Product Data for AI Shopping Assistants: A Prep Guide

Product data for AI shopping assistants comes down to one thing: a complete, accurate and current product feed, backed by matching markup on your product pages. An AI shopping agent (an assistant that researches and recommends products for a shopper, and in some cases completes the purchase) can only recommend what it can read and trust. This guide walks through the preparation steps, the platform differences, and what to verify before you rely on any claim, including some in this article.

A note on terms. MCP (Model Context Protocol) is an open protocol that connects an AI assistant to tools and data. It is not the same thing as a product feed. Feeds are how platforms like Google, OpenAI and Microsoft learn about your catalog. Protocols such as UCP and ACP deal with checkout. Microsoft states the split directly: feeds primarily enable discovery, while UCP enables transactions. If you want the protocol side, see "Agentic Commerce Protocol Compared: ACP, UCP, AP2 and MCP".

Why does product data decide whether AI assistants recommend you?

Assistants compare products using the data they can parse. If price, availability, shipping or returns information is missing or out of date, the system may skip you. Microsoft's playbook on AI shopping, as summarized by a secondary write-up (the original was not opened for this article), makes this point: incomplete or stale catalog, availability, price, shipping and return data may mean an AI system fails to recommend the merchant at all.

The practical consequence: product data quality is now a visibility issue, not only a feed-approval issue. It is also cheaper to fix than most other levers.

What do you need before you start?

• Admin access to your store platform and to the product catalog export.
• A Google Merchant Center account (most platforms accept or import a Google-style feed).
• Access to your theme or template, to check Schema.org Product markup (structured data that labels price, brand and availability for machines).
• A list of your policy pages: shipping, returns, privacy and contact.
• A spreadsheet to track gaps by product.

How do you prepare product data for AI shopping assistants, step by step?

1. Export and audit the catalog. Pull every active product and variant. Count how many lack a GTIN, brand, description, image or category. Sort by revenue so you fix the products that matter first.
2. Give every item a stable, unique ID. The ID should not change when you edit the title or price. Secondary guidance on OpenAI's feed lists a unique, stable id as a minimum field, alongside title, description, link and image link.
3. Add identifiers. Fill GTIN, brand and MPN (manufacturer part number) where they exist. See the GTIN section below.
4. Rewrite titles and descriptions as plain, factual text. Put the distinguishing facts (product type, brand, size, color, material) in the title. Keep descriptions free of HTML and marketing filler. An assistant matching "waterproof hiking boots, wide fit, size 44" needs those words in the data.
5. Fill variant and category attributes. Secondary sources on Google AI Mode recommend completing Google product category, product type, and variant attributes such as color, size, material, gender and age group.
6. Make price, availability, shipping and returns accurate and current. These are the fields an assistant uses to show a result and decide if it is safe to recommend. Sync stock and price changes as often as your platform allows.
7. Align structured data with the feed. Add Schema.org Product markup to product pages and make sure it matches the feed. A vendor guide notes that structured data helps Google match on-page product data with Merchant Center data. Mismatched prices between page and feed are a common cause of disapprovals.
8. Check images. Use clean, high-resolution main images and additional angles. Confirm each image URL is publicly reachable.
9. Publish policy and contact information. Returns policy, shipping terms and a support contact are listed among the requirements for Google's agentic checkout, per a secondary summary.
10. Check that automated access is not blocked. A secondary guide reports that OpenAI says some retailers block automated access and get skipped. Review your robots.txt and any bot-protection settings, and confirm with each platform which crawlers to allow.
11. Submit and monitor. Upload to each channel, review diagnostics, and fix errors weekly.

How do you optimize a product feed for AI?

Product feed optimization for AI is mostly the standard feed hygiene above, applied more strictly. The common base is Google's feed specification. Per the research behind this article, Perplexity ingests the Google Shopping spec, OpenAI accepts a Google-compatible file, and Microsoft accepts a Google Merchant Center import. Google also states that merchants can use existing Merchant Center shopping feeds for UCP, and secondary sources say the same feed that powers Shopping also powers AI Mode.

So the efficient approach is one master feed, built to Google's spec, then adapted per channel. Do not build four separate catalogs by hand.

Where AI changes the work is in descriptive depth. Traditional shopping ads lean on bids and keywords. Assistants answer natural-language questions, so attributes (material, fit, use case, compatibility) matter more than they used to. Fill in optional attributes that genuinely describe the product, and skip ones that do not apply.

Why do GTINs matter for AI shopping?

A GTIN (Global Trade Item Number, the number under a barcode) lets a platform recognize that your product is the same item sold by other merchants. For an AI shopping agent comparing offers, that matching is how it groups listings and compares price and availability.

Some vendor blogs claim missing GTINs significantly reduce eligibility. That claim was unsourced in the material reviewed, so treat it as unverified. The safer reading: GTINs are a long-standing Google requirement for many branded products, and there is no upside in leaving them blank.

Practical rules:

• Use the manufacturer's GTIN. Do not invent one.
• For custom or handmade items with no GTIN, follow each platform's rules for products without identifiers, and provide brand and MPN where applicable.
• Keep GTINs at the variant level, since each size and color usually has its own.

How do the main platforms differ?

Platform	What it reads	Notes to verify
Google (AI Mode, Gemini)	Merchant Center feed plus Schema.org markup (secondary sources)	UCP checkout scope, regions and the Buy button attribute: check Merchant Center Help
OpenAI (ChatGPT)	Product feed in a Google-compatible file or OpenAI's own format	Formats, refresh rate and size limits differ across sources: check OpenAI's current feed spec
Microsoft (Copilot)	Merchant Center, including Google imports	Microsoft says Merchant Center will support UCP and richer signals such as returns and support policies

Google

Google's UCP (Universal Commerce Protocol) page says it enables agentic actions on AI Mode and Gemini, starting with direct buying, and that the merchant remains Merchant of Record. Shopify's engineering blog says the merchant profile is published at /.well-known/ucp on the merchant's site. One retailer-platform article says only listings with a specific native commerce attribute show the Buy button, but this was not verified in Google's documentation. Regional availability is also unclear: sources conflict between U.S. only and a wider list. Check the current Merchant Center Help page before planning around it.

OpenAI and ChatGPT

OpenAI's key-concepts documentation describes a secure, regularly refreshed feed covering identifiers, descriptions, pricing, inventory, media and fulfillment options, with required fields for correct price and availability and recommended attributes (rich media, reviews, performance signals) that improve ranking and trust. Sources conflict on file format (CSV, TSV, XML, JSON or JSONL) and on refresh frequency (daily snapshots versus as often as every 15 minutes). The status of Instant Checkout has also changed over time, so confirm it on OpenAI's own pages. On Shopify, a secondary source describes the route as Sales channels, then Agentic, then ChatGPT, with eligibility conditions. Confirm in your admin.

Microsoft

Microsoft's official agentic commerce page says its Merchant Center will support UCP. Its feed attributes are evolving, so read the current documentation before changing your feed.

What can go wrong?

• Price or stock mismatch. The feed says in stock, the page says sold out. Expect disapprovals and lost trust.
• HTML or promotional text in descriptions. It may be rejected or displayed badly.
• Changing IDs. A platform may treat the product as new and lose its history.
• Fabricated GTINs. This can cause account-level problems.
• Blocked crawlers. Bot protection can hide your product pages from assistants.
• Outdated advice. This area changes monthly. Specs, regions and beta status in this article may be out of date by the time you read it.
• Over-reliance on any single claim. Statements such as UCP "becoming required" were single-sourced. Verify before spending budget.

Also keep in mind what feed work cannot do. It does not guarantee a recommendation, and no one outside the platforms controls how an assistant ranks results.

How does this connect to MCP?

Once your catalog is clean, an assistant connected through MCP can read it more reliably. For how that works on Shopify, see "Shopify MCP Explained: How AI Assistants Read Your Catalog". Clean data helps both channels: the platform feeds and your own assistant workflows.

A simple maintenance routine

• Weekly: review feed diagnostics in Merchant Center and any other channel.
• Monthly: re-run the catalog audit for missing GTINs, brands and attributes.
• Quarterly: re-read each platform's feed spec and eligibility notes for changes.
• After any theme or app change: re-test structured data on a sample of product pages.