Product Data for AI Agents
Why structured product data underpins agentic commerce, how it differs from SEO, AEO and GEO, and what good data, feeds and catalog APIs look like.
In short: AI agents can only recommend and buy what they can read accurately. Structured product data (clear attributes, variants, prices, availability, and standard identifiers) is the input every shopping agent depends on. Merchants supply it through feeds, on-page markup, and platform catalogs, and agents also reach it through cross-merchant catalog APIs.
Why data comes first
A human shopper can look at a product photo, scroll the reviews, and infer that “navy” and “dark blue” mean the same thing. An agent answering “waterproof hiking boots under $150 in a size 10 wide, delivered by Friday” needs that information in fields it can filter on. If waterproofing appears only in an image, or width is missing from the variant data, the product is likely to drop out of the answer.
Agents also act on the data. When an agent places an order, the price, stock status, and shipping promise it used become the terms the shopper expects. Stale or wrong data turns into failed checkouts, cancelled orders, and disputes (see consumer protection and liability).
Commerce platforms make the same point. In its September 2026 announcement of catalog enrichment tools, Commerce (the parent of BigCommerce and Feedonomics) wrote that AI agents can only answer questions about products “as well as your data allows.”
Product data vs. SEO, AEO and GEO
These terms often get lumped together, but they answer different questions. SEO, AEO and GEO are about being found and mentioned. Product data is about being understood and bought.
The three acronyms
- SEO (search engine optimization) is the long-standing practice of helping web pages rank in traditional search results, such as Google’s list of links. The unit is the web page, and success means someone clicks through to your site.
- AEO (answer engine optimization) became popular with featured snippets and voice assistants, which reply with a single direct answer instead of a list of links. The goal is for your content to be the source of that answer.
- GEO (generative engine optimization) is the newest label. It means getting mentioned or cited in answers written by generative AI, such as ChatGPT, Gemini, Perplexity, or Google’s AI Overviews and AI Mode.
The line between AEO and GEO is blurry, and many people use the two interchangeably. All three are mostly about content: web pages, articles, reviews and mentions elsewhere on the web that a search engine or AI model reads and then summarizes.
What product data does differently
Product data is the factual record of what you sell: each item’s name, price, sizes and colors, stock level, barcode numbers (identifiers such as GTINs), and shipping and return terms. It is usually supplied as structured fields, in a feed, an API or markup on the page, rather than written as prose.
An analogy: SEO, AEO and GEO are like getting your store mentioned in a city guide. Product data is the shelf label and the stock count inside the store. The mention gets an agent’s attention; the label tells it whether the item actually fits the request and can be bought right now.
| SEO | AEO / GEO | Product data | |
|---|---|---|---|
| Question it answers | Will my page rank? | Will an AI mention or cite us? | Can an agent understand, compare and buy this item? |
| Main input | Web pages and links | Content and mentions across the web | Structured fields: price, variants, availability, identifiers |
| Who reads it | Search engine crawlers | AI models and AI search crawlers | Shopping agents, feeds, catalogs and checkout systems |
| Success looks like | A click from a results page | A mention or citation in an answer | The product is chosen and the order goes through at the right price |
| If it’s wrong | You rank lower | You are left out or described inaccurately | Failed checkouts, wrong orders, returns and disputes |
Why the difference matters in commerce
When an AI only answers questions, a loose mention can be enough. When it shops, it has to act: filter by size, check stock, quote a price and place an order. That takes exact data, not persuasive copy. A brand can show up often in AI answers and still lose the sale if its feed says an item is in stock when it sold out yesterday.
The two also overlap. Good product data helps AEO and GEO, because AI search tools read the same structured markup and merchant feeds. Google, for example, says there are no extra requirements to appear in its AI features beyond normal Search eligibility. For most merchants the practical order is to make product data complete and current first, then work on visibility. Agent optimization covers the visibility side in detail.
What good product data contains
Core attributes
The fields every major feed specification requires or strongly recommends:
| Field | Purpose | Common problems |
|---|---|---|
| Stable ID | Ties every update to the same item | IDs that change on re-import |
| Title | Identifies the product | Promotional text, missing brand or model |
| Description | Factual details | Marketing copy without specifications |
| Brand | Filtering and trust | Missing or inconsistent spelling |
| Price and currency | Comparison and checkout | Differs from landing page |
| Availability | Whether it can be bought now | Not updated when stock changes |
| Product URL and images | Verification and display | Broken links, placeholder images |
| Category | Grouping and comparison | Store-specific labels only |
Category-specific attributes
Agents compare on attributes such as material, dimensions, compatibility, capacity, certifications, and care instructions. These are the fields shoppers ask about in natural language. Free-text descriptions help, but discrete key-value attributes are easier for an agent to filter and compare.
Variants
Size, color, and other options should be modeled as variants of one parent product, each with its own ID, price, availability, and image. Google Merchant Center uses item_group_id to group variants. OpenAI’s feed specification uses group_id and a variant_dict map. Flattening variants into one listing, or listing each variant as an unrelated product, makes it hard for an agent to answer “do you have it in blue, size M?”
Identifiers
The Global Trade Item Number (GTIN), the number behind UPC and EAN barcodes, lets an agent recognize the same product across retailers. That matters for price comparison and for merging reviews. Google’s specification asks for valid GTINs with correct check digits and disapproves products with incorrect ones. Where no GTIN exists, the manufacturer part number (MPN) plus brand is the usual fallback. Store identifiers as strings so leading zeros are not lost.
Policies and logistics
Shipping costs, delivery times, and return windows increasingly decide which offer an agent picks. These can be supplied in feeds and on-page as structured data.
How data reaches agents
Merchant feeds
- Google Merchant Center. Feeds Google’s Shopping Graph, which Google says contains more than 50 billion product listings, with more than 2 billion refreshed every hour. The same data powers AI Mode and Gemini shopping. In 2026 Google added optional “conversational attributes” for AI experiences, including
question_and_answer(up to 30 question-answer pairs per product),document_link(manuals and size guides),related_product,item_group_title, andvariant_option. Google describes these as optional and separate from the core specification. - OpenAI product feed specification. Used for discovery in ChatGPT under the Agentic Commerce Protocol. It accepts JSON Lines, CSV, and TSV files. Required fields include
item_id,title,description,url,brand,seller_name,image_url,availability, andprice, with optionalgtin,mpn,group_id, and eligibility flags for search and checkout. The specification asks merchants to update the feed when prices change and to omit unknown values rather than use placeholders. - Other channels. Microsoft, Perplexity, Meta, and marketplaces each accept feeds, often through platform integrations or feed management tools rather than direct submission.
On-page structured data
Schema.org markup, usually embedded as JSON-LD, describes a product on its own page using types such as Product, Offer, AggregateRating, OfferShippingDetails, and MerchantReturnPolicy. Crawlers and browsing agents can read it directly, and it gives a check against feed data. Prices and availability in markup should match the feed and the visible page.
Platform catalogs
Commerce platforms increasingly maintain a normalized catalog of their merchants and syndicate it to AI surfaces. Shopify Catalog standardizes products from millions of merchants into one taxonomy and feeds ChatGPT, Copilot, Google AI Mode, Gemini, and Meta. Developers can query it through Shopify’s Global Catalog and Storefront Catalog MCP servers. See commerce platforms.
Universal product databases and catalog APIs
A further category aggregates product data across many merchants and platforms into one searchable index for agent developers. These services crawl, ingest, and normalize listings so that an agent can search across retailers through one API instead of integrating with each merchant. Channel3 is one example: it offers an API for searching and retrieving products from a catalog it describes as more than 100 million products, with commissions on resulting sales. Google’s Shopping Graph and Shopify’s Global Catalog cover similar ground within their own ecosystems, and checkout-focused providers such as Rye also offer product data APIs. When evaluating any of them, developers typically look at category coverage, how fresh prices and stock are, variant handling, and how purchases are attributed (see for developers).
These catalog APIs work differently from the web search APIs that many AI agents use for research. Product search vs. web search compares the two.
Freshness
Data that was accurate yesterday can be wrong today. Freshness matters most for:
- Price. Sales starting and ending. OpenAI’s specification says to update the feed when a sale starts or ends.
- Availability. Agents that recommend out-of-stock items lose user trust, and agents that buy them create cancellations.
- Delivery promises. Cut-off times and carrier delays change daily.
Common approaches include scheduled full feeds plus incremental updates for price and stock, real-time inventory APIs that agents call at checkout, and checking that the landing page agrees with the feed. Protocols such as UCP and ACP confirm price and stock with the merchant during checkout, which catches errors before payment but not before a wrong recommendation.
Quality checklist
A practical audit for agent-readiness:
- Every product has a stable ID, brand, and either a valid GTIN or MPN.
- Titles follow a consistent pattern (brand, product type, key attribute, variant) without promotional text.
- Descriptions state specifications in plain language; key specs also exist as discrete attributes.
- Variants are grouped under a parent with per-variant price, stock, and image.
- Categories map to a standard taxonomy, such as Google’s product taxonomy.
- Price and availability in feeds, markup, and the live page agree.
- Shipping and return policies are structured, not only in a footer page.
- Images are real product photos of adequate resolution. Google’s specification sets a 500 x 500 pixel minimum, to be enforced from January 31, 2027.
- Common shopper questions are answered in structured FAQs where the channel supports them.
- Feeds pass each channel’s diagnostics with no unresolved disapprovals.
For how this data affects visibility in AI answers, see agent optimization. For brand-side priorities, see for brands.
Sources
- SEO starter guide — Google Search Central
- AI features and your website — Google Search Central
- Product data specification — Google Merchant Center Help
- How to use conversational attributes — Google Merchant Center Help
- Question and answer attribute — Google Merchant Center Help
- Read Sundar Pichai’s remarks at the 2026 National Retail Federation — Google
- Product feed specification — OpenAI Developers
- Agentic commerce on Shopify: how it works — Shopify
- Global Catalog MCP — Shopify Developers
- Commerce introduces data enrichment for BigCommerce and Feedonomics — GlobeNewswire
- Supported structured data attributes and values — Google Merchant Center Help
- Channel3 — Channel3
- Introducing the Rye Product Data API — Rye