6 November 2026International edition
Vol. I · No.
6 November 2026
AI IN FASHION
DAILY
The daily briefing on AI in the fashion business
Where fashion meets artificial intelligence.
Commerce & Marketing · Stories

Fashion product feed readiness for AI agents: 20 attribute checks

Structured product data determines whether apparel appears in ChatGPT Shopping and conversational discovery. These 20 attribute and format requirements prepare catalogues for agent ranking.

white printer paper on white textile
Photo: Laura Chouette / Unsplash

KEY TAKEAWAYS Summary by the editors

  1. OpenAI's product feed specification defines 79 fields with 18 required, including stable identifiers, price with currency code, availability status and product URLs.
  2. Each fashion variant requires a unique GTIN per colour and size combination, with branded apparel excluded from Google and ChatGPT surfaces without valid barcodes.
  3. AI agents parse material composition, fit type, measurements and care instructions as decision-critical attributes that filter products from conversational queries.
  4. Product feeds must update inventory and pricing within a 15-minute window, with inconsistency between feed data and on-page markup treated as quality failure by agents.

What product data do AI shopping agents require from fashion catalogues?

OpenAI's product feed specification defines 79 fields, of which 18 are required, whilst AI agents parse full Product schemas including AggregateRating, Review, Offer, ShippingDeliveryTime and MerchantReturnPolicy objects. Since March 2026 Shopify makes products discoverable in ChatGPT by default, and lists them in Microsoft Copilot, Google AI Mode and the Gemini app. Nearly half of consumers report AI use in shopping at some point in 2025, with 11 per cent saying they tried it for the first time whilst holiday shopping, and apparel and fashion prompts trigger shopping cards 62 per cent of the time.

Since March 2026 ChatGPT Shopping has been used for product discovery, with merchants sharing their product feed along with their promotions. AI agents pull from product feeds, reviews, brand content and third-party data, with brands that invest in structured product data more likely to be recommended. The operational requirement is strict: product feeds must update inventory state and pricing within a 15-minute lag window at most.

Which identifiers and commercial fields are mandatory?

Five core fields form the minimum viable product record. The ACP product feed specification requires item_id as a unique identifier with a maximum of 100 characters that must remain stable, title as product name with a maximum of 150 characters, description with a maximum of 5,000 characters, url as the product page URL that must resolve HTTP 200, and price as a number with ISO 4217 currency code. Availability must be declared as an enumerated value: in_stock, out_of_stock, pre_order, backorder or unknown.

OpenAI's file-upload product specification explicitly says to keep item_id, group_id and offer_id stable when price, stock, titles or images change, with the ACP product-feed RFC similarly requiring that variant IDs remain stable across feed updates. When on-page markup and feed data disagree, the agent treats it as a quality failure.

Why does each apparel variant need a separate GTIN?

Every sellable product variation, including each colour and size combination, requires a unique GTIN assignment. Products without GTINs are excluded from Google's trust-based layers and get reduced confidence from ChatGPT's commerce engine. One 2026 analysis estimated that roughly 60 per cent of ecommerce catalogues carry missing GTINs, inconsistent attribute naming or stale inventory states, all of which cause agents to down-rank or exclude products outright.

For fashion merchants the GTIN requirement affects every SKU. Branded clothing requires a valid GTIN, with products from Nike, Adidas, Levi's or any other established brand requiring a valid GTIN for each product. Google's requirements remain that GTIN is mandatory for all products that have one assigned by the manufacturer, with each unique product variant needing its own unique GTIN. The only exception: Meta and Google Shopping have an identifier_exists field, which can be set to false for handmade custom goods, though for any manufactured good, apparel or standard retail item a GTIN is required to maximise ad performance.

What apparel-specific attributes do agents parse for filtering?

Conversational queries map to structured fields. Conversational queries like "black wool coat under £300, ships by Friday" map cleanly to structured product attributes, making apparel retailers some of the highest-converting AI shopping traffic in 2026. Product attribute coverage measures how consistently the important attributes of a product are populated across a catalogue, with those attributes for fashion including colour, material, pattern, fit, length, sleeve type, neckline, occasion and style.

AI agents pull sizing information from structured data, product descriptions and customer reviews, and if sizing data lacks consistency and structure or mixes unlabelled body and garment measurements, agents cannot reliably recommend the right size to shoppers. Products with complete attribute data typically see 30 per cent higher conversion and 25 per cent fewer returns.

Schema.org markup reinforces feed data. The highest-ROI schema fields for apparel include color as exact colour name matching the product title, size using SizeSpecification with suggestedMeasurement for international sizing, material as fabric composition, and an image array of at least three images per product. Recommended Offer properties include availability, itemCondition, url, priceValidUntil, validFrom, validThrough, hasMerchantReturnPolicy and shippingDetails, whilst Product properties include brand.name, sku, mpn, the gtin family, description, category, color, size, material, pattern, audience, hasCertification, review, aggregateRating, inProductGroupWithID, isVariantOf and subjectOf.

How should fashion merchants structure variant relationships?

OpenAI's current feed format requires a separate row for each variant selection, with a distinct item ID, a shared group ID and a structured variant dictionary. The ACP specification includes item_group_id plus variant attributes for colour and size, defining parent-child grouping so agents resolve variants in one hop. This structure prevents the wrong variant being selected or variants collapsing into one ambiguous row.

One trouser style in seven colours, ten sizes and three lengths adds up to 210 SKUs, with each one carrying attributes worth capturing: material composition, fit, rise, length, care instructions and seasonality. Every description should carry at least five machine-parseable facts: explicit dimensions, weight or capacity with units, material composition with percentages, and at least one concrete use case.

What shipping and return data do agents surface to shoppers?

AI agents parse ShippingDeliveryTime objects, with hasMerchantReturnPolicy critical for agents that filter by return flexibility and shippingDetails required for agents surfacing delivery-time comparisons. These fields answer queries that specify fulfilment constraints, such as delivery by a particular date or free returns within a defined window.

Shipping data has become table stakes for conversational commerce. Queries including "ships by Friday" require structured shipping and availability data to be matched, whilst merchant listing markup targets pages where the visitor buys from the merchant, with apparel sizing, shipping and returns, and filling in the required merchant listing properties generally making the page eligible for product snippets too.

Which review and rating signals affect agent recommendations?

OpenAI researchers told reporters that the model takes organic reviews and other high-fidelity web content into account when recommending specific products and noted that they have steered the model specifically towards trustworthy sources. According to OpenAI representatives briefing reporters, the system explicitly considers "user experiences shared on Reddit may be considered more trustworthy than paid marketing or reviews posted on a product page".

Structured rating data remains essential. The highest-priority Schema.org Product markup elements for AI agent retrieval include aggregateRating, with product schema needing to include all the core fields: Product, Offer, Brand, AggregateRating, Review and Availability. Product reviews shown in ChatGPT come from various online sources, including publishers and user forums like Reddit.

Attribute classRequired fieldsAgent use caseFailure mode
IdentifierGTIN, item_id, SKUCross-merchant deduplicationProduct appears as multiple unrelated listings
CommercialPrice, priceCurrency, availabilityBudget and stock filteringItem drops from priced or in-stock queries
Variant structureitem_group_id, colour, sizeSeparating variants from distinct productsWrong variant selected or variants collapse
Material & fitMaterial composition, fit type, measurementsFiltering conversational queriesProduct excluded from attribute-specific searches
FulfilmentShippingDeliveryTime, hasMerchantReturnPolicyDelivery constraint matchingProduct hidden from time-sensitive queries
Trust signalsAggregateRating, Review, brand.nameRanking and recommendation confidenceLower position or exclusion from results

Twenty product data checks prepare fashion catalogues for agent discovery:

  • Assign stable item_id to every product, maintaining consistency across feed updates and price changes
  • Provide product title under 150 characters and description under 5,000 characters with machine-parseable facts
  • Include working product page URL returning HTTP 200 status
  • Declare price as numeric value with ISO 4217 currency code
  • Set availability to enumerated value: in_stock, out_of_stock, pre_order, backorder or unknown
  • Assign unique GTIN to each sellable variant, covering every colour and size combination
  • Register GTIN through GS1 for branded apparel sold through retail or marketplace channels
  • Populate item_group_id to link variants sharing the same style
  • Define colour using exact name matching product title, avoiding ambiguous or marketing terms
  • Specify size with SizeSpecification schema, including body or garment measurement clarification
  • Declare material as fabric composition with percentages
  • Add fit type as structured attribute
  • Include garment measurements or fit guidance to reduce sizing uncertainty
  • Specify care instructions as parseable text or structured field
  • Populate pattern, neckline, sleeve type, occasion and style where applicable
  • Add hasMerchantReturnPolicy with return window and conditions
  • Include shippingDetails with delivery time range
  • Provide aggregateRating and review markup on product pages
  • Update feed inventory and pricing within 15-minute window
  • Ensure on-page Schema.org markup matches submitted feed data

Frequently asked questions

Does every fashion product variant need its own GTIN for AI agent discovery?

Yes. Each unique product variant, including every distinct combination of colour and size, requires its own unique GTIN. Products without GTINs are excluded from Google's trust-based layers and get reduced confidence from ChatGPT's commerce engine. The only exception is handmade or custom goods, which can set identifier_exists to false in feed data.

Which product attributes are mandatory for ChatGPT Shopping?

OpenAI's product feed specification defines 18 required fields from a total of 79. The mandatory attributes include item_id as a stable unique identifier, title under 150 characters, description under 5,000 characters, product page url, price with ISO 4217 currency code, and availability status as an enumerated value. For fashion specifically, GTIN is required for all branded apparel.

How quickly must fashion product feeds update for AI agent accuracy?

Product feeds must update inventory state and pricing within a 15-minute lag window at most. When on-page markup and feed data disagree, AI agents treat it as a quality failure. Stale pricing or availability data causes agents to down-rank or exclude products, particularly for queries with delivery or budget constraints.

Stories