How to Measure AI Visibility – Share of Voice

A business owner asks ChatGPT a question, sees their own name and says "fine, we're visible". The next day they ask the same question and the name is gone. Which is true? Both. Generative engines don't give the same answer every time; that's why the answer to a single question isn't a measurement, it's an anecdote.

Share of voice is the way to turn that uncertainty into a manageable number.

Definition

Share of voice is the percentage of all answers to a given set of questions in which your business appears.

A simple example: you asked 20 different questions to 4 engines and got 80 answers. Your name appears in 24 of them. Your share of voice is 30%. If your competitor appears in 40 of the same 80 answers, their share is 50%, and the gap between the two is what gets discussed at the marketing meeting.

Why share, not rank

In a search engine, "we're third" means something, because the results page is fixed and ordered. A generative engine's answer is neither fixed nor ordered: sometimes it names three businesses, sometimes one, sometimes none. In that structure, "what position" is meaningless and "in how many answers" is meaningful.

Share also captures two things at once: whether the model recognizes you, and whether, when it does, it matches you to the question. A high share shows both are happening.

How to calculate it

There are four decisions, and all four affect the result.

1. The prompt set. The questions customers actually ask; not the ones you wish they asked. Not "best glass balcony company" but "I want glass fitted on my balcony in Karşıyaka, who would you recommend?" Keep the set to 15–30 questions, with variety across service, area and customer type. Define the set once and freeze it: if you change the questions at every measurement, the results can't be compared.

2. The engines. The engines used in your market: in Turkey, ChatGPT and Gemini lead, with Perplexity and Claude behind. Each engine is measured separately; you can combine them afterwards if you like.

3. Repeats. Ask the same question to the same engine at least twice, preferably three times. A single answer is a coin toss.

4. The counting rule. What counts as "appearing"? Your name being mentioned, or you being recommended? Count the two separately: mention share and recommendation share. The gap between them shows how big the "recognizes but doesn't recommend" problem is; for most businesses, that's where the real issue lies.

Comparing with competitors

A share on its own says little; is 30% good or bad? Count your competitors on the same prompt set. A table with three or four competitors says more than a single number: who's strong on which engine, which question types nobody shows up for (open ground), which area questions you lead on.

Freeze the competitor list too. Swapping competitors from one measurement to the next is moving the goalposts.

Common mistakes

• One question, one day. As above; anecdote, not measurement.
• Questions that contain your own name. "Is Protan any good?" doesn't measure visibility; you've already said the name.
• Expecting results immediately. Changes reach the engines over weeks, sometimes months. Monthly measurement is a sensible rhythm; weekly measurement measures noise.
• Tracking the share alone. Going from 30% to 35% is good; but on which engine, on which question type? A number with no breakdown produces no decisions.

A simple worksheet

Columns: question, engine, repeat number, your state (recommends / recognizes / doesn't mention), competitor A, competitor B. Fill in the same table every month. After three months you have a curve rather than a single number; and the curve tells you whether the changes you made are working.

protan.ai keeps this table for you: the prompt set is frozen, the engines are queried regularly with the same questions, and the state and share are reported for each engine. If you do it yourself, the method is the same; only the time is yours.