GEO, Measurement
Invisible is one failure. Mentioned but not recommended is another. The third is being described inaccurately, and nobody checks it. Why we score representation accuracy separately, why a friendly wrong answer counts as negative, and the four causes of a wrong description.
Most people worry about one failure mode: the AI has never heard of me. It is the obvious fear, and it is the easiest to test. Ask ChatGPT, Perplexity, Gemini or Claude a question your buyers ask, and see whether your name appears at all.
The second failure mode is subtler. Your name appears, but as background furniture rather than as the answer. We wrote about that in three ways a brand appears in an AI answer, and only one counts.
There is a third, and it is the one almost nobody checks. The engine names you, recommends you warmly, and describes a business that is not yours.
Here is the shape of it. A buyer asks for a plumber in their town. The engine says your firm is excellent, then adds that you specialise in commercial boiler installs and cover a city 40 miles away. You do domestic work. You do not cover that city.
Every word of that answer is positive. The tone is a glowing recommendation. And it sends the buyer somewhere else, because they are a homeowner in a town you were not credited with covering.
This is why our rubric classifies a hallucinated description as negative regardless of tone. Hallucination is the industry word for a confident, fluent statement that is simply not true. A sentiment score that rewarded that answer would be measuring the wrong thing. Warmth is not the outcome. Getting the right buyer to the right door is the outcome.
So we score representation accuracy as its own number, separate from whether you were mentioned and separate from whether you were recommended. Three different questions, three different answers:
You can score well on the first two and badly on the third. In practice, that combination is more common than most owners expect.
When we put this in front of small business owners, we expected the headline number to be the draw. It was not.
The thing they rated most valuable was "does the AI describe you correctly". It also turned out to be the thing none of them had considered before we asked. Owners had thought about being invisible. They had not thought about being visible and wrong, and once they saw it, it was the finding they wanted to act on first.
That reaction changed how we build. A score tells you where you stand. A wrong description tells you what to go and fix this afternoon.
An honest number needs enough samples to be a number rather than a coin flip.
If we ask an engine 3 questions about you and 1 answer is inaccurate, that is not "33% wrong". Re-run the same 3 questions and you might get 0 wrong, or 2. Language models are not deterministic, which means the same question can produce a different answer on a different day. Small sample sizes turn that variance into noise that looks like a finding.
So we refuse to display the accuracy percentage below 5 samples. Under that threshold we show you the individual answers instead, because reading 3 raw answers is genuinely useful, while averaging 3 raw answers is not. We would rather show you less than show you something that sounds precise and is not. It is the same instinct behind why we throw away AI competitor guesses rather than dressing them up as market data.
Wrong descriptions are not random. In the answers we read, they usually trace back to one of four causes, and each has a different fix.
If your own site never states plainly what you do, who you do it for and where, the engine has to infer it. Inference is where invention starts.
The fix is one plain sentence, high on the page, in ordinary text. Category, customer, place, and one concrete detail. "We are a domestic plumbing firm covering Leeds and Wakefield, Gas Safe registered since 2009." Boring is the point. Boring is liftable. If your homepage opens with a slogan instead, there is nothing for the engine to copy, so it writes its own version of you.
Engines read your directory entries, your review profiles, your social bios and your old press coverage alongside your site. When those disagree about your category, your service area or your opening hours, the engine picks one, and it will not always pick the current one.
The fix is dull housework. Make the name, category, address and one-line description identical everywhere you appear. Not similar. Identical. This matters most for local firms, where engines are often place blind and will happily attach the right service to the wrong town.
If another company shares your name, or something close to it, expect the engines to blend you. This is the most common cause of the friendly wrong answer, because the warmth comes from the other company's reputation and the buyer never learns they read about someone else.
The fix is differentiation in text. Always pair your name with your category and your location in the same sentence, on your own site and in every listing. "Northgate Plumbing, Leeds" is harder to merge with "Northgate Plumbing, Bristol" than "Northgate Plumbing" is on its own. You cannot make the other company disappear. You can make yourself easier to tell apart.
An old services page you never deleted. A 2019 pricing page still sitting on the site. A location page for a branch you closed. Engines read what is published, not what is current in your head.
The fix is to delete or update, not to hide. A page that is still reachable is still readable. And if it is possible the engines cannot read your current pages at all, start further back with what AI search engines actually read on your site.
These are not separate disciplines. Ahrefs studied 75,000 brands and found branded web mentions correlate with AI visibility at between 0.66 and 0.71, well ahead of backlinks at 0.218. A correlation is a measure of how tightly two things rise together, where 1 is lockstep and 0 is no relationship.
The point for accuracy is this: the same body of text that makes engines mention you is the text they use to describe you. Feed it consistent facts and you get both. Feed it fragments and you get a confident guess. That is also why a mention count on its own is a thin measure, and why we split traffic reporting into two numbers rather than one.
We keep coming back to two questions. Readiness asks whether the door is unlocked. Visibility asks whether anyone is walking through it. Representation accuracy is the third: when someone is sent to your door, are they being sent to the right one?
You can start without us. Ask each of the four engines to describe your business in three sentences, then mark each answer as correct, partly correct or wrong. If you get 5 or more answers, count them. If you get fewer, read them and resist the urge to turn them into a percentage.
Or let us do the counting. Run a free scan at whocanfindme.com. No signup. Takes about ten seconds.
When a buyer asks an AI one question, the engine quietly turns it into about ten. Seer Interactive measured 10.7 hidden sub-queries per prompt across 501 prompts on Gemini 3. Most sites answer the headline question and none of the ten. Here is how to find your gaps.
اقرأ المزيدThe same question, same engine, same day can produce different answers. That is the nature of AI search, not a bug. The 3 rules we use to avoid showing you fake statistics, and how to read a visibility score without fooling yourself.
اقرأ المزيد