How to prepare fashion product feeds for ChatGPT, Gemini and Perplexity shopping
AI assistants recommend products from structured feeds. A practical guide to the fields, variant logic and update discipline fashion retailers need to appear accurately in ChatGPT, Gemini and Perplexity.
KEY TAKEAWAYS Summary by the editors
- AI shopping assistants rely on structured product feeds, so missing or inconsistent attributes reduce the chance that a fashion item is recommended accurately.
- OpenAI's product feed specification requires fields including item_id, title, description, url, brand, image_url, availability and price, and uses group_id and variant_dict to describe size and colour variants.
- Google Merchant Center requires colour, size, gender, age group and item group ID for apparel in markets including Germany, France, the UK and the US.
- Both OpenAI and Google tell merchants not to invent GTINs; where no valid identifier exists, the specifications provide other routes such as brand plus MPN or identifier_exists set to no.
- Feed freshness matters as much as completeness: OpenAI asks merchants to update feeds when sales start or end and when stock changes.
To appear accurately in ChatGPT, Gemini and Perplexity shopping results, a fashion retailer needs a complete, variant-level product feed with correct sizes, colours, prices and stock, submitted through each platform's merchant programme and refreshed whenever prices or availability change. The work is mostly product data hygiene rather than new technology.
How do AI assistants get fashion product data?
Assistants combine web crawling with structured data that merchants submit. OpenAI publishes a product feed specification for ChatGPT shopping. Google uses Merchant Center data across Search, AI Mode and the Gemini app, and in January 2026 announced dozens of new Merchant Center attributes designed for conversational commerce, such as answers to common product questions, compatible accessories and substitutes. Perplexity launched a free Merchant Program in November 2024 through which large retailers share product specifications, and uses Shopify for product information from US Shopify merchants.
Crawled pages are still used, but a feed gives you control over what the assistant reads: the exact price, the current stock per size and the canonical product link.
Which feed fields matter most for fashion?
The core fields overlap across platforms, although names differ. The table compares the OpenAI format with the Google-compatible names.
| Purpose | OpenAI field | Google field | Fashion note |
|---|---|---|---|
| Unique item | item_id | id | One row per purchasable size and colour combination |
| Parent product | group_id | item_group_id | Required for variants; links all sizes and colours of a style |
| Variant options | variant_dict, size, color | size, color | Colour must match the image; size labels need a size system |
| Size system | size_system | size_system | EU, UK, US or JP labels are not country codes |
| Audience | gender, age_group | gender, age_group | Required for apparel in major Google markets |
| Stock | availability | availability | OpenAI accepts in_stock, out_of_stock, pre_order, backorder, unknown |
| Price | price, sale_price | price, sale_price | OpenAI format: amount plus ISO 4217 code, for example 79.99 USD |
| Identifiers | gtin, mpn, brand | gtin, mpn, brand | Never invent a GTIN |
How should size and colour variants be structured?
Variant logic is where fashion feeds most often fail. OpenAI's specification asks for one row per purchasable item or variant, each with its own price, availability, URL and images. Every variant row carries a group_id that differs from its item_id, sets listing_has_variations to true, and includes a variant_dict such as colour Black and size 10. Google's specification similarly requires item_group_id for variants in markets including Germany, France, the UK and the US.
For apparel, Google requires colour, gender and age group for products targeted to those markets, and size for clothing and shoes. If a style comes in several colours, each colour needs its own image so that the assistant does not show a navy jacket when the shopper asked for black.
What are the most common feed mistakes?
- Invented identifiers: both OpenAI and Google say not to guess or make up GTINs. Use brand plus MPN, or set identifier_exists to no in Google-compatible feeds where no identifier exists.
- Parent-only rows: listing a style once without size rows means the assistant cannot tell whether a size is in stock.
- Stale availability: OpenAI's feed rejects rows with missing or unrecognised availability values, and asks merchants to update availability when stock changes.
- Unsynchronised sale prices: OpenAI asks for feeds to be updated when a sale starts or ends; sale windows in the feed do not schedule changes automatically.
- Thin titles and descriptions: a title such as 'Jacket 4471' gives an assistant nothing to match against a request for a 'waterproof lightweight rain jacket'. OpenAI suggests titles of up to 150 characters and descriptions of up to 5,000.
- Prices in offer IDs: OpenAI's specification explicitly says never to include price in an offer ID.
How often should a fashion feed be updated?
OpenAI does not set a fixed cadence, but its guidance ties updates to events: when a sale starts or ends and when stock changes. For fashion, where single sizes sell out quickly, that implies at least daily full feeds plus more frequent stock and price updates for fast-moving lines. A recommendation for an out-of-stock size damages trust in both the assistant and the brand.
What else should be in the feed beyond the basics?
Assistants answer questions that shoppers would otherwise ask a sales associate. Material composition, care instructions, fit description (slim, regular, oversized), length, rise, sleeve type, occasion and sustainability certifications all help, provided they are accurate and evidenced. Google's new conversational attributes point in the same direction. Include delivery times and return conditions where the specification allows, because shoppers increasingly ask about them in the same conversation.
How do you measure whether the feed is working?
Track feed health first: the share of rows accepted, rejected and warned in each merchant dashboard. Then segment web analytics by AI referral source to see traffic, conversion and returns from assistant visits. Finally, run regular manual tests: ask each assistant for typical customer requests in your categories and check whether your products appear with correct prices, sizes and images.
Ownership matters as much as format. In many fashion companies, product data is split between design, merchandising, e-commerce and marketing teams, and feeds are maintained by whoever runs paid search. Agent feeds deserve a named owner with authority to fix data at the source in the product information system, rather than patching values in a feed tool. That owner should also coordinate translations, because assistants answer in many languages and a missing localised title or material description reduces visibility in that market. A short monthly review of rejected rows, top missing attributes and customer queries that returned no product is usually enough to keep feed quality improving steadily. Treat each improvement as reusable: the same clean attributes improve on-site search, marketplace listings and wholesale data, so the investment is not tied to any single AI platform succeeding.
Frequently asked questions
How do I get my products into ChatGPT shopping?
OpenAI publishes a product feed specification for merchants. The feed needs fields such as item_id, title, description, url, brand, image_url, availability and price, with group_id and variant_dict for variants. Participation in specific features such as checkout requires separate enablement.
Does Gemini use Google Merchant Center data?
Yes. Google uses Merchant Center product data for shopping in Search, AI Mode and the Gemini app, and announced new conversational attributes for these surfaces in January 2026. Apparel needs colour, size, gender, age group and item group ID in major markets.
Can fashion retailers sell through Perplexity?
Perplexity launched a free Merchant Program in November 2024 for large retailers to share product data, alongside a Buy with Pro checkout for selected US merchants. It also uses Shopify for product information from US Shopify merchants.
What if my products have no GTIN?
Do not invent one. OpenAI says to submit an MPN together with the brand instead, and Google-compatible feeds allow identifier_exists to be set to no when no identifier is available.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
