What is AI share of voice?
AI share of voice (AI SoV) is the percentage of AI assistant answers, across a fixed set of buyer prompts, that mention your brand versus your competitors. It is the visibility metric for generative search. In a world of AI answers, share of voice now comes before share of market.
Traditional metrics like keyword rankings or impressions assume the user sees a list of links and chooses one. AI assistants do not work that way: they read many sources and return a single synthesized answer that may name a handful of brands, or none. Your share of voice is simply how much of that conversation belongs to you. If an assistant consistently recommends three vendors and you are never one of them, you are invisible to a buyer who never scrolls past the answer.
Because the answer is the destination, being named is the new equivalent of ranking. AI share of voice turns that into a number you can track, compare against rivals, and improve over time, the same way teams have tracked organic visibility for years.
AI SoV vs traditional SEO share of voice
Classic SEO share of voice estimates how visible you are across ranked links, weighted by keyword search volume and position. AI share of voice measures how often you are actually named inside the generated answer, where a clickable link often does not exist at all. The first is about ranking among links; the second is about being included in the answer. They overlap, but they are not the same, which is why a brand can rank well on Google and still be missing from ChatGPT: roughly 80% of the URLs AI assistants cite do not rank anywhere in Google for the same query (Ahrefs, 2025).
Why is AI share of voice the new marketing KPI in 2026?
Discovery is moving into AI answers faster than budgets are moving with it. Assistant adoption is now measured in hundreds of millions of weekly users, while more searches end without a click. When the answer replaces the link, the only honest measure of visibility is how often the answer names you.
The shift has two engines. First, adoption: AI assistants have reached mainstream scale. Second, zero-click behavior: even on Google, an AI summary reduces the chance that anyone clicks through to a website. Together they mean a growing share of buying research happens inside a generated answer your analytics never sees.
| Signal | What it shows | Why AI SoV matters |
|---|---|---|
| Assistant scale | ChatGPT reached about 800M weekly active users in 2025. | Your buyers are already asking assistants about your category. |
| AI in Google | Google AI Overviews reach roughly 1.5B users per month. | AI answers now sit above the classic links you optimised for. |
| Search displacement | Gartner predicts a ~25% drop in traditional search volume by 2026. | Visibility leaks from search engines into assistants. |
| Fewer clicks | With an AI summary, ~8% of users click a link vs ~15% without. | Being named in the answer matters more than ranking below it. |
Sources: OpenAI via TechCrunch, 2025; Alphabet, 2025; Gartner, 2024; Pew Research Center, 2025.
Sources: OpenAI via TechCrunch, 2025; Alphabet, 2025; Perplexity via TechCrunch, 2025; Gartner, 2024; Pew Research, 2025; SparkToro / Datos, 2024.
Which metrics make up AI visibility?
AI visibility is not one number but a small family of metrics: citation rate, mention rate, average position, share of voice and sentiment. Each answers a different question, from "are we linked?" to "how are we described?". Together they turn vague "AI presence" into something you can measure and improve.
One important nuance: many assistants name brands without giving a clickable source. Research on Gemini found it provided no clickable source in roughly 92% of responses (Strauss et al., arXiv 2508.00838, 2025). That is why mention rate, not just citation rate, is essential, otherwise you would undercount the times an assistant talks about you without linking out.
| Metric | Definition | Why it matters |
|---|---|---|
| Citation rate | Share of answers that link to your site as a source. | Shows the assistant trusts your content enough to attribute it. |
| Mention rate | Share of answers that name your brand, with or without a link. | Captures visibility even when no source is clickable. |
| Average position | Where you appear in the answer (first, mid, or last brand named). | Being named first carries more weight than a passing mention. |
| Share of voice | Your mentions as a percentage of all brand mentions in the set. | The comparative headline metric against competitors. |
| Sentiment | Whether you are described positively, neutrally or negatively. | Being named is not enough if the framing is unfavourable. |
Note on citations vs mentions: Strauss et al., 2025 (Gemini provides no clickable source in ~92% of responses).
How do you build a representative prompt set?
Build a fixed set of 20 to 50 prompts that mirror the real questions your buyers ask an assistant, balanced across intents and platforms. The prompt set is the foundation of every measurement: if it does not reflect real demand, your share of voice number describes a conversation nobody is actually having.
Good prompts are phrased the way people talk to an assistant, longer and more conversational than a typed Google query. Cover the full buyer journey rather than only branded searches, and keep the wording stable so you can compare results over time.
- Category questions — "what is the best tool for X", "alternatives to Y" — where you compete head-to-head.
- Problem questions — "how do I solve X" — where you can be recommended as a solution.
- Comparison questions — "X vs Y" — where position and sentiment matter most.
- Branded questions — "is your brand any good" — to check how you are described.
A simple way to keep the set balanced is to allocate prompts across intents on purpose, rather than letting one type dominate. The blueprint below shows a workable split for a 40-prompt set; adjust the weighting toward whichever stage of the journey matters most for your business.
| Intent type | Example prompt | Suggested share of set |
|---|---|---|
| Category | "best tools to measure brand visibility in AI search" | ~35% |
| Problem | "how do I track brand mentions in ChatGPT" | ~30% |
| Comparison | "option A vs option B for AI share of voice" | ~20% |
| Branded | "is [your brand] good for AI visibility tracking" | ~15% |
Illustrative split for a 40-prompt set; tune the weighting to your own funnel.
How many prompts is enough?
For most brands, 20 to 50 well-chosen prompts give a stable read without becoming unmanageable. Below 20, a single odd answer skews the whole percentage; far above 50, you add maintenance without much new signal. Start focused on your highest-value buyer questions, then expand into adjacent intents as you learn which ones move your share of voice.
How do you calculate AI share of voice (step by step)?
Run every prompt across each assistant, count how many answers mention your brand, and divide by the total number of answers in the set. Multiply by 100 to get a percentage. Repeat the whole process on each assistant and at regular intervals so you can see the trend, not just a snapshot.
AI share of voice = (answers mentioning your brand ÷ total answers in the prompt set) × 100
Here is a labeled example. Suppose your prompt set has 40 prompts and you run each one once on a single assistant, giving 40 answers. Your brand is named in 12 of them.
- Total answers: 40 prompts × 1 run = 40 answers.
- Answers mentioning your brand: 12.
- AI share of voice: (12 ÷ 40) × 100 = 30% on that assistant.
To make it comparative, run the same count for each competitor across the same 40 answers, then express everyone as a percentage of all brand mentions. Finally, repeat the run on every assistant you care about and average or report per platform, because a 30% share on Perplexity can sit next to a 5% share on Gemini for the very same brand.
Why report distributions and confidence intervals, not single numbers?
AI answers are volatile: the same prompt can return different brands from one run to the next. A single query is noise, not a measurement. The honest approach is to sample, run each prompt several times across days, and report a range rather than one precise-looking percentage that hides the uncertainty.
The volatility is not only about which brands appear, it is also about whether the assistant gets attribution right. An audit by the Tow Center for Digital Journalism found AI search engines were wrong on more than 60% of source attributions, with error rates spanning roughly 37% on Perplexity to about 94% on Grok-3 (Columbia Journalism Review, 2025). If the engines themselves are inconsistent about sources, a single reading of your share of voice cannot be trusted either.
Practical rule: repeat each prompt several times per assistant, aggregate, and report your share of voice as a range with a central value, for example "28%–34%, median 31%". A trend line over several cycles tells you far more than any one number.
How do you track it continuously across platforms?
Manual checks do not scale: a meaningful prompt set across six assistants, repeated weekly with several runs each, is thousands of answers to read and classify. The only sustainable approach is automated, per-platform, over-time monitoring that records mentions, position and sentiment so you watch a trend instead of guessing.
Doing this by hand also introduces inconsistency, the wording drifts, the timing varies, and nobody logs the results the same way twice. Continuous tracking fixes the prompt set, runs it on a schedule across each assistant, and stores every result so you can compare this week to last and react when a competitor surges or you slip.
If your share of voice is low, the next question is diagnostic: why are you missing from the answers in the first place? That is covered in our companion piece on why your brand isn't showing up in ChatGPT. For the underlying tactics, see the GEO guide and how it works.
Sources
- Aggarwal et al. (2024). GEO: Generative Engine Optimization, KDD 2024. arxiv.org/abs/2311.09735
- Tow Center for Digital Journalism, Columbia (2025). We compared eight AI search engines. cjr.org
- Strauss et al. (2025). arXiv 2508.00838. arxiv.org/abs/2508.00838
- Gartner (2024). Search engine volume to drop 25% by 2026. gartner.com
- Alphabet (2025). Q1 2025 earnings, AI Overviews ~1.5B monthly users. blog.google
- OpenAI via TechCrunch (2025). ChatGPT at 800M weekly active users. techcrunch.com
- Perplexity via TechCrunch (2025). 780M queries per month. techcrunch.com
- Pew Research Center (2025). Users less likely to click when an AI summary appears. pewresearch.org
- SparkToro / Datos (2024). 2024 zero-click search study. sparktoro.com
- Ahrefs (2025). AI search overlap with Google rankings. ahrefs.com
Frequently asked questions
Is AI share of voice the same as SEO share of voice?
No. SEO share of voice estimates your visibility across ranked links in search results, usually weighted by keyword and position. AI share of voice measures how often your brand is actually named inside generated answers, where there is often no link to click. They are related but measure different surfaces, so you should track both.
How often should I measure AI share of voice?
Measure on a recurring schedule rather than once. AI answers are volatile, so a single check is noise. Most brands run their prompt set weekly or every two weeks to see trends, and increase frequency after publishing new content or a competitor move. Consistent timing matters more than the exact interval.
Which AI assistants should I include?
Include the assistants your buyers actually use. Today that usually means ChatGPT, Google AI Overviews, Gemini, Perplexity, Claude and Grok. Coverage matters because each engine retrieves and cites sources differently, so a brand can be strong on one and invisible on another. Measuring only one platform gives a misleading picture.
Can I track competitors with AI share of voice?
Yes, and you should. Share of voice is comparative by definition: your percentage only means something against the set of brands competing for the same answers. Track which competitors get named, how often and in what position, so you can see who owns each buyer question and where you can realistically gain ground.
How long until AI share of voice results are reliable?
A single run is directional, not reliable. Reliability comes from sampling: repeating each prompt several times, across assistants, over multiple weeks, then reporting ranges instead of one number. After a few measurement cycles you can distinguish a real trend from day-to-day variance and act on it with confidence.