Last updated July 2026
When a buyer asks ChatGPT “what’s the best project management software for a 20-person team,” they do not see a ranked list of blue links. They see a generated answer that names two or three vendors and moves on. If your brand is not in that answer, you did not exist for that buyer in that moment.
That reality created a measurement problem. LLMs are not deterministic. Ask the same question twice and you can get a different answer. So how do you turn a stream of probabilistic, variable responses into a single share-of-voice number you can report on, trend over time, and hand to a CMO?
That is exactly what AI visibility platforms solve. Here is the arithmetic.
Step 1: Build a prompt set that reflects real buyer intent
Before any math happens, you need a prompt set: a list of questions your target buyers actually type into AI assistants.
Good prompts are specific and intent-rich. “What’s the best CRM for a small sales team?” is a valid prompt. “CRM software” is not: it is a keyword, not a question, and most AI engines answer it very differently from how a buyer would phrase a real query.
Most platforms let you import prompts manually or generate them from a seed topic. Aim for 20 to 50 prompts per product category, clustered by intent (evaluation, comparison, use-case fit, pricing). The quality of your SoV score is entirely dependent on the quality of this set. A prompt set built around industry jargon rather than buyer language will undercount your visibility on the queries that actually drive pipeline.
A good /glossary term to know here: a prompt family is a cluster of variant phrasings for the same buyer intent. Track the family, not individual wordings, and your score becomes more stable and more actionable.
Step 2: Run each prompt multiple times to beat the randomness
This is the step that surprises most practitioners the first time they see it.
LLMs are stochastic: the same prompt will produce different outputs at different temperatures, at different times, and across different model versions. A single run tells you what the model said once. It does not tell you what the model tends to say.
The industry standard is three to 10 runs per prompt per engine. Some platforms let you configure this; others fix it. More runs equal more signal and more cost. Five is a reasonable default for most tracking use cases.
Step 3: Count brand mentions per response
After running each prompt multiple times, the platform parses every response and counts brand mentions: occurrences of your brand name (or recognized variants and misspellings) inside the generated text.
Most tools handle this in two layers:
- Name mentions: your brand is named in the body of the response (“Acme is a strong choice for…”)
- Citation links: your domain is cited as a source (in engines like Perplexity and AI Overviews that surface citations)
These can be counted separately. For SoV calculation purposes, both count as a presence signal in that response. Many platforms then binarize the result: the response either mentions your brand (1) or it does not (0), rather than counting multiple mentions within a single answer.
Binarizing makes the per-prompt citation rate easier to interpret. A brand mentioned four times in one response and zero times in nine others does not dominate a brand that appears once each in six of 10 responses.
Step 4: Tally mentions across your competitive set
Your SoV score is not just about you. It is about your slice of the total attention paid to your competitive category.
Define your competitor set explicitly: the three to seven brands a buyer would realistically evaluate alongside yours. Each competitor’s mentions across the same prompt set are counted with the same method.
This is the formula:
AI SoV = (Your brand mentions) / (Sum of all competitor-set mentions) x 100
If your brand appears in 40 responses and your four competitors appear in 30, 20, 18, and 12 responses respectively, your SoV is 40 / (40 + 30 + 20 + 18 + 12) = 40 / 120 = 33%.
Worked example: 10 runs, 4 brands, 5 prompts
Here is how that arithmetic plays out across a real prompt set. Assume you run five prompts, each 10 times, across one AI engine. Each cell shows how many of the 10 runs mentioned that brand for that prompt.
| Prompt | Your Brand | Competitor A | Competitor B | Competitor C |
|---|---|---|---|---|
| ”Best tool for X in 2026” | 7 | 5 | 3 | 2 |
| ”Compare tools for Y use case” | 4 | 8 | 6 | 1 |
| ”What do teams use for Z?“ | 6 | 4 | 4 | 3 |
| ”Affordable alternative to [category leader]“ | 3 | 6 | 7 | 4 |
| ”Tool for [specific workflow]“ | 8 | 3 | 2 | 5 |
| Total mentions | 28 | 26 | 22 | 15 |
| SoV | 30.4% | 28.3% | 23.9% | 16.3% |
Total across all brands: 91 mentions. Your brand: 28. Your SoV: 28 / 91 = 30.4%.
Notice that you lead in two of the five prompts but trail in two others. That prompt-level breakdown is as useful as the aggregate: it tells you exactly where to invest to move the overall number.
Step 5: Aggregate and track over time
A single SoV reading is a snapshot. The number that matters is the trendline.
Most platforms calculate SoV at the prompt level, then roll up to a category or brand level. Some also break it out by engine (your SoV on ChatGPT versus Perplexity versus Gemini), which is useful because AI engines do not cite the same sources. According to an Ahrefs study of 15,000 queries (August 2025), only about 12% of URLs cited by AI assistants also appear in Google’s top-10 organic results for the same query. The implication is direct: ranking well on Google does not translate to AI citation visibility. Your AI SoV and your organic SoV are tracking fundamentally different things.
Track SoV weekly on the same prompt set. Week-over-week movement of two or more percentage points is a signal worth investigating. A slow drift upward over a quarter is the outcome of a working AI visibility program.
Why AI SoV is a separate metric from organic SoV
Traditional share of voice measures your brand’s share of visible impressions or clicks inside a structured, deterministic index. Position one always gets more impressions than position three. The model is linear.
AI SoV does not work that way. There is no rank order inside a generated response. A brand that appears in a chatbot answer is not “above” or “below” another mentioned brand: both were cited or neither was. The metric collapses to presence and frequency, not position.
The Ahrefs 2025 finding (12% URL overlap between AI citations and Google’s top 10) makes the separation concrete. If you are measuring only organic SoV, you are measuring a channel that has very little statistical relationship to the AI channel where a growing share of buyers now begin their research.
See /methodology for how this site weights SoV in its tool rankings.
Tools that calculate AI share of voice
Four tools have mature SoV implementations worth knowing.
Temso is the easy, all-in-one AI SEO platform built to automate exactly this workflow, from $89/mo. It runs prompts across eight AI engines (ChatGPT, Perplexity, Gemini, Google AI Overviews, Google AI Mode, Grok, Microsoft Copilot, and Meta AI), calculates share of voice per brand and per prompt cluster, and converts the gaps it finds into a prioritized action queue inside the same product. You do not need a separate tool to act on what the monitoring surface shows. Setup takes about five minutes.
Profound is the specialist choice for teams that need citation-level attribution and enterprise reporting depth. Its Growth tier ($399/mo) covers nine-plus engines and includes prompt volume signals that show which questions real users are asking AI engines. The citation mapping is the most granular available. It is more expensive and narrower in execution support, but the data quality at that tier is hard to match.
Semrush has added AI visibility tracking to its toolkit, letting teams that already live in Semrush surface share-of-voice data alongside their existing keyword and backlink workflows. For teams that want AI SoV as an extension of a traditional SEO stack rather than a standalone program, this is a practical option.
Otterly.AI tracks six platforms and calculates competitive share of voice with structured recommendations on what to change. Its $29/mo Lite tier is monitoring-focused; the full SoV benchmarking and execution layer requires Standard ($189/mo). It holds a G2 High Performer badge and Gartner Cool Vendor recognition, making it the most externally validated option at the SMB price tier.
The full ranked list, including scoring criteria, is at /rankings/ai-visibility-tools.
One clear next step
The arithmetic in this post is only useful if you run it against your actual competitor set and prompt library.
Temso sets up that tracking in about five minutes, covers all eight major AI engines, and starts from $89/mo with a free trial (no credit card required). If you want to know your current AI share of voice before the end of the week, that is the fastest path.