AI Research Division · AI Visibility
Hallucination Rate by Model: A Local-Business Benchmark
Across 1,000 controlled prompts about real Houston businesses, we measure how often each frontier model fabricates verifiable facts — locations, hours, certifications, ownership.
By Shayne Beavan
Founder, Deep AI Solutions · Inventor of record, 5 USPTO filings
7 min
Setup
We constructed 1,000 prompts asking each of the five frontier models specific, verifiable facts about real Houston businesses — locations, hours, services, certifications, ownership, founding year. We then verified every claim against the business's own public surfaces.
Hallucination rates
| Model | Hallucination rate |
|---|---|
| Claude | 1.9% |
| Perplexity | 2.4% |
| ChatGPT | 3.6% |
| Gemini | 4.1% |
| Grok | 5.8% |
Common failure modes
- Inventing additional locations for single-location practices.
- Conflating two similarly-named businesses.
- Asserting certifications, awards, or affiliations that the business does not hold.
- Fabricating extended hours or 24/7 service when none exists.
Defensive posture
Hallucination is reduced — not eliminated — by publishing complete, consistent, machine-readable facts. The model fabricates least when the truth is unambiguous and easy to retrieve.