You're in first place in Google's search, and there's no guarantee that ChatGPT, Perplexity or Gemini will mention you in answer to "best XX tool". The AI engine recommends a stand-alone list of people who are almost out of line with natural search. Competing AI Share of Voices analysis is putting out this invisible ranking: in your class, AI re-recognizes which families and for what reason, and you're in what or not.
What's the difference between AI Share of Voice and traditional SEO?
Traditional SEO Share of Voice calculates how much visibility you have in the key wordspace, in "connection." AI changed a set of rules: the user asked a question, and the model went on to say a few words and named a few brands. Your molecule is "AI mentions you in related questions" and the denominator is "all brands are mentioned in total". Even more troublesome is that the answer to the same question is not necessarily the same, and the model itself is random. This means that you can't just conclude once, and the sampling method is more important than traditional rankings.
First step: build a first set of questions, not a key list
The tradition of SEO is to start with key words and GEO is to start with "problems". What you're going to list is that potential clients actually get into ChatGPT's sentence when they make their decisions. A SaaS, a B2B electronic signature, is not the word "electronic signature" but a whole set of questions. Split them into three categories, cover them more fully.
- Type reference: "What are the good electronic signature software? "A common contract signature tool for the Taiwan B2B team."
- Contrast: "A and B, which is better for small and medium enterprises? "B, what's the alternative?" -- which one is the most vulnerable to your competition?
- The situational issue: "The electronic signature tool for budget limitations that require accompaniment of API", "The Signing Service in accordance with Taiwan Law" -- AI, cutting off half of your brand here, can see if you're in the right situation.
A class grab 20 to 40 is enough to start. It's in the user's language, not in your inner language, because AI is learning how to answer the questions.
Step 2: Cross-engine, repeat multiples
Just ask the ChatGPTs, get the wrong sample. At least cover ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews because of the data sources and sorting logic behind them. Different: Perplexity focuses on immediate search and reference sources, and ChatGPT relies more on training and searching for language, which may vary significantly. Each one runs at least three to five times in each engine, shuts down personal and historical records, uses clean working sessions, and avoids the model remembering your last sentence selling yourself.
The first time a lot of clients did this was "Ai didn't even count us in." It's not that I dropped the ranking. It's that list. This is more worrying than going to page two.
Step three: Split the answer to calculable data
The sample can't just be assumed. Each response is presented in a table, with a sample, and the fields contain at least the engine, the questions, the list of brands mentioned, the order in which each brand appears, and an AI description of each brand. The brand name needs to be regularized and the "Tenten" "tenten.co" should be the same entity, otherwise the volume will be broken down and the number will be low. When this step is done, you have the original data that can be calculated over time, not a bunch of screenshots.
Step four: calculate four really useful indicators.
With the data sheet, it's crushed into four numbers that can be taken directly to a meeting. Read the number of times referred to, it's a misunderstanding.
- Incidence: What percentage of class you were mentioned. It's a visible ground, with a lower than 30% presence, and all the other indicators go ahead.
- Voice is: Your number of references divided by the total number of references in the whole class, which is the lowest-known AI Share of Voice, which is the most impressive combination of the top three competitions.
- Average sorting position: You usually rank first when you're mentioned. One of the first names in the AI answer, and the fifth convinced that it's far from the best.
- Situational correctness: AI, when referring to you, did you get the location right? To be recommended to the wrong situation is to help the competition.
Count these four indicators against every major competition, and you'll get a cruel but useful picture of who's in AI's eyes the default answer for this kind of thing, who's just a role model, who's completely invisible.
Step five: Read the "justification" of AI to recommend the competition.
Volume numbers tell you how big the gap is, and then why. By putting together all the words that AI describes each competition, you will find that it repeats some words in order to locate the leader: it may be "high integration" "with SOC 2 authentication" "clear file". These words are not empty, but they respond to statements that have been extensively repeated in competition networks, third-party evaluations and community discussions. When your brand is absent, it's usually not because it's poor, but because it's public information that is drawn from AI, and you don't write, or it's not sufficiently structured.
From point to gap
This analysis is not a one-time physical examination, the model is updated, the competitions are added, the queue is moving every month, so it's worth running back to the same problem set on a fixed weekly basis, tracking its volume up or down. The real job started after the plate was finished: for the reasons that AI recommended the competition, to complete your own evidence, location, and clean extracts, so that the next round of sampling would have a reason to name you. If you want to see where your AI voice gap is in the class, you can expect a 30-minute GEO diagnosis, and we'll use your own class set, and we'll run an initial competition for you.
