Methodology
How the AI Visibility Score is computed.
Last updated 2026-09-28.
Score
What is AI Visibility?
AI Visibility measures how often AI assistants mention your company, meaning the right company with something real to say, when they are asked the kinds of questions a buyer in your category would ask. It is not an SEO ranking, not a click-through prediction, and not a measure of online sentiment at large. It answers one question: when ChatGPT or Claude gets the buyer questions from your analysis, how often does your company actually show up in the answer? The AI Visibility Score is a number from 0 to 100. A score of 35 means your company was visible in 35 of every 100 questions that don't name it. It is a share of questions, not of individual AI answers. The score is computed on the fixed set of questions in your analysis, so you can re-run the same questions later and track the change.
How the score is calculated
The score is a share of questions: round(100 × visible questions ÷ questions that don't name your company). Questions that contain your company name are shown separately, under "Questions that name your company", and are not part of the score. A question counts as visible when at least one provider (ChatGPT or Claude) clearly talks about your company in at least 2 of its 3 answers on the paid plan, or in its one answer on the free plan. Visible on either provider is enough. A small AI step reads each answer and decides whether it is really about your company — not a different company with the same name, and not an answer that says it doesn't know you. These never count: answers about another company with the same name, answers where the AI says it doesn't know enough about you, and answers we couldn't verify are about you. Position and sentiment don't change the score: a first-place and a fifth-place mention count the same, and so do positive and negative ones. On the paid plan, sentiment is shown next to each mention in the report and in the CSV export, and the CSV also shows your position in each answer. To check the score yourself from the CSV (paid plan): use only the rows where names_brand is false. Each question has one row per AI assistant. A question is visible if any of its rows has classification = visible. Divide the number of visible questions by the number of questions, and round to a whole number.
Providers
Which AI assistants we ask
Lumialo works with two AI assistants: ChatGPT (OpenAI) and Claude (Anthropic). A free analysis asks ChatGPT. The paid plan asks both, which gives you a second, independent reading of every question.
Versioning + change cadence
AI models change. The same question asked of "Claude" in March can return a different answer in June because Anthropic shipped a new model version in between. That is normal, and it is exactly why a reproducible visibility score has to record which version of each provider produced each answer. Lumialo records the exact AI model version that answered every question. The versions are listed at the bottom of your report. If you compare two runs months apart and your score moved, those versions help you tell whether the move was the AI changing, your brand presence changing, or both. The AI assistants' own apps may run other model versions than the APIs we ask, so an answer you see yourself can differ from the one in your report.
Sample size
Questions per run
A free run analyses up to 10 questions; a paid run analyses up to 30. The questions are grouped into question groups you and Lumialo defined together in the wizard, so the set is specific to your business rather than a generic industry checklist. More questions give you a wider read on your category, which is why the paid plan opens up the larger set. The free plan is a trial: fewer questions, one answer from ChatGPT per question and no recommendations, so you can see how the analysis works before paying. The 30-question paid ceiling is a cap on how much work one run does — the cost of an analysis follows the number of questions, so the ceiling is what keeps a single run from growing without limit; it is not a statistical lower bound. Treat the numbers in your report as descriptive of how the AI answered those specific questions, not as a survey of the wider internet at large.
Runs per question (statistical confidence)
AI models vary — ask the same question twice and you can get two slightly different answers. To control for that, the paid plan asks each question multiple times on each provider and treats the brand as mentioned for a question when at least two of those runs surface it. The default on the paid plan is 3 runs per question per AI assistant. The free plan uses 1 run per question on a single assistant, which is enough to demonstrate the analysis but is more exposed to the fact that AI answers can vary day to day.
Evidence
How many recommendations you get
Recommendations are part of the paid plan; a free analysis does not include them. We only write recommendations we can back with your data; a strong result produces fewer. Every recommendation has to pass these checks before it appears in your report: 1. It points to a question from your analysis, quoted exactly, that doesn't name you and where you were not visible. 2. It names a competitor the AI brought up on its own in that answer. 3. It quotes the AI's answer word for word as evidence. 4. It agrees with your report: it never says you were missing where you were visible, and it only uses percentages that appear in the report. 5. It states no facts about your company that aren't in your description or in an AI answer that names you. A question is never taken as proof of what you offer. 6. It starts with a concrete action and names what to create or change, and no more than two recommendations suggest the same kind of action. If we find no gaps (you showed up wherever competitors did), there are no recommendations, and the report says so. If we find gaps but only some recommendations pass these checks, you get the ones that passed — we would rather send you three you can trust than five you cannot. The rest of the report is delivered as normal.
How recommendations are validated
Each draft recommendation is written by an AI model and then checked automatically against the rules above, using your report's own data. If a draft fails, the generator tries again, and on the third attempt it uses a stronger model. Drafts that still fail are dropped rather than weakened. Lumialo would rather give you a few recommendations you can act on than many you can't.