What Does an AI Assistant Actually Know About a Business?

A business can have an accurate website and still be described wrongly elsewhere. Its own site can contradict itself. Directory records can conflict. Old descriptions stay visible to search engines and AI assistants. The question is not whether an assistant sounds confident. It is whether its claims trace to current, authoritative evidence.

What our September study shows

In September 2026 Lehvel studied four businesses: 25 AI answers and 135 factual claims. Of those claims, 69 matched the business's recorded facts, 2 contradicted them and 64 could not be checked. The three groups add up to 135. The result is narrower than a headline like “half of AI answers are wrong.” Unknown means there was no recorded fact to check against, not that the claim was false. Four businesses cannot set a rate for the population.

The same study found the business named in 13 of 13 answers when asked by name, and in six of 12 when asked for the best provider of its service. That describes this set of questions. It does not show a general preference by any assistant, or an effect of optimizing.

Why the distinction matters

Four questions are easy to mix up. Did the assistant mention the right business? Did it describe it accurately? Could a reader verify the description? Could someone safely act on it? An assistant can pass the first and fail the rest, for example by finding a seller and crediting it with manufacturing or installation it does not offer.

The unit of analysis is the claim. Each one needs a reference fact, a source, a date, the assistant and question that produced it, and a rule for sorting it as supported, conflicting or unknown. The rules are set before anyone reads the answers. Otherwise reviewers drift toward the result they hoped for.

Why discovery is changing

Pew Research Center tracked the browsing of 900 U.S. adults in March 2025 and analyzed 68,879 Google searches. People clicked a traditional result in 8% of visits with an AI summary, against 15% without one. They clicked a link inside the summary in 1% of visits that showed one. These are observations from one search environment. They do not show that any business lost a given amount of traffic. Source: Pew Research Center, July 22, 2025.

With fewer clicks, what an answer says about a business matters even when nobody visits. That does not show that a citation causes a sale, or that any tactic works.

How to check your own business

Start with the business's approved identity, offerings, roles, locations and restrictions, and who has authority over each. Collect the public versions: the website and the directory records. Then put a fixed set of questions to named AI systems and sort every factual claim in the answers against the dated record.

Keep four outcomes apart: being found, being accurate, being verifiable and being safe to act on. Keep the questions, dates and exclusions. Repeat the check over time instead of replacing the first snapshot.

Limits

An assistant may answer from sources that are incomplete but reasonable. A business's own records may be wrong. Different questions can fairly produce different descriptions, and not every unknown claim needs action. A better source record is necessary for checking. It does not guarantee that outside models will use it.

Method and sources

The study counts come from Lehvel's own research record and should be recomputed independently before anyone treats them as audited. The Pew figures follow its published methodology. This article claims no representative sample and no measured business outcome.