Amazon Admits AI Shopping Agent Is Inaccurate, Plans Global Product Database

Amazon's internal memo admits its AI agent Rufus gives wrong answers. Now they want to scrape every product on Earth to fix it.

Business · Source: Reuters

What happened

Amazon has a secret plan called World Product Graph. The goal is to collect data on virtually every item for sale on Earth. This includes products Amazon does not even stock. An internal document reveals Amazon wants to stop shoppers from leaving its site to research products on Google or brand websites. They want to crush Google's dominance in search and fend off Walmart.

The internal memo admits Amazon's own AI search agent is failing. The document states that many answers from Rufus are factually wrong or incomplete. It also notes that customers lack the expertise to detect these errors. Rufus reportedly handles 83 million searches a week. Running these queries is highly expensive for Amazon.

Amazon claims the leaked document is an outdated draft. The company says its AI is 97 percent accurate globally. But the data grab is already underway. Coca-Cola and Bayer have agreed to contribute product data. Amazon even held a trial summit with Procter and Gamble. The ultimate goal is an October 2027 launch to build a moat against rival AI shopping agents.

Key facts

Why it matters

Data is the only moat left in the AI era. Amazon realizes that external AI agents like Meta's Muse are a massive threat. Muse already has 10 million users. If shoppers use external agents to find products, Amazon loses its lucrative advertising revenue. By indexing the entire internet's product catalog, Amazon forces all AI shopping agents to rely on its ecosystem. They are building a walled garden for global commerce.

The cost of running AI search is bleeding big tech companies. Amazon explicitly noted that having more structured data allows them to use cheaper AI models. Better data means less compute overhead. This is a massive shift from relying purely on massive language models. It proves that clean, structured data is more valuable than raw compute power.

For builders

Proprietary data beats large models

Amazon is building a massive data graph to run cheaper models. If you are building AI tools, stop paying for massive compute. Build a proprietary data moat so you can use smaller, cheaper models to get better accuracy. The company with the best data wins, not the biggest model.

AI agents threaten search monopolies

Amazon blocked Meta's Muse and sued Perplexity because AI agents are stealing top of funnel product searches. If you build consumer AI agents, expect aggressive blocking and lawsuits from incumbents. You must find creative ways to access data when platforms shut off access.

Brand trust is your leverage

Amazon wants to own all product relationships and become the default research hub. Brands that hand over their data to Amazon risk losing direct contact with buyers. Founders building e-commerce tools should help brands own their customer relationships before Amazon intercepts every transaction.

My take

Amazon is terrified of AI agents stealing their traffic. They admit their own AI is hallucinating, so their brute force fix is to monopolize all product data on Earth. Never trust an AI shopping agent blindly. Always verify the facts on the brand's actual website before you buy.

Original reporting: Reuters. This is my rewrite and opinion.

More AI news for builders