AI shopping: how engines pick the products they recommend

Most advice about AI search is written for informational questions. Product questions work differently. Ask an engine which running shoe suits a wide foot and the reply is not ten links to compare. It is a short set of named products, usually carrying a price, a stock status, and increasingly a checkout button. The unit of selection is a product rather than a page, and the data no longer comes only from the crawl. Google's AI features guide states that "Using products like Merchant Center (such as Merchant Center feeds) and Google Business Profiles can help your products and services to be visible in both AI responses and other Google Search results." Those inputs are covered at concept level in local and merchant data in AI answers. This page goes down to the fields.

The structured inputs, company by company

Google

Google documents two inputs. The product data specification makes these attributes required: id, title, description, link, image_link, availability and price. Three more are conditional: availability_date for preorder and backorder items, brand for new products outside movies, books and music, and mpn when there is no manufacturer-assigned GTIN. The spec ties the feed to the page: the title must "match the title from your landing page", and the price must "match with the price from your landing page".

On the page itself, merchant listing structured data requires only name, image and offers, with price and priceCurrency inside the offer. The detail sits in the recommended list: availability, brand.name, gtin, sku, itemCondition, shippingDetails, hasMerchantReturnPolicy, isVariantOf, review and aggregateRating. Google's product documentation explains why both exist: "Providing both structured data on web pages and a Merchant Center feed maximizes your eligibility to experiences and helps Google correctly understand and verify your data." Merchant listings require a price greater than zero and apply only to pages where a shopper can buy the item.

This does not contradict Google's statement that structured data is not required for generative AI features. Product markup earns product experiences, not AI placement. The general case is in structured data.

OpenAI

OpenAI publishes a product feed specification for ChatGPT. Required fields are item_id, title, description, url, brand, image_url, price, availability, seller_name, seller_url, target_countries and store_country, plus two switches the merchant sets: is_eligible_search and is_eligible_checkout. Discovery and purchase are separate decisions in the same file. Recommended fields include gtin or mpn, size, color, group_id, listing_has_variations, sale_price, reviews and q_and_a. The key concepts guide describes onboarding as an initial sample feed for validation followed by daily snapshots. How ingested products are ranked against each other is not documented.

Microsoft

Microsoft documents the program rather than the format. Its agentic commerce page says Copilot Checkout "lets shoppers complete purchases directly inside Microsoft Copilot", and states that "Only English-language merchants who sell to US buyers are eligible at this time (supporting USD)." A post dated January 8, 2026 states that "product feeds from MMC help inform organic Copilot results" and that Merchant Center "is not required to sell through Copilot Checkout". There is no field-by-field ingestion spec for Copilot answers.

Review data

Reviews are recommended input on both sides. Google's review snippet documentation requires marked-up review content to be "readily available to users from the marked-up page", forbids fake or undisclosed incentivized reviews, states "Don't aggregate reviews or ratings from other websites", and makes a page ineligible for the star review feature where "the entity that's being reviewed controls the reviews about itself".

Agentic checkout, as documented in August 2026

An agent that completes a purchase needs machine-readable availability, unambiguous variant identity and a checkout it can call. Three efforts are published, months apart.

• Agentic Commerce Protocol. Its site states that "Stripe and OpenAI developed the Agentic Commerce Protocol to define a common language for how agents and businesses transact" and that it is open source under the Apache 2.0 license. OpenAI splits it into a product feed, a checkout spec merchants implement, and a delegated payment spec, and documents that merchant systems, not OpenAI's, handle validation, tax, payment and the decision to accept or decline an order.
• Universal Commerce Protocol. Google announced UCP on January 11, 2026 as "a new open standard for agentic commerce", co-developed with Shopify, Etsy, Wayfair, Target and Walmart. Google's Merchant Center documentation says it powers a checkout button on eligible listings in AI Mode and in Gemini, that the merchant "will remain the seller of record", that a native_commerce(checkout_eligibility) attribute controls the button, and that eligibility covers the United States, Canada and Australia for "select merchants at this time".
• Copilot Checkout, on the terms quoted above.

These are separate specifications with overlapping goals, all in staged rollouts to selected merchants and markets. Anything written about agentic checkout dates quickly, including this section.

What a merchant controls

1. Feed accuracy. Both feeds carry price and availability, and Google's spec asks them to match the landing page. A feed that disagrees with the page gives an engine two answers.
2. Markup that matches the page. Google's structured data policies state "Don't mark up content that is not visible to readers of the page" and "Your structured data must be a true representation of the page content."
3. Honest availability and pricing. In stock, out of stock, preorder and backorder are defined values with defined consequences.
4. Review integrity, on the rules above.
5. Content that answers the buying question. The feed already carries the spec sheet. What it cannot carry is the comparison, the fit, and the reason a buyer should pick one variant over another.

What a merchant does not control

Which products an engine names is not a setting. Google describes its AI features as using query fan-out and grounding in the core Search ranking systems, and none of the three companies publishes a ranking specification for product selection. Google also warns: "Be wary of third-party tools that promise ranking success or claim to use 'internal' Google metrics. No third-party tool has access to our internal ranking or AI systems." That applies to product placement as to citations.

Measurement

Track product and category prompts as their own sets, separate from informational prompts: different phrasings, different competitors, different seasonality. The method is in designing a prompt set, and the sampling discipline behind comparable runs is in measuring AI search visibility honestly.