LLMs create uneven city info layer, 47.5% of restaurants ignored

A new study from researchers Lin Chen, Guangyuan Weng, and Esteban Moro reveals that large language models (LLMs) act as a selective informational layer over cities, shaping which places people discover and visit. The team audited restaurant recommendations from three major LLMs across 304 neighborhoods in five U.S. cities using 320 synthetic user profiles varying in income, age, sex, and residential status. They found that LLMs both fabricate nonexistent venues and systematically overlook real ones. Fabrication is concentrated in neighborhoods with weaker digital and physical footprints, but disappears when models are provided with verified venue lists.

However, invisibility persists even with accurate lists: 47.5% of real establishments are never recommended, and 31.9% of these blind spots are shared across all three model families. This indicates the problem is not just missing knowledge but stable patterns of selective attention rooted in shared visibility patterns, not model-specific errors. The bias extends to users as well. Within identical venue pools, higher-income users receive recommendations for more expensive and less popular venues, while tourists are directed toward costlier but more socially diverse establishments than local residents.

Simulating the economic impact, the researchers found that widespread reliance on LLM recommendations would redirect visits and revenue away from chain and quick-service restaurants toward independent and full-service dining. This shift could have significant consequences for local economies and urban inequality, as LLMs amplify existing disparities in digital representation and consumer access. The findings highlight a critical need for transparency and fairness in AI-driven urban information systems.