Brand Radar did it in one sentence: treating the AI engine as an entity that gives different answers every time, throwing the same questions over and over again, and counting the rates of reference to your brand. It doesn't check the key word ranking, it doesn't look at web traffic, it's "how high and how trusted you are when potential clients ask who AI is to find." The success of the entire process depends on a sample design, not on the technology behind it.
We need to figure out what Brand Radar is measuring.
The traditional SEO tool gives you an address in Google number one. The AI engine has no fixed number. You ask ChatGPT, "Who are the GEO agents in Taiwan for B2B SaaS?" It's probably three today, five tomorrow, and the order will change. So visibility cannot be expressed single-handedly, but only as "the probability that you will be included in a lot of repeat queries". Brand Radar breaks down this rate into quantifiable indicators: reference rate, reference rate, average location, speech orientation, and the volume of the competition.
Step 1: Turn the buyer's real questions into a collection of hints.
The most common mistake here, starting with the collection of hints, is to use the key words you want to be searched for. Buyers will not ask for "GEO Audit Service" and they will ask "Ai's search never mentioned our company, who should deal with it" or "how do you know that ChatGPT is recommending my product?" We build this list from three sources: what business has actually been asked by clients, the original words of the client's and sales conversations, and the self-help and advice of various engines. It takes enough size to get a steady ratio, but not as big as possible, too much to dilute the real business problem. In a Leegie B2B class, it's usually enough to catch between 40 and 120 core hints, to overwhelm perceptions, to compare, to make decisions, and the point is that each one has a real shopping scene.
Step 2: Why do you have to do it over and over again?
The answer of the AI engine is random. The same problem, the same model, ten runs could give you eight words, and in different order. It's only once. You've got a message, not a signal. Brand Radar's approach is to repeat each hint, repeat the sample on each engine, then take a statistical distribution, and you're mentioned seven times as a 70% reference rate, which is the number that holds. The sample is going to cross the engine because ChatGPT, Perplexity, Gemini, Google AI Overviews have a much different source preference, and the same brand may be far from visible in different engines.
Step three: Retrieving responses and distinguishing between "mentioned" and "quoted"
Every response that comes back from the sample is to have two different interpretations. The first is to mention: does your name, product name or key person appear in the text of the answer? The second is a quote: does the source link, footer or reference list at the bottom of the answer point to your website? The two are often confused, but the meaning is completely different, the text is mentioned to represent the model that remembers you, and the reference is being used as evidence in this inspection. Healthy brands are both, except for references that are not quoted, which usually means that you're holding on to your old reputation, and you're undercovered. There is also a trap for recognition: brand names are easy to misread, other companies with the same name are mixed in, models may spell names wrong or only say half. So taking the steps requires a rule of non-discrimination that separates what is really said about the brand from the homogeneity of information, or the scores that are calculated later are false.
Step four: Count -- how do you calculate the visibility score?
With clean references and references, they are then reduced to traceable scores. The core logic is a weighted one, not a genre:
- Reference rate: In all samples, the brand is mentioned in the text.
- Reference rate: When responding to the source, the source points to the proportion of your website, reflecting the extent to which the content is taken as evidence.
- Location: Is the brand listed at the front end of the answer or is it a pre-eminence sentence?
- Voice: AI describes you as positive, neutral, or as missing, directly affecting the real value of this exposure.
- Voice to voice: compare you to the main competition in the same set of hints, see who is more often selected by AI.
These indicators will then be based on the commercial value-added of the hint, deciding on the issue in the decision-making phase, with a clear purchase plan, and giving priority to the issue of general awareness. In the end, it is a condensed visibility score that facilitates communication, but what is really used to make decisions is often a piece of paper, as they point to different patches: content and structured evidence with low reference rates, and voice-to-mouth to topics often cited in competitions.
Step 5: Turn a single snapshot into a trend.
The score for a single moment can only tell you the situation, not the cause or the effect. Values are constantly being tracked: by reruning the same hints on a fixed weekly period and drawing the scores into a time series, "we added three comparisons last month" to "Perplexity's reference rate starts to go up two weeks later." The trend also exposes risks early in the day, when the volume of a competition rises three times in a row and is eating your answers, a signal that can be seen before the traffic on your website is reflected.
These numbers, how to use them in practice.
Visibility scores are not a vanity indicator for reporting. Its purpose is to tell you what the next article is about, which engine's most important gap, and whether the power you spend really makes AI more frequent.
Walking through Brand Radar's logic, you'll find it's actually a model method of measurement that's put on an AI engine that's a variable Matrix. The difficulty is not to capture the data, but to match the hints, clean up the details, and set the score back to the action that can be executed. If you don't know where you're talking to and where you're going with the main AI engine, where the gap is stuck, you can expect a 30-minute GEO diagnosis, and we'll run a round of practical problems of your type, and we'll clear up the drop points and priority patches.
