Amazon's World Product Graph: 83 Million Searches, 97% Accuracy

By Sophie Lindqvist

Amazon's internal memo claimed about 83 million Rufus searches a week, while Amazon says Alexa for Shopping is “roughly 97% accurate globally” and calls its usage data significantly out of date.

Reuters reported on Nov. 13 that Amazon is developing a proposed World Product Graph, citing an internal document and people familiar with the matter.

What would Amazon build?

The proposal would collect product information from around the world, including items Amazon does not sell, while brands and retailers contribute data and Amazon crawls rival websites to fill gaps.

Amazon wants cleaner catalog records. It also wants better Rufus answers. The wider goal is to counter shopping agents that direct customers to competing stores.

Coca-Cola and Bayer have signed up to contribute product data. Procter & Gamble attended a trial summit. Those steps show interest. They do not establish feed volume, geographic coverage, update frequency, or commercial terms.

The document gives the project an aspirational release date of Oct. 1, 2027. Reuters reported that Amazon could still modify or abandon it.

Amazon told Reuters the document was an “outdated brainstorming memo.” The company said it did not necessarily reflect current plans. A date in that document establishes no launch schedule, development budget, or rollout capacity.

Can richer data fix Rufus?

The memo says many Rufus answers are factually wrong or incomplete. The memo put weekly searches at about 83 million. Answering those queries carries a high cost. According to the document, richer product data could let Amazon answer the same questions with less expensive models, potentially lowering the cost of serving that weekly workload.

Amazon disputes that portrayal. It calls the usage figures significantly outdated. Reuters did not report a test method, error categories, or a common measurement period.

The two claims measure different things. Search volume shows workload. Accuracy indicates answer quality. Neither figure proves that a new graph will reduce costs at Amazon's stated scale.

The document says a complete map of products and prices could answer almost any shopping question, including requests Amazon cannot fulfill from its own inventory.

Amazon renamed Rufus Alexa for Shopping this year. The memo uses both names. That choice can blur whether its criticism applies to the current service, an earlier configuration, or both.

How far does Alexa travel?

Earlier this month, Amazon invited U.S. customers to use Alexa for Shopping by text message. Those customers could search, compare, and buy through the assistant. The source gives no customer count, conversion rate, or shopping-volume figure.

Amazon also wants to restrict outside agents on its storefront. In November 2025, the company sued Perplexity. It accused Perplexity's Comet browser agent of disguising automated sessions as ordinary Chrome traffic, conduct Amazon sought to prevent as those agents browsed its storefront.

A preliminary injunction followed in March 2026. A federal appeals court vacated that order in August. The underlying case continues.

The crawling proposal runs in the opposite direction. Amazon would collect information from other retailers' sites.

What is the commercial test?

eBay appears to have blocked third-party agents from browsing its marketplace. Meta's Muse was among them. The word “appears” matters. The reported information gives no enforcement method or technical scope.

United Airlines this week barred AI agents from booking flights. The ban covered its site and travel agencies.

Those actions show a commercial tension. Amazon wants product and price data beyond its own inventory. Other retailers want control over automated access to their stores and inventory.

The business case rests on three unproven links. The reported commitments do not settle them.

• Voluntary data must arrive at useful scale and update frequency.
• Site crawling must remain reliable despite retailer restrictions.
• Richer records must improve answers without adding factual errors.

Oct. 1, 2027 remains aspirational. Amazon's denial gives the target little certainty.

Against eBay's apparent blocking and United's booking ban, the graph could give Amazon a data advantage it currently lacks, but only if retailers permit crawling or supply feeds. If Amazon advances, richer data could lower model costs and extend Alexa's reach beyond its own inventory. Any launch must then prove that better records produce fewer errors, not merely more searchable catalog entries.