EN AR

Amazon plans global product database to blunt AI shopping agents

Tamara khadrah

Key Points

  1. Reuters reports Amazon is building World Product Graph, a database covering nearly every product sold globally.
  2. An internal memo cites inaccurate answers from Amazon's own AI search tool and the high cost of running it.
  3. People familiar with the plan say the database could keep AI shopping agents tied to Amazon and protect ad revenue.

The latest:

Amazon is working on an initiative to collect data on virtually every item for sale worldwide, including products it does not stock, Reuters reported after reviewing an internal document. Called World Product Graph, the effort aims to stop shoppers leaving the site to research purchases on Google, brand pages and review forums. Amazon said draft documents often paint an incomplete picture of its current plans.

Details:

  • The document: The plan is laid out in what Amazon calls a PR/FAQ, a roughly six-page memo written as a press release setting out the rationale and business case for a new product. It carries an aspirational release date of October 1, 2027. People familiar with the matter said the project could still be modified or abandoned.
  • How it works: According to the document, the system depends on discovering how brands organize products on their own websites. Some brands and retailers would feed data to Amazon continuously, while Amazon would independently crawl other retailers’ sites. The document does not set out the crawling method or whether permission would be sought.
  • Partners signed: Coca-Cola and Bayer are among companies already contributing product data, the document says, and Amazon held a CEO-level summit with Procter & Gamble for a trial. The companies did not respond to requests for comment. The initiative sits inside Amazon’s Selection and Catalog Systems unit under vice president Robert Tekiela, who also did not respond.
  • The AI problem: The memo says many answers from Rufus, Amazon’s AI shopping search tool renamed Alexa for Shopping in May, are factually wrong or incomplete, and that customers lack the expertise to detect the errors. It puts weekly Rufus searches at about 83 million and calls running them expensive. Richer data would allow cheaper models.
  • Amazon’s response: In a statement to Reuters, Amazon said the usage figures in the memo are significantly outdated and that its own data shows Alexa for Shopping is roughly 97% accurate globally. The company said it constantly works on new ideas to make its catalog comprehensive, accurate and helpful.
  • Search share: A 2025 Morgan Stanley report found 16% of online shoppers begin product research on Amazon, against 57% on Google. When buyers already know the item they want, Amazon leads with 41% of searches starting there, compared with 23% for Google and 18% for Walmart.
  • Agent rivalry: Meta’s Muse agent has drawn more than 10 million users in the United States and Canada since launching in September. Amazon blocked Muse from its retail site that same month and last year sued AI firm Perplexity to prevent similar access, a case that remains open. JPMorgan called Muse potentially the most widely used consumer AI application since ChatGPT.
  • Jassy’s line: On the third-quarter earnings call, chief executive Andy Jassy played down the impact of shopping agents, saying few customers shop that way yet and that advertising has been minimally affected. He said Alexa for Shopping has been used by more than 400 million people over the past year. Executives privately remain wary, people familiar said.
  • Musk reaction: Responding last month on X to a post saying xAI’s Grok agent had been blocked on Amazon, Elon Musk wrote that perhaps a competitor to Amazon should be started. Employees describe World Product Graph as an extension of existing Buy for Me and Shop Direct services, which list goods sold elsewhere with Amazon as intermediary.
  • The stock: Amazon shares were down just over 1% at $264.12 at midday Friday, part of a wider selloff that followed the resignation this week of Nvidia chief executive Jensen Huang.

Background:

Amazon already lists products sold on other sites through Buy for Me and Shop Direct, acting as an intermediary. World Product Graph would widen that logic from transactions to information, covering goods Amazon itself does not sell.

Between the lines:

The commercial logic runs through the search-share gap: Amazon wins when a buyer knows the item, Google wins when they are still researching. A database covering unstocked goods targets exactly that research stage. People familiar with the matter said the same data would make Amazon unavoidable for third-party shopping agents, which matters because Amazon has already blocked Muse and sued Perplexity rather than let agents route traffic elsewhere.

What’s next

Watch whether the October 1, 2027 target survives internal review, how the Perplexity suit proceeds, and whether brands beyond Coca-Cola, Bayer and Procter and Gamble agree to supply catalog data.

What to read next