product from the menu, and does the category page rank on Google? In 2026 a third question decides: when ChatGPT, Google AI Mode or Amazon Rufus breaks "a linen shirt for hot summer weather that does not crease and ships to Europe" into sub-queries, in which of those sub-queries is your category retrieved?

We first published this article in September 2025 under the title AI-assisted product categorisation; the causes and consequences of mis-categorisation from that version still hold. What changed is the customer of the category page. Google AI Mode's query fan-out, the Google product taxonomy update of January 2026 and the product_category field in the OpenAI product feed turned category from a matter of on-site order into a matter of whether the agent finds you. Our thesis at Webtures: a category is no longer a shelf, it is an answer to one of the sub-questions the agent asks.

How does an agent retrieve a category page?

Google AI Mode splits a single request into sub-queries that run in parallel; this is called query fan-out. "A linen shirt for summer" becomes "men's linen shirt", "non-crease summer fabric", "Turkish brands that ship to Europe", "190 gram linen shirt price"; each part is retrieved separately and the results are merged into one answer. On the shopping side that retrieval runs through the Shopping Graph: 50 billion product listings, 2 billion refreshed every hour. Your category is represented in that graph as the intersection of product attributes, taxonomy node and merchant signals.

The practical consequence cuts both ways. First, your category page can now be retrieved not for one keyword but for dozens of sub-queries that fall into that category; visibility is probabilistic, not binary. Second, industry measurements show that e-commerce category pages and informational queries are the page types most exposed to AI Mode, while branded and transactional queries hold steady. The classic traffic of the category page shrinks while the category content the agent retrieves gains. Which side you are on is decided by how parseable your category page is for the agent.

On the ChatGPT side the mechanism is similar: the product_category field in the OpenAI feed carries a path from broad to narrow category; the agent parses the user's request into product type, attributes and constraints and matches it against that path. Amazon Rufus predicts the category node from the listing data; a product in the wrong node never appears for the right question.

What does mis-categorisation cost in the agent era?

The causes of mis-categorisation have not changed: manual error and fatigue, product data that is missing or reduced to a model code, the impossibility of checking tens of thousands of products by hand, and differences of language and terminology between markets. A slipper is footwear in one market and homewear or a beach product in another. The consequences got heavier:

• Dropping out of the cluster. Google AI Mode groups the same product from several sellers into one cluster. A wrong google_product_category or a missing GTIN leaves the product outside the cluster; the agent does not see you when it compares.
• Missing the fan-out. If the sub-query is "non-crease summer shirt" and the product sits in a broad node such as "men's tops", the match stays weak at retrieval time.
• Filter loss. Size and colour filters on the category page drop a mis-categorised product from the list; for the agent that means the product does not exist in that combination.
• Feed warnings. Merchant Center removed some paths in the January 2026 taxonomy update; products carrying the old text path get warnings, products carrying the numeric ID are unaffected.
• Loss of trust. As with product descriptions, contradiction is worse than invisibility: a case listed under "phones" leads the agent to recommend the wrong product to the user.

Which taxonomy should you use?

Keeping a single mapping table between three taxonomies is the cheapest step in agent discovery.

The rule is simple: the site category tree is written in the shopper's language, the feed category in the platform's language, and the mapping between the two lives in one table. Our team's Agentic Commerce Assessment Set asks this directly: "To what extent does the data conform to the shared schema expected by the target marketplaces and agent interfaces?" The answer is a schema conformance report, not a feeling.

How do you build a category page for an agent?

1. Open with a definition sentence. The first paragraph of the category page says what the category is, who it is for and how many products it holds. "Men's linen shirts: 100% linen and linen blends, sizes 38 to 46, 24 styles." An agent can cite that paragraph on its own.
2. Publish filters as attributes too. Fabric, size and colour filters exist not only in the interface but in the product data as product_detail and in the schema as additionalProperty. A fan-out sub-query is very often a filter value.
3. Answer the sub-questions under sub-headings. Below the listing, 3 to 5 question headings with short answers: "Does linen crease?", "Which weight suits summer?". The section serves the shopper and the fan-out at once.
4. Use CollectionPage and ItemList schema. Make what the category page is and which products it lists machine-readable; connect every product card to its own Product node.
5. Solve pagination and infinite scroll in raw HTML. Most major AI bots do not execute JavaScript; if the products on page 2 are not in the raw HTML they do not exist for the agent.
6. One canonical, one language per category. URLs generated from filter combinations canonicalise to the main category; on a multilingual site the category name uses the target market's terminology.
7. Close empty and thin categories. A category with fewer than three products is noise for the agent and thin content for the search engine; merge it or noindex it.

How does AI-assisted categorisation work?

Category assignment is already being solved as an AI problem on the big platforms: Shopify suggests the category from its standard taxonomy as you add a product, Merchant Center maps feeds to its own taxonomy, Amazon predicts the node at listing time. In your own catalogue the same approach is built in three steps: an embedding is generated from the product title, description and attributes; the embedding is compared with the taxonomy nodes and a category is suggested with a confidence score; low-confidence suggestions go to human review. Image classification steps in where text falls short (titles that carry only a model code).

Two numbers need measuring here: category accuracy (what share of sampled assignments are correct) and coverage (how many products sit only in a broad parent node). Both feed directly into the "data completeness" and "platform fit" dimensions of our Readiness Score.

How should cross-border brands set up their categories?

On ready-made platforms such as Ticimax, ikas or İdeasoft the category tree is usually designed in Turkish while the feed category is mapped automatically. The first job is to audit that automatic mapping: whether the google_product_category values in Merchant Center are numeric IDs, whether any products fall on paths removed in the January 2026 update, whether the category name carries the right term in the target market's language. In our own export matrix the focus in B2C markets such as the United States and the United Kingdom is the Merchant Center feed and enriched category content; in B2B markets such as Germany it is a PIM-based single record. In the Gulf, Arabic category and attribute data is almost empty; the brand that enters looks unrivalled.

Category architecture is designed together with the product description: if the product has a description agents can read, the category mapping comes out right too. We covered the visibility side of category pages in category pages in GEO and the protocol layer in the protocol map. You can see the score of your category and feed conformance with the free scan on our Agentic Commerce Readiness page, and audit how agents navigate your site with our Agent Experience team.