Reverse Engineering How Meta's Muse Shops

Meta has introduced its new AI agent, Muse, which acts as an advanced virtual assistant capable of more than just answering queries. Each Muse instance operates on a dedicated Linux virtual machine, complete with a web browser and functionalities to run commands and manage files. Launched in early September 2026, Muse has quickly gained traction, reaching over 2.8 million installs in just 12 days, outpacing the early adoption of ChatGPT. One of Muse's notable features is its shopping capability, allowing users to search for products like clothing and kitchen items, wherein it generates product cards from an internal catalog.

Recent investigations into Muse's shopping algorithm revealed crucial insights into how it ranks and displays products. The AI relies on a proprietary catalog system, accessing and processing product data without using Google’s ranking system for direct searches. Instead, it employs internal rankings, which include factors like seller quality, to determine which items are displayed to users. Researchers noted that minor linguistic changes in queries could drastically alter the products returned, suggesting Muse has a nuanced understanding of context and brand recognition. While the exact ranking formula remains confidential on Meta's servers, these findings underscore Muse's potential impact on e-commerce and the wider AI/ML community, presenting new challenges and opportunities in product discovery and digital shopping experiences.