How Georgye measures

1. How answers are collected

• AI models: ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic) and Google AI Mode. Each company chooses which ones to track.
• Web search on: questions are asked with web search enabled, the way most people use these tools today.
• Pinned model version: we use a dated version of each model. When a provider ships a new version, the switch is recorded, so a change in numbers is not mistaken for a change of model.
• Language and country: each question is sent in the company’s language; for Google AI Mode, also with the market’s country.
• Frequency and samples: from weekly to daily, with 1 to 5 samples per question in each collection, depending on the plan. The same question can get different answers each time; samples measure that variation.
• What is left out: failed answers and answers where the AI asked a question back instead of answering are stored but do not count in the metrics.

2. How brands are identified

Each tracked brand has a name and aliases (for example, “Comanda Mais” and “Comanda+”). We look for these names in the answer text, ignoring case and accents. The order in which brands appear sets each one’s position.

3. Visibility

Visibility = answers that mention the brand ÷ answers collected. Each answer counts once, even if it mentions the brand several times. Example: 580 of 1,106 answers mention the brand = 52%.

4. Average position

Average position = the average place where the brand appears among tracked brands, in the answers that mention it. Lower is better. Questions that already contain the brand’s name are left out: there the brand comes first only because it was asked about.

5. Sentiment

Sentiment = the average tone of mentions, from 0 to 100. Each mention gets a tone from the passage where the brand appears: positive is 100, neutral 50 and negative 0. The tone is classified by an AI model; until that runs, an initial keyword-based classification applies.

6. Share of voice

Share of voice = the brand’s mentions ÷ mentions of all tracked brands, calculated in each AI; the final number is the average across AIs. Each AI weighs the same, even the one that answered more. That is why share of voice depends on which competitors a company tracks.

7. Margin of error and “can’t tell yet”

Every percentage comes with a 95% margin of error, calculated with the Wilson interval. We only state a number with at least 30 answers and a margin of at most 15 points; below that, the screen shows the value as indicative or says it cannot tell yet.

Up or down? A difference between two periods is only treated as a change when it is larger than the sum of both margins. A smaller difference is noise, and alerts do not fire for it.

8. Sources

Every link cited in an answer is recorded with its domain, address and position. Sources are classified as the brand’s own site, a competitor’s site or third parties (editorial, reviews, forums, video, reference and others). A domain’s share is the number of answers citing it ÷ answers collected.

9. What Georgye does not measure

• Answers personalized by each user’s history, memory or login. Georgye measures the answer of an account with no history.
• Differences between each AI’s app and its API. Collection uses providers’ official APIs with web search on, which may differ in details from what the app shows.
• Voice or image answers.
• Every possible question: metrics apply to the questions being tracked. That is why choosing the prompts matters.