When an AI agent shops for someone, it can only buy what it can read: the names, prices, stock and specs behind every listing. In our reading, that product data is the contested layer of AI shopping. Amazon, Google and OpenAI each hold a different piece of it. Google says its Shopping Graph holds more than 50 billion product listings. A 2025 Morgan Stanley report, as Reuters cited it, found 57% of online shoppers start product research on Google against 16% on Amazon.
This explainer maps the layers between a brand’s website and an agent’s checkout, who holds each one and how money moves through it. The latest test came on 13 November, when Reuters reported an internal Amazon plan to gather data on virtually every item for sale, which we covered in Amazon’s plan against AI shopping agents. On Forter’s order data, Meta’s Muse agent bought at 74% of stores with any agentic order.
Business Pill · Agentic Commerce Explained
An evergreen Business Engineer explainer on agentic commerce, when the shopper is an AI agent. It teaches the concept and does not discuss the companies in this explainer.
The key insight: Six layers sit between a brand’s product data and an AI agent’s checkout: Amazon charges referral fees on its catalog, Google says its Shopping Graph holds more than 50 billion listings, and OpenAI says merchants pay a small fee on completed purchases.
Read from the bottom up, there are six layers. First, brands and merchants hold the original data on their own websites and feeds. Second, marketplaces such as Amazon turn it into a catalog they control. Third, search engines such as Google merge listings from many sellers into a shopping graph.
Fourth, agent checkout protocols, such as OpenAI’s Agentic Commerce Protocol and Google’s Agent Payments Protocol (AP2), carry an order from the agent to the merchant. Fifth, payment credentials from Visa, Mastercard and crypto rails such as x402 let the agent pay. Sixth, the agent itself, ChatGPT, Gemini, Muse or Amazon’s Rufus, talks to the shopper. The layer names are our own framing; each owner below is named in its own announcement or in reporting we cite.
The six layers between a brand’s product data and an AI agent’s checkout, who holds each and how money moves. Layer names are our framing; figures from Google (May 2025), Amazon’s seller pricing page, OpenAI, and Reuters citing Morgan Stanley.
Layer 1 and 2: Brand Data and the Marketplace Catalog
Brands still hold the source. The Amazon document Reuters reviewed says its planned World Product Graph “depends on discovering how brands organize their products on their own websites”. Some brands and retailers would contribute data continuously, and Amazon would also crawl other retailers’ websites, Reuters reported. The document says Coca-Cola and Bayer, among others, have signed up to contribute product data.
Amazon’s seller pricing page lists two plans, $0.99 per item sold or $39.99 a month, plus referral fees that vary by category: a percentage of the total price or a minimum amount, whichever is greater. The page lists 8% for computers, 15% for footwear and eyewear, and a $0.30 minimum in those categories.
The internal Amazon document Reuters reviewed explains why Amazon wants more than its own listings: “Amazon cannot be the go-destination for every shopping mission if customers must leave to research product relationships on Google, brand websites, and review sites”. On Morgan Stanley’s 2025 figures, as Reuters cited them, 41% of searches start on Amazon when shoppers know exactly what they plan to buy, against 23% for Google and 18% for Walmart.
Layer 3: The Search Engine’s Shopping Graph
In May 2025 Google said its Shopping Graph “now has more than 50 billion product listings”, with details like reviews, prices, colour options and availability, and that more than 2 billion of those listings are refreshed every hour. That is about 4% of the graph an hour, or the equivalent of the whole graph every 25 hours, by our arithmetic.
Google’s same post introduced an agentic checkout in AI Mode: a shopper taps track price, sets a size, colour and budget, and can confirm a buy for me purchase when the price drops. On the Morgan Stanley figures Reuters cited, 57% of shoppers start research on Google, about 3.6 times Amazon’s 16% by our arithmetic.
Layer 4 and 5: Protocols and Payment Credentials
OpenAI’s Instant Checkout runs on the Agentic Commerce Protocol, which it co-developed with Stripe and open-sourced. OpenAI says “Product results are organic and unsponsored, ranked purely on relevance to the user.” It also says “Merchants pay a small fee on completed purchases”, that the merchant stays the merchant of record, and that ChatGPT considers availability, price, quality and whether Instant Checkout is enabled when ranking merchants that sell the same product.
Google announced AP2 on 16 September 2025 with more than 60 organizations, including Adyen, American Express, Coinbase, Etsy, Forter, Mastercard and PayPal. Its job, Google says, includes “Proving that a user gave an agent the specific authority to make a particular purchase.” AP2 uses signed mandates: an intent mandate for the request and a cart mandate for the exact items and price.
Mastercard announced Agent Pay and its Agentic Tokens on 29 April 2025, with Microsoft as an early partner; we covered its Agent Pay for machines in June. Visa launched Intelligent Commerce on 30 April 2025 with partners including Anthropic, Microsoft, OpenAI, Perplexity and Stripe. Its chief product and strategy officer, Jack Forestell, said: “These agents will need to be trusted with payments, not only by users, but by banks and sellers as well.”
For payments too small for cards, Chainalysis counted more than 138 million x402 transactions on Base, as we reported in x402 payments on Base top 138M.
Layer 6: The Agent, and Why Protocols Reach Few Stores
Agents do not have to wait for protocols. Forter’s data split agents into protocol agents, which check out through commerce protocols, and scraping agents, which drive a store’s own pages. Forter’s explanation: “They can shop anywhere, while the protocols can only reach those stores that have integrated them.” Muse bought at 74% of stores with any agentic order between 1 August and 7 October, against 24% for OpenAI’s Instant Checkout and 5% for Google, as we reported in Muse buys at 74% of stores.
Scraping agents went from 1% of agentic orders in late July to 77% in the week of 19 October, Forter says. Amazon blocked Muse from its site in September and is in court with Perplexity over similar access, Reuters reported. a16z’s agent economy map counts 33 payments and checkout logos, by our count, the largest infrastructure section.
Why the Catalog Matters, as We Read It
As we read the six layers, the money sits at different points. Amazon’s page lists referral fees on its catalog, OpenAI says merchants pay a small fee on completed purchases, and Google holds the search page where, on the Morgan Stanley figures Reuters cited, 57% of research starts. In our reading, every agent route, protocol or scraping, depends on accurate product data.
Amazon’s own document draws the same link, as Reuters reported: more complete data would let Amazon use less expensive models more efficiently, and the people familiar said the most extensive store of product data could compel agents to seek products on Amazon.com. The document also says “Many Rufus answers are factually wrong or incomplete, but customers lack expertise to detect these errors”. For the basics of the concept, see agentic commerce explained.
The Structural Read
As we read it, each company in this explainer holds a different layer of the same data. Google says more than 2 billion of its 50 billion listings are refreshed every hour, about 4% an hour by our arithmetic. Amazon’s document, as Reuters reported it, plans data on virtually every item for sale. Forter’s data shows agents that skip protocols reaching more stores, which in our reading puts the weight on whoever holds readable, current product data.
Forter, AI Agent Order Data, 2 November 2026
“They can shop anywhere, while the protocols can only reach those stores that have integrated them.”
Three Implications
FOR BRANDS Amazon’s document, as Reuters reported it, says World Product Graph depends on how brands organize products on their own websites, and Coca-Cola and Bayer have signed up to contribute data.
FOR MERCHANTS OpenAI says merchants pay a small fee on completed Instant Checkout purchases and stay the merchant of record; Amazon’s page lists referral fees such as 8% for computers and 15% for footwear.
FOR PAYMENT NETWORKS Visa and Mastercard announced agent programs in April 2025, and Google says AP2 launched with more than 60 organizations, including Mastercard, American Express and PayPal.
The framework says: “If you understand how each data pipeline becomes a moat — and how each moat decays — you understand the strategic structure of the entire AI stack.”
As we read it, product data for shopping agents is one such pipeline: Google refreshes it hourly, Amazon plans to gather it from brands and other retailers, and the protocols and card networks carry orders built on it. The framework predates this explainer and does not discuss AI shopping.
Where online shoppers start, from a 2025 Morgan Stanley research report as reported by Reuters on 13 November 2026. Google’s 57% is about 3.6 times Amazon’s 16% by our arithmetic.
What Is Not Established
OpenAI does not state the size of its merchant fee in the announcement we read, and Google’s AP2 post sets no fee. The 50 billion and 2 billion Shopping Graph figures are Google’s from May 2025; we did not find a later count. The Amazon figures come from a document Reuters reviewed and we did not see; Amazon told Reuters the usage figures are significantly outdated.
Forter’s figures cover its own network, not the whole market. Referral fees vary by category, and the ones named here are examples from Amazon’s page.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Product data passes through six layers before an AI agent can buy: brand sites, marketplace catalogs, search graphs, checkout protocols, payment credentials and the agent itself. Amazon charges referral fees on its catalog, Google says its Shopping Graph holds more than 50 billion listings, OpenAI says merchants pay a small fee on completed purchases, and on Forter’s data the agents that skipped protocols reached the most stores.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read OpenAI’s Instant Checkout and Agentic Commerce Protocol announcement, Google Cloud’s AP2 announcement, Visa’s Intelligent Commerce and Mastercard’s Agent Pay releases, Google’s May 2025 shopping post and Amazon’s seller pricing page in full on 13 November 2026. Amazon document and Morgan Stanley figures come through Reuters, agent order figures through Forter and x402 counts through Chainalysis, as reported in our earlier pieces. The six layer names and the refresh arithmetic are ours. We did not contact the companies for this explainer.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
