An internal Amazon memo describes a plan to collect data on "virtually every item for sale on Earth, whether Amazon stocks it or not." Greg Bensinger reported it for Reuters on November 13, 2026. Amazon calls the plan World Product Graph. If it happens, a shopper asking Amazon's AI assistant for the best stand mixer could get an answer that includes your mixer, even if you have never listed a thing on Amazon.
The plan is only a proposal. Amazon told Reuters the memo was a brainstorm that "doesn't necessarily reflect our current plans." Even so, it shows where some people inside Amazon think shopping is going, and how Amazon would get your product information: partly from brands that send it, and partly by reading other stores' websites. Below is what the memo says, the odd position Amazon is in with outside AI agents (AI assistants from other companies that can shop and buy for people), and what to check on your own product pages now.
What the memo describes
Amazon's World Product Graph
Reuters reviewed an internal Amazon memo for a project called World Product Graph. Amazon calls this kind of memo a PR/FAQ: a mock press release followed by questions and answers, typically six pages long, written before Amazon decides whether to build something. World Product Graph would gather data on products sold anywhere, including items Amazon doesn't carry. Some brands would send their product data to Amazon on an ongoing basis, and Amazon would also crawl other retailers' websites (send a program to visit their pages and copy the product details). The memo says Coca-Cola and Bayer, among others, have signed up to contribute data, and it mentions a trial with Procter & Gamble. It gives October 1, 2027, as a hoped-for release date. Reuters could not find out whether or when the plan will go ahead.
The project sits in Amazon's Selection & Catalog Systems unit, under vice president Robert Tekiela. In the memo, Tekiela says customers "have always been able to find products on Amazon." What's missing, he writes, is knowing how products relate to each other and which one is right for a given shopper.
Employees told Reuters they see World Product Graph as the start of making Amazon a one-stop shop for everything. It would build on two features Amazon already runs, Shop Direct and Buy for Me. Reuters says both show products from other websites inside Amazon. With Buy for Me, Amazon's AI buys the item for the shopper from the seller's own website.
Why Amazon wants it, according to the memo
Rufus is Amazon's AI shopping assistant, which Reuters notes was renamed Alexa for Shopping in May. The memo is blunt about Rufus's mistakes: "Many Rufus answers are factually wrong or incomplete, but customers lack expertise to detect these errors." The memo says people run about 83 million Rufus searches a week, and that answering them is expensive. Its argument is that better product data would let cheaper AI models (the programs that write the answers) give more accurate answers.
Amazon disputes part of that. It told Reuters the usage figures in the memo are "significantly outdated," and that its own data shows Alexa for Shopping is "roughly 97% accurate globally." Amazon CEO Andy Jassy has said more than 400 million people used Alexa for Shopping in the past year, according to Reuters.
There's a business worry too. The memo admits Amazon can't be where shoppers go for every purchase, and it names Google, brand websites, and review sites as the places they go instead. People familiar with the plan told Reuters that if Amazon holds the most complete product data, AI agents will look for products on Amazon instead of on other retailers' sites. Shoppers would stay on Amazon, where brands pay to put their products in search results. Reuters describes that advertising business as a growing one.
Amazon would read other stores' sites but keeps outside agents off its own
World Product Graph would crawl other retailers' websites to learn about their products. We find that hard to square with how Amazon treats agents that try to shop on Amazon. Amazon blocked Meta's Muse agent (an AI assistant that can open a browser and act for its user) in September, the same month Muse launched, Reuters reports. Reuters puts Muse at more than 10 million users in the US and Canada.
Amazon also sued Perplexity, an AI search company, to stop its Comet browser (a web browser with an AI agent built in that can shop for its user) from shopping on Amazon, and won a temporary court order against it. On August 4, 2026, the Ninth Circuit (a US federal appeals court) threw that order out, Engadget reported, though the case itself goes on. A user on X (the site formerly called Twitter) said Grok, the assistant from Elon Musk's company xAI, was blocked too. Musk replied, "Maybe we should start a competitor to Amazon," Reuters notes.
On Amazon's third-quarter earnings call (its quarterly update for investors), Jassy played down outside agents: "few customers are doing much shopping this way yet." Amazon executives remain wary of agents that could send shoppers to sites with better product data, according to Reuters.
Whether Amazon keeps blocking outside agents or ends up letting them in, some of them can't reach your Amazon listings at all right now. Your own website is different: almost every agent can try to read it, Amazon's included, unless you block it. That makes your product pages a source Amazon, ChatGPT, Muse, and Perplexity can all draw on, as long as they can make sense of them. Our post on ChatGPT shopping and product feeds shows ChatGPT moving, almost overnight, from finding products on the open web to taking most of them from product feeds (files of products, prices, and stock that retailers send straight to ChatGPT).
What a product page needs if Amazon or any agent reads it
None of this is specific to Amazon. An agent comparing products, whoever built it, needs the same facts from your page, and it can only use what it can find and trust.
Put the facts in the page your server sends. Price, stock, sizes, colors, shipping cost, and return policy should all be in the page's HTML (its text and code) as your server first sends it. Your server is the computer that hosts your site. Those facts shouldn't depend on scripts (small programs that run in the visitor's browser and fill in parts of the page after it loads). If a crawler (a program that visits pages and copies what's on them) doesn't run those scripts, it sees an empty spot where your price should be. To check, open a product page, right-click, and choose View Page Source (in Safari, turn on the Develop menu first and choose Show Page Source), then search for the price.
Label them with structured data. Structured data (labels in a page's code
that state facts such as the price or whether an item is in stock) lets an
agent read a value without guessing what it means. Use the Product label for
the item and the Offer label for its price and availability. Both come from
Schema.org, a shared list of labels that Google documents and uses. How much
AI assistants rely on them is less clear, but they're the most direct way to
state a price in code. Our
structured data guide shows how.
Say which version is newest. The memo says richer data would help shoppers tell which model or color of a product is the latest. If you sell the 2025 and 2026 versions of something, say so plainly on each page, with the model year or release date in the text, and link the two pages to each other.
Decide which crawlers you let in. robots.txt (the file that tells bots which pages they may visit) is where you allow or block them by name, though it's a request a bot can ignore. Amazonbot is one of Amazon's crawlers, but Reuters couldn't find out how World Product Graph would crawl or whether it would ask permission. Check that your bot protection (a setting from your host or a security service that turns away automated visitors) isn't blocking the agents you do want. Our AI crawler controls guide lists the names.
Test whether an agent can finish a purchase. Some agents already buy straight from store websites. Amazon's Buy for Me does, and so does Perplexity's Comet. Meta's Muse and OpenAI's Dots also use a browser to go through a site the way a shopper would, and our post on them shows how. Go through your own checkout using only the keyboard (Tab to move, Enter to choose). It isn't exactly how an agent works, but it's a quick way to find the steps that only work with a mouse. Note where a pop-up, a slow-loading button, or a required account sign-up would stop you.
If you make your own products, keep one version of the facts. Coca-Cola and Bayer have already signed up to contribute data, according to the memo. Reuters doesn't say whether other brands could join. If Amazon opens it up, the brands with the least work to do would probably be those whose product names, specs, and prices already match across their own site, their retailers, and their product feeds.
References
- Reuters, Greg Bensinger, "Exclusive: Amazon's plan to own the world's product data and fight AI shopping agents", November 13, 2026
- Engadget, "Appeals court vacates the order blocking Perplexity's Comet from Amazon", August 4, 2026