Being Named Is Not the Same as Being Described Correctly

AI visibility accuracy is whether an engine describes your business correctly when it names you. Not whether it names you. Whether it gets you right.

Those are different questions, and until recently we were only measuring the first one.

What our own guide was missing

Our published guide to measuring AI search visibility lists three metrics: mention rate, citation share, and competitive share-of-voice. It is a good guide. We stand behind the method, and we still think a spreadsheet and two hours a month beats most paid tools.

But read those three back. Mention rate counts how often you are named. Citation share counts which of your pages get used as a source. Share-of-voice counts your mentions against your competitors' mentions.

All three count presence. Not one of them asks whether the sentence attached to your name was correct.

That is a gap, and it is ours, so we are going to fix it in public the same way we did last time. A few weeks ago we wrote up how we got our own AI visibility testing wrong by running queries while logged in, which made the numbers unrepeatable. This is the next thing we found when we went looking. The first problem was that our readings could not be reproduced. The second is that even a clean, reproducible reading was measuring the wrong half of the question.

How often do AI engines get UK business details wrong?

Often enough that it should be on your dashboard.

The most detailed UK dataset we have seen comes from Searchable, an AI visibility platform, published on 13 July 2026. One press release, no published methodology, and the number of retailers tested is not disclosed. They put more than 72,000 questions about UK high street retailers to ChatGPT, Google Gemini and Perplexity, then graded every answer against each retailer's verified information.

Around one answer in 16 came back false. Roughly two in three of the businesses tested had at least one false fact returned about them. The most common error was a wrong postcode, at about one answer in 10, and in around 15% of those cases the location given was more than twenty miles from the real address. The median gap between the real and the given location was just over a kilometre, so the twenty-mile cases are the tail, not the norm. About one website answer in 15 pointed at a dead link, a lookalike, or a different business altogether. On engine-by-engine accuracy, Perplexity was wrong in about 10% of cases, Gemini around 5%, ChatGPT around 4%. That ordering comes from a single vendor test and other datasets rank the engines differently, so do not build a procurement decision on it. Note also that Claude was not in the sample, so if you test four engines the way we do, this covers three of them.

Three caveats, because most of the coverage carries none of them. Searchable sells AI visibility software, so this is vendor research and should be read as directional. It is a retail sample, so a broker or a law firm should not assume the same rates apply to them. And the national dataset was published on 13 July 2026. Searchable has since released regional cuts, so it may be resurfacing in your feed as new. It is not.

What survives all three caveats is the shape of the finding. Engines are confident and they are wrong at a rate that would not be tolerated from any other channel you rely on.

Why smaller firms get this worse than big brands

Because engines describe you using other people's pages, and big brands have far more of them.

This is the part of the Searchable work that made me sit up, because it is the same argument we make in every audit. Be careful with it though. In the retail release, the small-versus-large point is a founder quote, not a published result. The numbers behind it sit in a separate Searchable study from late June: 13,000-plus prompts about 165 London businesses, checked against Companies House and official company profiles, comparing SMEs against firms of 500-plus staff. That one found 50% of SMEs got at least one fabricated fact against 32% of large companies, and brand-name confusion running at 4% for SMEs against 0.7% for large brands. It is cross-sector rather than retail, which makes it more relevant to a broker or a law firm than the high street data. It also ran as sponsored content, so the same directional-only rule applies.

Think about what that shape implies, allowing that no one has run this test on insurance or legal firms. A national insurer is being described from hundreds of independent pages. If one directory has the wrong phone number, it is outvoted. A regional broker or a five-partner firm might be described from a handful of sources, so one stale Companies House-derived listing, one old directory entry with a previous address, or one aggregator that has merged you with a similarly named firm can become the answer.

That is why we keep pushing third-party authority sources in audits rather than just telling people to tidy their own website. Your own site is one voice. Accuracy is a vote.

It also connects to something we wrote about earlier in August, when a fake version of a firm ranked better than the real one . A lookalike site is the extreme version of this problem. One website answer in 15 pointing at a dead link, a lookalike or a different business is the everyday version, and it is far more common.

What the visibility tools actually count

Almost all of them count mentions.

A practitioner write-up doing the rounds this month tested more than twenty AI visibility tools, including the well-known ones, and landed on a conclusion worth repeating: most count how often your brand appears, and very few check whether the engine describes you correctly or has confused you with a competitor. It is one person's field test posted publicly rather than a controlled study, so I am not going to put a number on it. It is also not something a reader can verify, which is a fair thing to hold against it. But it matches what we see, and it matches what our own guide did.

There is a second finding in that write-up that we have now built into our method. What a tool pulls through an API can differ from what a buyer sees in the live web interface. So a dashboard number is not the buyer's experience. It is a proxy for it, and proxies drift.

The fourth metric we are adding

Accuracy rate: of the answers where you are named, what percentage describe you correctly?

Here is how we are running it. For each priority query, we already record whether the brand is named. We now also grade the description that comes with it, against four checks:

Identity. Is this actually you, or has the engine merged you with a similarly named business? This is the one that catches the most damage and the one nobody looks for.

Location and contact. Address, area served, phone, email. Wrong postcode was the single most common error in the Searchable retail data, and it is the easiest thing in the world to verify.

What you do. Are the services described the ones you sell? Engines frequently attach a plausible adjacent service that you do not offer, which generates enquiries you have to turn away.

The link. Does the URL it gives resolve to your live site? Not a redirect, not an old page, not a lookalike.

Score each answer pass or fail on all four, report accuracy as a percentage alongside mention rate, and list the specific wrong facts rather than just the score. The list is the useful part, because each wrong fact is a fixable job with a named source behind it.

We are also testing the live engine interface for a sample of priority queries rather than trusting the API output alone, and noting where the two diverge.

None of this is clever. It is the check that should have been there from the start.

What you can check yourself this week

Twenty minutes, no tools, no spend.

Log out first. Open a clean session, the way we described last time, because a logged-in test tells you about you rather than about the market. Then ask three questions in the plainest language a customer would use:

1. Who is [your business name] and what do they do?
2. Where are [your business name] based and how do I contact them?
3. Who are the best [your category] in [your area]?

Read the answers against your own verified details, line by line. Check the link resolves. Do it on two engines, not one, because they source differently even when they use the same label for the answer box.

Most firms that do this find at least one wrong fact. The common ones are an old address, a phone number from a previous system, a service you stopped offering, or a description borrowed from a competitor with a similar name.

Then go and find where the wrong fact lives. It is almost never on your own website. It is usually a directory, an aggregator, an old press mention or a data broker feeding several of them.

What to ask anyone selling you a visibility number

Last time we gave four questions: clean session or logged in, what the exact query was, how many runs over what period, and which engines, held constant.

Here is the fifth, and it is now the first one I would ask.

Does this number tell me whether the engine got me right, or only whether it mentioned me?

If the answer is only presence, the report is half a report. It might be a perfectly good half. But a firm with a 40% mention rate where every mention carries the wrong postcode is in worse shape than a firm at 25% that is described accurately, and no dashboard built on mention rate alone will ever tell you that.

We were selling the half without noticing. That is the honest version, and it is why this is a blog post rather than a quiet change to the method.

If you would rather have someone run the accuracy check for you under controlled conditions, the AI Visibility Audit now includes it as standard, and you can book a 45-minute call to walk through the method before you spend anything.