Local businesses get uneven visibility in AI: 47.5% of venues are never recommended
Models both invent venues that do not exist and systematically overlook real ones — and visibility tracks digital footprint more than quality.
Key numbers
- Audit scale: three major models across 304 neighbourhoods in five US cities, using 320 synthetic user profiles, evaluating restaurant recommendations.
- Models fabricated venues and systematically missed real ones; fabrication concentrated in neighbourhoods with weaker digital and physical footprints and disappeared once models were given verified venue lists.
- Invisibility persists even with real options: choosing from a fixed set of real venues, 47.5% were never recommended, and 31.9% of those blind spots were shared across all three model families — selective attention rather than isolated model error.
- Selectivity extends to users: within identical venue pools, higher-income users received more expensive and less popular venues, and tourists were directed to costlier but more socially diverse places than residents.
Method and sample
arXiv paper “Large language models create an uneven informational layer over cities” (2026-07-07): 320 synthetic user profiles used to audit restaurant recommendations from three major models across 304 neighbourhoods in five US cities, measuring fabrication, never-recommended rates, cross-model blind-spot overlap and profile-level differences.
What it means for brands
Whether a local business surfaces in AI has less to do with how good it is than with its digital footprint: with thin information, no structured data and few third-party reviews, the model either invents something or says nothing.
How SourceSight handles it
For local projects SourceSight asks the questions locals actually ask (nearby, directions, opening hours) and reports mentions and position per engine; the off-site list prioritises maps, review platforms and local directories before the website.
Limits
The sample covers US restaurants, so it does not transfer directly to other industries or countries; synthetic profiles are not real users, and the exact rates shift with model versions.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages
Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.
Industries · 2026-10-07Home services: 22% of homeowners already use AI to find a provider, and 80% of businesses don't know how to be found
The local services funnel is moving upstream: before a homeowner opens a website or calls, AI has already synthesised reviews, reputation and service descriptions into one answer.