The City Guide Becomes the Visibility Layer
A July 2026 arXiv paper audits restaurant recommendations from three language-model families across five U.S. cities. The useful warning is concrete: when a model answers "where should I go?", it can allocate local visibility before the person ever sees a map.
The Paper
The paper is Lin Chen, Guangyuan Weng, and Esteban Moro's Large language models create an uneven informational layer over cities, arXiv:2607.06260 [cs.CY, cs.SI]. The arXiv API lists version 1 as submitted on July 7, 2026, and the PDF metadata reports a 27-page paper. The arXiv HTML lists Network Science Institute and Khoury College of Computer Sciences affiliations at Northeastern University and a CC BY 4.0 license.
The paper belongs beside this site's work on smart-city refusal, urban intelligence, AI search and answer engines, recommender systems, and answer-engine reputation. Its fresh angle is not the city as dashboard. It is the city as a generated shortlist.
The Spiralist point is that local discovery is governance. A restaurant that does not appear in a model's answer is not formally banned, ranked down, or delisted. It is simply absent from the first version of the city that the user is handed.
What They Audited
Chen, Weng, and Moro query three model families: GPT-4o-mini, Llama-3.3-70b-instruct, and Gemini-2.0-flash, accessed through the OpenRouter API. The audit covers restaurant recommendations in 304 neighborhoods across Boston, New York, Chicago, Houston, and San Francisco.
Each query is paired with a synthetic user profile crossing four demographic dimensions: eight age levels, five household-income levels, two sex labels, and four residential-status labels from local resident to tourist. That creates 320 profiles per neighborhood. The authors match model outputs to verified venue records and characterize venues and neighborhoods with SafeGraph, Spectus, Yelp, and American Community Survey data.
The design matters because it separates two problems. In the open-ended setting, the model can invent places or choose from whatever it appears to know. In the candidate-constrained setting, it receives a list of real local venues and must select from that list. If grounding fixes the problem, hallucination should disappear and visibility should become broad. The paper finds only the first half.
Fabrication and Invisibility
In open-ended recommendations, the paper reports an average hallucination rate of 36.8 percent across cities and models. The authors find that fabrication is not evenly distributed. It is associated with weaker public information footprints, including fewer real-world visitations and fewer Yelp reviews.
When the models are constrained to verified venue candidates, hallucination falls to zero in the reported experiment. But invisibility remains. Even in the candidate-constrained setting, 47.5 percent of establishments are never recommended across all demographic profiles, and 31.9 percent of the overlooked venues are shared across all three model families.
This is the important distinction. Better local data can remove invented venues. It does not automatically make the model distribute attention across real ones. The paper calls that an allocative failure rather than only an epistemic failure: the model is not merely missing facts; it is selecting from known facts in patterned ways.
The Personalized City
The same selectivity appears across users. Within identical venue pools, the paper reports that higher-income users receive more expensive and less popular venues, while tourists are directed toward costlier and more socially mixed establishments than local residents. The model is not only answering a place query. It is projecting expectations from demographic and contextual cues into the user's choice set.
The authors also simulate potential demand shifts. Their analysis suggests that broad reliance on LLM recommendations would redirect visits and revenue away from chain and quick-service restaurants toward independent and full-service dining. That effect is not automatically good or bad. Independent restaurants may gain visibility; affordable or habitual venues may lose mediated access. The governance issue is that the shift would be produced by an opaque answer layer rather than by an accountable urban policy.
The paper's limits are material. It studies restaurants, five large U.S. cities, three model families, synthetic profiles, and a controlled prompt design. It does not prove the same pattern for clinics, libraries, repair shops, schools, smaller cities, non-English contexts, or every deployed assistant. Some underlying commercial datasets are not fully public because of provider restrictions. The paper is strongest as an audit of a mechanism: generated local guidance can turn recommendation into spatial allocation.
Governance Reading
A search result page lets the user scroll past the first answer and see that alternatives exist. A conversational recommendation often hides the boundary of omission. It sounds like help, but it can function as a gate. The missing restaurant is not shown as "rank 34." It is not shown at all.
For city agencies, chambers of commerce, tourism boards, delivery platforms, and model providers, that changes the audit target. The question is not only whether a model names real places. It is whether the system can explain which real places it did not name, whether omissions cluster by neighborhood or user profile, and whether the model amplifies existing digital-footprint advantages.
The Receipt
An urban-visibility receipt should name the model, provider, date, prompt, system message, location, radius, candidate set, data source, venue taxonomy, matching rule, hallucination rate, omitted real venues, shared blind spots across models, user-profile fields, demographic effects, visit or revenue simulation assumptions, unavailable datasets, and appeal path for affected businesses.
The practical rule is simple: do not call a model a local guide until it can show the city it left out. A useful urban assistant should help people navigate place without quietly turning digital prominence into the geography of possibility.
Sources
- Lin Chen, Guangyuan Weng, and Esteban Moro, Large language models create an uneven informational layer over cities, arXiv:2607.06260 [cs.CY, cs.SI], submitted July 7, 2026.
- arXiv experimental HTML for arXiv:2607.06260v1, checked for affiliations, license, abstract, methods, prompts, results, discussion, limitations, data availability, and supplementary notes.
- arXiv API record for arXiv:2607.06260, checked for title, authors, subject classes, submission date, and version metadata.
- arXiv PDF for arXiv:2607.06260, checked as the 27-page PDF source and for PDF metadata.