AI share of voice
AI share of voice is your slice of the brand mentions that AI engines give when people ask for recommendations in your category. This is the definitive guide to what it is, how it differs from AI visibility, the formula, and how to measure it across every engine, backed by 20.4 million real AI citations.
What is AI share of voice?
AI share of voice is the percentage of AI answers that mention or recommend your brand, relative to your competitors, for a defined set of category questions. When someone asks ChatGPT "what is the best project management tool for remote teams?", the assistant does not return ten blue links. It names a handful of brands and explains why. Your AI share of voice is how big your slice of those named brands is, across all the questions your buyers actually ask.
It is the direct descendant of traditional media share of voice, the metric brands have used for decades to compare their presence in advertising, PR, and search against the rest of the category. The channel is new. The idea is not. What has changed is that the "results page" is now a generated paragraph, and the brands inside it were chosen by a model, not ranked by a position.
That single shift changes everything about how the metric behaves. In classic search you hold a fixed rank: first, second, third. In an AI answer there are no slots. A brand can be named first, recommended outright, buried in a list, or actively cautioned against, and the same question asked twice can return a different set of brands. So AI share of voice is never one number from one query. It is a distribution: measured across many prompts, many engines, and many runs, then summarised.
The one-line definition to quote
Why it matters now, not later
ChatGPT has more than 800 million weekly users. Perplexity answers millions of queries a day. Google AI Overviews now sit above the organic results on a growing share of searches. For a large and growing slice of your market, the AI answer is the research step. If your brand is not in it, you are invisible at exactly the moment a buyer is forming their shortlist, and unlike classic SEO, where you can see your rank in Search Console, this has been a blind spot. Measuring your AI share of voice is how you turn that blind spot into a number you can move.
AI share of voice or AI visibility: which is the better metric?
This is the question buyers ask most, and the honest answer has a point of view: they measure different things, you need both, and which one leads depends on where you are. Treating them as the same number is the most common mistake teams make when they start tracking AI search.
AI visibility is an absolute measure: how often you appear, full stop. It answers "am I showing up at all?" AI share of voice is a relative measure: your appearances divided by everyone's. It answers "am I winning the category?" The trap is that the two can move in opposite directions. Your visibility can climb while your share of voice falls, because the whole category got more visible faster than you did. You felt like you were winning. You were losing ground.
- / Absolute: how often you appear
- / Answers "am I present?"
- / A clean trend line, hard to game
- / Best when you are climbing off zero
- / Blind to what competitors are doing
- / Relative: your slice of the category
- / Answers "am I winning?"
- / The truer competitive signal
- / Best once you are reliably present
- / Sensitive to your competitive set
The verdict
If you can watch only one number, the order matters. While you are still climbing off zero, lead with visibility: your first job is simply to exist in the answer, and visibility is the cleaner, less gameable trend line for that. Once you appear reliably, manage to share of voice, because at that point the only question that matters is whether you are taking ground from competitors or losing it to them, and share of voice is the only one of the two that can tell you. Visibility is your health metric. Share of voice is your growth metric. The teams that conflate them either celebrate a rising visibility number while quietly losing the category, or chase a share-of-voice figure that is noise because they are barely present yet.
Trakkr tracks both, side by side
Visibility score and competitive share of voice on one dashboard, per engine and per prompt.
The AI share of voice formula
The base formula is simple. The judgement is in what you count.
AI share of voice = (your brand mentions / total brand mentions across all competitors) x 100
Say you track 100 category prompts. Your brand is named in 40 of the answers. Your named competitors are collectively mentioned 200 times. Your raw AI share of voice is 40 / (40 + 200) = 16.7%. That is the number most tools report, and it is a fine starting point.
But a raw count treats "named first and recommended" the same as "mentioned once at the end of a list." In AI answers, position carries real weight: the first brand named is the one a busy reader remembers. That is why serious measurement uses position-weighted share of voice, where a first or explicitly recommended mention is worth more than a trailing one. The formula is the same; you just weight each mention by its prominence before you sum. The result is a share that reflects influence, not just presence.
Try it
Drop in your own counts to see the raw share. Then read it as a floor: the position-weighted version is usually what you manage to.
Count one mention per brand per answer. For a sharper number, weight a first or recommended mention more than a passing one (that is position-weighted share of voice).
Your brand is named in 35 of 240 total brand mentions. That is your raw share of the category's AI voice.
Why one engine is not the category
The single biggest measurement error is to check share of voice on one assistant and assume it represents "AI." It does not, and we can prove it with our own data. Across 825,343 category prompts run through eight engines, the major models agreed on their answer only 43.3% of the time on average, and produced a perfectly identical set of brands just 4.0% of the time. Different training data, different retrieval behaviour, different source preferences. Your share of voice in ChatGPT can look healthy while you are absent from Gemini entirely.
The engines also differ in how readily they name brands at all. Below is how often each one returns a brand-level answer when asked a category question. Meta AI and ChatGPT almost always name brands. Google AI Overviews name a brand in barely more than half of the same prompts, which means your AI Overviews share of voice has a much smaller denominator and deserves its own line, not a blended average.
Source: Trakkr model-divergence dataset, 825,343 category prompts across 8 engines (Aug 2025 to Mar 2026). Brand-answer rate is the share of category prompts in which the engine named at least one brand.
What good cross-engine coverage looks like
What the data actually shows about AI share of voice
Most writing on this topic is theory. Trakkr measures it. We index 20.4 million AI citations across 313,000 sources, sample hundreds of thousands of category prompts a month, and, because trakkr.ai sits behind Cloudflare, we log every AI crawler that fetches our own pages. Three findings from that data should change how you read any share-of-voice number.
1. Citations are volatile, so a snapshot lies
In our citation study of 857,138 reports, 73.5% of brand citations were one-and-done: they appeared once and never again. The median citation lifespan was zero days. The typical brand lost half its AI citations within about 31 days. A share-of-voice figure measured once is a photo of a moving target. The number that matters is the trend across weeks, not any single reading.
2. Being trained on is not the same as being cited
This is the distinction that almost no one measures, and it is the one that decides whether AI search sends you anything. An engine can crawl your site to train a model (it reads you once, in bulk, and you may never surface again) or it can fetch your page live to answer a specific user (which is what actually puts you in an answer with a citation). We can see the difference in our own crawler logs. Over 90 days, AI bots fetched trakkr.ai about 4.4 million times, and the split between "take" (training) and "bring" (live citation) is roughly three to one toward training, but it varies enormously by vendor.
Read across that table and a strategy falls out. OpenAI brings more than it takes: ChatGPT-User pulled our pages 257,403 times live, at 98% success, and OpenAI's live retrieval reads outnumber its training crawl. Anthropic's ClaudeBot takes roughly 70 times what it brings back live. ByteDance and Amazon take heavily and bring nothing yet. The practical lesson for share of voice: the engines where live-retrieval bots read you are the engines where fresh, well-structured content can move your share this quarter. The training-only crawlers reward you on a much slower clock. And the trend is the encouraging part, live retrieval across all engines grew every month we measured, from 151,690 fetches to 285,220 to 309,415.
The Trakkr-only angle
Sources: Trakkr Data citation index (20.4M citations, 313K sources); citation half-life study (857,138 reports, 108,650 citations); first-party Cloudflare AI-crawler logs (90 days, ~4.4M fetches). Figures current as of June 2026.
Explore the open dataset
Live AI brand rankings, citation sources, and crawler activity, free to browse on Trakkr Data.
Five ways an AI share of voice number lies to you
The wrong competitive set
Share of voice is a fraction, and the denominator is your competitors. Include too many irrelevant brands and your share looks artificially small; include too few and it looks inflated. Define the set deliberately, the brands a buyer would actually weigh against you, and revisit it as new entrants appear.
One snapshot, treated as truth
With 73.5% of citations being one-and-done and a 31-day brand half-life, a single measurement is noise. Always read the trend across weeks. A drop from one reading to the next may just be the volatility, not a real loss.
One engine, called "AI"
Models agree only 43.3% of the time. A strong ChatGPT share tells you nothing about Gemini or Google AI Overviews. Measure each engine separately, then weight by where your buyers are.
Counting mentions, ignoring position
Being named first and recommended is not the same as a trailing mention. A raw count flattens that. Use position-weighted share of voice so the number reflects influence, not just presence.
A vanity denominator
If your category only produces a handful of brand mentions across your prompts, a high percentage is meaningless. Check the absolute volume (your visibility) before you trust the share. Share of voice on tiny numbers is a vanity metric.
How to act on your AI share of voice
A metric you only watch is a vanity metric. Here is the loop that turns AI share of voice into growth.
Find where you lose, by prompt
Break your share down to the prompt level. The prompts where competitors win and you are absent are your highest-leverage targets, far more useful than a single category average.
See who is taking your share, and why
For each losing prompt, look at which competitor is named and what source the engine cited to name them. That source is usually the thing to answer better: a review, a comparison, a piece of documentation.
Earn the citation, then watch the live readers
Publish or improve the page that answers the losing prompt, with the structure and facts an engine can lift. Prioritise the engines whose live-retrieval bots actually read you (OpenAI and Perplexity in our data), because those move first.
Measure the trend, not the snapshot
Track share of voice weekly and tie each change back to what you shipped. Given the volatility in the data, only the trend will tell you whether a move worked.
Frequently asked questions
AI share of voice is the percentage of AI answers that mention or recommend your brand, relative to your competitors, for a defined set of category questions. If 10 brands are named 240 times across your tracked prompts and your brand is named 35 of those times, your AI share of voice is roughly 14.6%. It is the AI-search equivalent of traditional media share of voice, applied to the answers ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews now give instead of a list of links.
They answer different questions, so you need both, but they are not interchangeable. AI visibility measures how often you appear in absolute terms ("am I showing up at all?"). AI share of voice measures your appearances relative to competitors ("am I winning the category?"). Visibility can rise while your share of voice falls, because the whole category got more visible faster than you did. If you can only watch one number, watch visibility first while you are climbing off zero, then manage to share of voice once you are reliably present, because share of voice is the truer competitive signal.
Define the category queries that trigger an AI Overview for your market, run them on a fixed schedule, and record which brands the Overview names and in what order. Google AI Overviews name a brand far less often than chat assistants do (a brand-level answer appears in about 56.5% of our tracked category prompts on AI Overviews versus 85.4% on ChatGPT), so your AI Overviews share of voice should be tracked as its own line, not blended into a single cross-engine average.
A ranking is a fixed position on a results page. AI share of voice is your share of a generated answer, where there are no fixed slots. A brand can be named first, buried mid-paragraph, recommended, or cautioned against, and the same question answered twice can return different brands. That is why share of voice is measured by mention frequency and position across many runs, not by a single position number.
Weekly at minimum, because AI citations are volatile. In our citation dataset, 73.5% of brand citations are one-and-done (they appear once and never again), the median citation lifespan is zero days, and the typical brand loses half its citations within about 31 days. A single measurement is a snapshot of a moving target. A weekly trend line is what connects your content and PR work to changes in your share.
Because the engines genuinely disagree. Across 825,343 category prompts, the major models agreed on their answer only 43.3% of the time on average, and gave a perfectly identical brand set just 4.0% of the time. Each model has different training data, retrieval behaviour, and source preferences. Measuring share of voice on one engine and assuming it represents "AI" will mislead you. Track each engine separately, then roll up.
You can get a rough baseline by hand: pick 20 to 50 category prompts, run them across two or three engines, and tally the brands. Trakkr also has a free AI share of voice tool that simulates a run and shows the position-weighted maths. The limits of any manual or free approach are scale, consistency, and history: AI answers drift week to week, so a one-time count goes stale fast. Continuous tracking across engines is what turns share of voice into a metric you can act on.
Track your AI share of voice
See your share against competitors across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. Per prompt, per engine, every week.
Free 7-day trial. Cancel anytime.