Large language models create an uneven informational layer over cities

When people use AI like ChatGPT to find restaurants, the AI doesn't just make mistakes. It also systematically ignores many real places. It treats different people differently based on their income or age. This could change how much money local businesses make by pushing people away from fast food and toward independent restaurants.

As large language models (LLMs) become integrated into daily life, they act as informational layers. They guide how we navigate urban spaces. Previously, we relied on search engines and review platforms. While these legacy systems have biases, they offer transparency. A lower-ranked business on a search results page is still discoverable. In contrast, an LLM provides a curated list. This list hides the boundary between what is included and what has been omitted. This creates a fundamental question. Do LLMs act as neutral mirrors of urban reality, or do they act as selective filters that reshape the economic landscape of our cities?

The failure of digital footprints

The transition from searchable databases to generative interfaces introduces two distinct types of error. These are epistemic and allocative failures. Epistemic failure, or "grounding failure," occurs when a model hallucinates. This means it fabricates the existence of a venue that does not exist in the physical world. Allocative failure, or "coverage deficit," occurs when the model possesses accurate information but chooses not to surface it. This makes real businesses invisible to the user.

The authors of this study argue that current understandings of LLM bias are insufficient. They do not account for how these errors map onto existing socioeconomic inequalities. In the open-ended setting—where the model suggests restaurants without a provided list—the authors report a hallucination rate averaging 36.8% [Figure 1d]. This error is not random. The paper finds that fabrication is concentrated in neighborhoods with weaker digital footprints. These include areas with fewer Yelp reviews or lower population density. Essentially, when a neighborhood lacks a robust online presence, the model fills the void with fiction.

Auditing the urban information layer

To untangle these failures, the researchers designed a dual-track experiment. They studied five major U.S. cities using three distinct model families: GPT-4o-mini, Llama-3.3-70b-instruct, and Gemini-2.0-flash. They used a factorial design to create 320 synthetic user profiles per neighborhood. These profiles varied by age, income, sex, and residential status. This allowed them to simulate how different people experience the same urban space.

The methodology relied on two specific experimental configurations: 1. Open-ended queries: The models suggested ten restaurants within a 5-km radius with no external data. 2. Candidate-constrained queries: The models chose ten restaurants from a verified list of 100 real venues.

By comparing these settings, the authors isolated the cause of "invisibility." They wanted to know if it was a lack of knowledge or active avoidance. They employed geo-semantic matching. This is a process using text embeddings (mathematical representations of word meanings) and geographic distance to verify matches.

Persistent blindness and demographic steering

The results reveal a striking asymmetry between truthfulness and coverage. Providing verified local data effectively eliminates hallucinations [Figure 2a]. However, it does almost nothing to solve the problem of selective attention [Figure 2b]. Even when forced to choose from real venues, the "Neighborhood Algorithmic Invisibility Rate" (NAIR) remains high. On average, 47.5% of establishments are never recommended [Figure 2b]. Most importantly, 31.9% of these "blind spots" are shared across all three model families [Figure 4a]. This suggests the bias is a systemic issue in the digital ecosystem rather than a single model error.

The selectivity is even more granular regarding individual users. The study finds that LLMs perform "demographic steering" .

The model infers a user's lifestyle from their profile and adjusts recommendations. For instance, the authors report that higher-income users are steered toward venues that are more expensive and less popular [Figure 3a]. Conversely, tourists are directed toward pricier but more socially diverse establishments than local residents [Figure 3d]. This suggests LLMs do not just find what is available. They attempt to predict what a person "should" want based on their demographic profile.

Economic redistribution and digital echoes

These selective filters have real-world economic implications. The authors performed a counterfactual simulation. They imagined a scenario where 10% of dining decisions follow LLM recommendations. They report that this would trigger a massive redistribution of revenue .

The models show a structural preference for independent, full-service dining over quick-service chains. Under this simulation, independent restaurants could gain an average of $12,101 per store per year. Meanwhile, major chains like McDonald's could lose upwards of $41,405 per store [Figure 4f].

This creates a potential feedback loop. LLMs prioritize venues that already have high visitation volumes and significant online discussion [Figure 4b]. This naturally amplifies existing popularity hierarchies. Successful businesses gain more visibility. Smaller or less digitally active establishments fall deeper into algorithmic invisibility. This could create a "rich get richer" dynamic in urban commerce. The informational layer could eventually dictate the physical survival of businesses.

Assessing the informational layer

The findings of this study are limited by their scope. The research focuses only on the restaurant sector in five major U.S. cities. The authors note that these patterns might differ in smaller cities or non-English speaking contexts. Furthermore, the study does not explore how multi-turn conversations might deepen these biases.

However, the core conclusion is significant. LLMs are becoming a new form of urban infrastructure. Just as highways determine which parts of a city are accessible, an LLM determines which parts are visible. For practitioners building recommendation engines or urban planners, the takeaway is clear. Providing more data to a model may fix factual errors. It will not necessarily fix the tendency to favor the prominent over the local. Code to reproduce these findings is reportedly available; see the paper for the canonical link.

Figures from the paper