Searchable found that 93% of London businesses in its study were described inaccurately by major large language models. The research focused on small and medium-sized enterprises and compared their results with those of larger companies.
It tested ChatGPT, Perplexity and Gemini more than 13,000 times on 165 London businesses, asking about services, contact details, staffing and other identifying facts. The responses were then checked against official LinkedIn profiles and Companies House records to determine whether the information was correct, incomplete or missing.
The results suggested a sharper problem for smaller firms. Half of SMEs received at least one false fact from an LLM, compared with 32% of large companies, a 56% higher rate of fabricated information for SMEs.
Across all prompts, 11 in 100 questions about SMEs returned false or missing information on key brand facts, compared with 7 in 100 for large companies. The study also found that LLMs were twice as likely to fabricate information about an SME, with a 5% fabrication rate versus 2% for larger brands.
Brand confusion was another issue. The findings showed that LLMs misattributed or confused SME brand names at a rate of 4%, compared with 0.7% for large companies, suggesting smaller businesses are more vulnerable when users rely on AI tools for basic discovery queries.
Discovery gaps
The weakest areas for accuracy were company size, website, founding year, phone number and services. These are among the details prospective customers often seek when deciding whether to contact a business.
Inaccurate responses included verifiably false facts, incomplete listings that covered at least 30% fewer essential services, or answers that failed to retrieve any verifiable existing information about a company.
The analysis adds to a wider debate about the reliability of consumer-facing AI tools as they become part of online search and recommendation behaviour. For smaller businesses, errors in basic facts could affect discovery when users are looking for a supplier, a contact number or confirmation that a company offers a specific service.
Chris Donnelly, co-founder of Searchable, linked the problem to the way language models are trained on public web data. He said smaller companies may be less visible to such systems because they are mentioned less often online than larger brands.
"Right now, if someone asks ChatGPT about a local company, there's a very real chance the AI either makes something up or draws a blank on key information. That's the equivalent of customers and revenue walking straight past businesses they should be connecting with.
The issue is related to how LLMs work. They're trained on publicly available web data, which often skews toward larger, more widely referenced brands. A London SME with a smaller digital footprint is less likely to register among what AI is reading and citing to its users," Donnelly said.
His comments reflect concern that AI systems may reproduce an existing imbalance in online visibility. Large businesses typically have broader media coverage, more backlinks, more directory listings and stronger digital footprints, all of which can make them easier for models and search systems to identify consistently.
At the same time, the findings suggest the gap is not fixed. Searchable argued that AI-driven discovery does not necessarily entrench incumbent brands in the same way as traditional search rankings, creating room for smaller companies to improve how they are represented.
"At the same time, AI doesn't cement big brands as favourites in the way that traditional search engines do. The playing field can be levelled more quickly for a well-optimised SME that understands how to make itself visible to these systems," Donnelly said.
How it was tested
The study covered 165 London businesses and generated 13,365 AI responses. Questions covered location, business size, services, specialties, website, founding year, phone number and other identifying facts.
A response was classed as inaccurate if it contained false information, omitted a significant share of essential services or failed to return verifiable information that existed in public records. That means the research measured not only outright hallucinations but also cases where AI systems appeared unable to surface established facts about a business.
The results indicate that while AI assistants are becoming a common route for gathering information, their performance on local business discovery remains inconsistent. For SMEs that depend on accurate digital visibility, the data suggests a missing phone number, a wrong service description or a mistaken identity can still appear when a potential customer asks an AI a simple question.