We ran 75 questions a buyer might ask about three real UK businesses through AI search. The businesses were named three times. Their own websites were cited as a source zero times.
And when the answers did mention them, a fair amount of what they said was wrong.
This is the part of AI visibility most businesses never check. They ask “do we show up?” and stop there. The better question is “when we show up, is it true?”
What we tested
Three businesses we audited in October 2026. We haven’t named them.
| What it does | How long trading | Its website | |
|---|---|---|---|
| Business A | Sells DIY building kits online | New, domain bought in 2026 | Built in JavaScript, so AI crawlers received two words |
| Business B | Event production and logistics, including national charity events | More than 20 years, by its own account | Clean static site with full company markup, founder included |
| Business C | Event and entertainment hire, B’s sister company | Incorporated 2005 | 583 pages in its sitemap |
For each, we ran 25 prompts on Claude web search: eight discovery questions, five comparisons, five problem-led questions, five about the business by name and two about buying safely. We logged every answer, every company named and every source cited, with a timestamp.
The headline numbers
- 75 questions asked
- 3 times a business was named, all for Business C, all when the question already had its name in it
- 0 times any of the three websites was cited as a source
- 6 distinct kinds of error
None of the three appeared for a single question that didn’t already contain its name. A buyer who didn’t know them would never have heard of them.
The six kinds of error
1. “I can’t find it”
Asked about Business A by name and by domain, the answers said they couldn’t find the product and “can’t confirm whether it exists”. Asked who founded Business B, the answer couldn’t find the company in its own town.
Not being found is bad. Being told a business might not exist, by a tool the buyer trusts, is worse.
2. Someone else, presented as you
Asked to compare Business B with other suppliers, the answer profiled a different firm with a similar name in another part of the country, base and coverage area included, as if it were the business we’d asked about. In another answer it suggested a local competitor: “It’s possible you mean this firm under a different name.”
3. The name means something else
Business A’s product name is close to a famous piece of office software. Ask the comparison question and the whole answer is about software, with no mention of buildings. One of its model names belongs to a software company too, and was read that way.
Ask whether Business C is insured and the answer is about a South Korean insurance company with the same initials. Ask for reviews of its domain and you get film review sites.
4. An old fact stated as current
One answer told the buyer that Business C’s filings “look out of date”. They don’t. Companies House shows its accounts filed and up to date, with nothing overdue. The answer had picked up an older snapshot and presented it as fact.
5. Described wrongly
Another answer called Business C “a small or newer business with little online presence”. It was incorporated in 2005. It also gave the company’s legal name slightly wrong.
6. Somebody else gets the credit
Business B runs the production and logistics for a national charity event with, by its own figures, a six-figure number of participants. Asked who runs that event’s logistics, the answer credited the charity’s in-house team. Business B’s name never came up.
That’s the most expensive error of the six. It’s the business’s best proof, and it was handed to someone else.
The best website did no better
You’d expect the technical winner to come out on top. It didn’t.
Business B has the cleanest site of the three: static HTML that every AI crawler can read in full, company markup on every page with the legal name, founder and address, and priced product pages. It was cited in none of its 25 answers and confused with a similarly named firm.
Where the answers did know something, they got it from somewhere else. Companies House, a business directory, a charity’s own pages. The businesses’ websites weren’t the source of anything.
That matches the bigger picture. Across 75,000 brands, Ahrefs found that mentions of a brand on other sites lined up with AI Overview visibility far more closely than backlinks did. It’s a correlation, not proof, but it explains this result. A good website gets you read. It doesn’t get you believed.
Why this matters more in regulated sectors
Every one of these errors has a regulated version.
- “Filings look out of date” becomes a wrong statement about registration or authorisation status.
- The wrong company becomes a firm with a similar name on the same register.
- The wrong legal name becomes a buyer checking the wrong entity.
- Lost credit becomes your track record attributed to a competitor.
In a regulated firm, a wrong claim about status or history doesn’t stay a marketing problem for long. It ends up with compliance. That’s why we treat accuracy as part of visibility, not a separate concern. If you’re working on AI visibility for a law firm or a financial services firm, the accuracy check comes first.
What a mis-statement audit is
A mis-statement audit records every factual claim AI assistants make about a business, then checks each one against the source. Founder, legal name, location, filing status, insurance, who did what work, and which company the answer is actually describing.
It runs alongside the usual visibility questions, because they answer different things. Visibility tells you whether you’re named. The mis-statement audit tells you whether what’s said is true.
If you’d like us to run both for you, start with a free AI visibility audit.
What to do this week
You don’t need us to start. Do this once:
- Ask what a stranger would ask. Your name, your domain, your founder’s name, “is [you] insured”, “[you] reviews”, and “who does [what you do] in [your area]”.
- Write down every claim and check each one. Wrong is worse than missing.
- Search your own name for collisions before you name a product, and again now.
- Put the boring facts on your site: legal name, company number, registered address and one phone number, the same everywhere.
- Fix the records the answers lean on. Companies House, your main directory listings, and any page that describes your work.
- Get your best work told by someone else. The client’s page, the case study they publish, the trade press write-up.
- Run it again in three months. The answers move.
For the step-by-step on correcting what’s already out there, see how to fix AI giving wrong information about your company.
How we ran it, and the limits
We used Claude web search only. ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews weren’t tested here, and their answers will differ.
Business A was tested on 8 October 2026, and Businesses B and C on 12 October 2026. Three businesses and 75 prompts is enough to show the kinds of error, not to say how common each one is across the UK.
AI answers change week to week, so this is what was said on those days, logged with timestamps. We haven’t named the businesses, and we’ve kept out details that would identify them.
Related
Part of a series where we publish our own findings: how long AI citation actually takes, organic at 9% of our traffic and 86% of our leads, and what UK GEO agencies actually charge.
Common questions
How often does AI search name a business when buyers ask about its category?
In our test, almost never. We ran 75 buyer questions about three UK businesses through Claude web search in October 2026. The businesses were named three times, all for the same business and only when the question already contained its name.
None of the three websites was cited as a source in any of the 75 answers.
What kinds of mistakes does AI make about businesses?
We logged six kinds: saying the business can’t be found, describing a different company as if it were the one asked about, confusing the name with something more famous, stating an old fact as current, describing the business wrongly, and giving credit for its work to someone else.
Does a well-built website stop AI getting your business wrong?
Not on its own. The best-built website of the three was cited in none of its 25 answers and was confused with a similarly named firm in another city.
The answers leaned on Companies House and directory listings, not on the businesses’ own sites.
What is a mis-statement audit?
The part of an AI visibility audit that records every factual claim AI assistants make about a business, then checks each one against the source.
It’s the check most businesses never run, because they only ask whether they appear, not whether what appears is true.
How do I check what AI says about my business?
Ask the questions a stranger would ask: your name, your domain, your founder’s name, whether you’re insured, whether you’re any good, and who does what you do in your area.
Write down every claim and check it. Then run the same questions again in three months, because the answers move.