What is an AI visibility score, and how do you calculate it?
A plain definition of the AI visibility score, a worked example, per-engine versus blended scores, and why the prompt set matters more than the headline number.
An AI visibility score is the share of an agreed set of buyer prompts in which an AI answer names your brand. That is the whole definition. If you track 50 questions your buyers ask and ChatGPT names you in 12 of its answers, your ChatGPT visibility score for that period is 24%.
The simplicity is deliberate. A percentage is easy to explain in a board meeting and easy to compare month to month. The catch is that the number is only as honest as the prompts behind it, and most of the effort in measuring visibility well goes into choices that never appear in the headline figure. Below we cover the calculation, the decisions sitting underneath it, and how we run it for clients.
The basic calculation
You need a fixed list of prompts, a list of AI engines, and a rule for what counts as being named. With those agreed, the formula for each engine in each period is:
Visibility score = responses that name the brand ÷ responses collected
We count responses rather than prompts because AI answers are not deterministic. Ask ChatGPT the same question twice and you can get two different shortlists. So each prompt gets checked more than once during the month, and every answer counts.
Here is a worked example. The numbers are invented for illustration only. Say you track 50 prompts in ChatGPT and each one is checked four times in the month. That gives 200 responses. Your brand is named in 46 of them, so the score is 46 ÷ 200, or 23%.
The prompt-level view of the same data is more useful than the percentage. In this hypothetical, maybe 15 prompts named you at least once, 9 named you every time, and 35 never named you at all. Those 35 prompts are the work plan.
"Named" needs its own definition. We count a response when the brand appears by name in the answer text, including product names and agreed aliases such as a former company name or a common abbreviation. A link to your site with no mention of your name is a citation. We track that separately, because it means something different.
Per-engine scores versus a blended score
Each engine gets its own score because each one builds answers in its own way. ChatGPT's search features draw on pages fetched by OpenAI's search crawler. Google's AI Overviews and AI Mode use Google's index, and Google's documentation states that a page must be indexed and eligible to show a snippet before it can appear as a supporting link. Perplexity retrieves pages and lists its sources with the answer. A brand can be well represented in one engine and missing from another.
Continuing the hypothetical example:
| Engine | Responses naming the brand | Responses collected | Score |
|---|---|---|---|
| ChatGPT | 46 | 200 | 23% |
| Perplexity | 78 | 200 | 39% |
| Gemini | 14 | 200 | 7% |
| Blended | 138 | 600 | 23% |
The blended figure looks respectable. It also hides the fact that Gemini barely knows this brand exists, which is a separate problem with a separate fix. If you only reported the blend, nobody would go looking.
A blend can still be useful as an orientation number for leadership. If you use one, decide how it is weighted and write that down. An unweighted blend treats every engine as equally important. A weighted blend might lean on whichever engines actually send you visitors, based on your own referral data. Either is defensible. Changing the weighting between reports is not, because the trend line stops meaning anything. Our default is to lead with per-engine scores, show an unweighted blend underneath, and only weight it when a client has referral data that justifies the weights.
Why prompt selection matters more than the number
Our agency name is a small joke about this. You could score 99% tomorrow by filling your prompt set with questions that contain your own brand name. "What does [Brand] do?" names [Brand] nearly every time. The score would be accurate and completely useless.
A meaningful prompt set reflects how buyers ask before they know you. That usually means category discovery ("best payroll software for a 20-person company in Germany"), comparisons between named alternatives, and problem-led questions where the buyer describes a situation rather than a product type. We keep a small group of brand prompts too, but only to check that engines describe the company correctly, and we report them apart from the main score so they cannot prop it up.
The raw material for good prompts is usually already inside the business. Sales call notes, support tickets, on-site search logs and Search Console queries all show the language real buyers use. Marketing copy is the worst source, because it describes the product the way the company wants it described.
Two more rules keep the score honest. First, the set stays fixed for the reporting period. If ten prompts change halfway through the quarter, the before and after numbers are measuring different things. We add and retire prompts at agreed points and note every change in the report. Second, size is not quality. Fifty prompts that real buyers would type beat three hundred keyword variants, because the variants mostly test the same answer over and over.
Mention rate, share of voice and citation rate
These three get mixed up constantly, so it is worth separating them.
Mention rate is what this post calls the visibility score: responses that name you divided by responses collected.
Share of voice is your mentions as a share of all brand mentions within a named competitor set. In the hypothetical ChatGPT data, suppose that across the same 200 responses you were named 46 times, competitor A 80 times and competitor B 54 times. That is 180 brand mentions in total, and your share of voice is 46 ÷ 180, about 26%. Notice that share of voice depends on who is in the comparison set. Add a fourth competitor and your percentage drops without a single answer changing.
Citation rate is the share of responses that cite a page on your domain as a source. If 20 of the 200 responses linked to your site, the citation rate is 10%.
You can be cited without being named, when an engine pulls a statistic from your blog and then recommends someone else. You can be named without being cited, which usually means the engine's picture of you comes from other sites. Each pattern points to different work. We report all three and use mention rate as the headline, because it is closest to what matters commercially: a buyer reading the answer and seeing your name.
How often to measure
We report monthly but collect data throughout the month. A single check per month is a snapshot of a system that varies from one answer to the next. Spreading checks across the month gives a steadier picture.
Small week-to-week movements are mostly noise. What deserves attention is movement that holds for at least two reporting periods and concentrates on the prompts you actually worked on. Engines also update their models and retrieval systems without notice. When every prompt moves on the same day, that tells you more about the engine than about your content.
How we measure it in Promptwatch
Every client score we report is calculated in Promptwatch. It is the platform we run every score on, and the reasons are practical.
Its prompt tracking attaches search volume and difficulty scores to each prompt, which helps when deciding what belongs in the set. It shows query fan-outs, the related sub-questions an engine generates on the way to an answer, which often reveal phrasings we would not have thought to track. Topics and tags let us group prompts by buying stage, and personas plus country-level tracking let us measure the answer a specific buyer in a specific market would see. State and city tracking is available too, from the Business plan and on every agency plan.
Promptwatch monitors the actual interfaces of ChatGPT, Gemini, Claude, Perplexity, Copilot, Google AI Overviews, AI Mode and several other engines, not only their APIs, so the answers we score are the ones the product returns. Share of voice and sentiment are calculated against the competitor set we configure. Citation analytics shows which pages and domains each engine cites, including Reddit and YouTube citations, which is where the citation rate comes from. Prompt trends then show how each prompt's visibility moved between checks and what changed in the answer.
If you want to try the calculation on your own brand first, Promptwatch has a free Explore plan that covers 10 prompts on ChatGPT only. It is enough to see how much a single engine's answers vary from one check to the next, and why one percentage on its own never tells the full story.