AI Share of Voice: Benchmark Competitors in ChatGPT and Perplexity

AI Share of Voice: Benchmark Competitors in ChatGPT and Perplexity

Bottom line: AI share of voice is the percentage of brand mentions you own inside answers from ChatGPT, Perplexity, and Google AI Overviews for a fixed set of buyer prompts. You benchmark it by running the same prompt set across each engine on a repeating cadence, logging every brand named, and dividing your weighted mentions by the total.

Most teams measure AI share of voice the lazy way: they type their brand name into ChatGPT, see a flattering paragraph, and call it a win. That tells you nothing. The question that matters is what the engine says when a buyer doesn't know you exist yet - when they ask "best tools for X" and an LLM returns five names. If yours isn't one of them, you have a visibility problem that no amount of keyword ranking will surface. We see this split constantly in audits: brands that rank page one on Google but appear in fewer than one in five relevant AI answers.

Key Takeaways:

• —AI share of voice measures your slice of brand mentions across a fixed prompt set, not your position on a single query - it's a category metric, not a keyword metric.
• —A credible benchmark needs three inputs: a prompt set, the engines you run it against, and a position weight so a first mention counts more than a buried one.
• —Vendor guides disagree on prompt-set size - averi.ai-your-category-in-chatgpt-perplexity-and-google-ai-mode) suggests 20-50 buyer-intent queries while pepper.inc runs 150-300 - so start small and expand where your category is noisy.
• —The same brand often scores well on ChatGPT and poorly on Perplexity; measure per engine or you'll miss the gap.
• —Reading the scorecard by prompt category - discovery, comparison, alternative-to, problem-based - tells you where you're losing, which is the only thing you can act on.

What is AI share of voice?

AI share of voice is your brand's share of all brand mentions an AI engine produces for a defined set of prompts, measured over time. Think of it as traditional share of voice - the slice of category conversation you own - ported to generative answers instead of ad impressions or SERP real estate.

The mechanics are different from classic SEO. A blue-link ranking is deterministic: you're at position 4 for a query, and everyone searching sees position 4. An AI answer is generative and probabilistic. Ask ChatGPT the same question twice and the brand list can shift. That volatility is exactly why a single check is worthless and a repeated, averaged benchmark is the only honest number. One vendor glossary frames the measurement cleanly: record whether your brand appears, whether competitors appear, and whether your own site is cited as a source (joinmassive.com).

This matters more every quarter. AI Overviews now appear on 88% of the commercial keywords Semrush tracked, up from 46% earlier in the cycle (Semrush). When nearly nine in ten commercial searches surface an AI answer, the brands named inside that answer own the consideration set before a user ever reaches a website. If you want the deeper mechanics of getting pulled into these answers, our generative engine optimization guide covers the citation signals in detail.

Why does your Google ranking not predict your AI share of voice?

Here's the uncomfortable finding from audit after audit: ranking position and AI mention rate correlate far less than teams assume. A brand sitting at #1 for its head term can be invisible for the unbranded, intent-rich prompts buyers actually type into ChatGPT - because AI engines synthesize across many sources, weight third-party mentions heavily, and favor pages structured as clean, liftable answers.

Two forces drive the gap:

• —Source diversity. AI engines don't just read your site. They pull from Reddit threads, review roundups, comparison pages, and expert articles. A competitor with weaker Google rankings but heavier third-party coverage can out-mention you. We dig into this in brand mentions vs backlinks.
• —Answer structure. Engines preferentially lift content that's already shaped like an answer - tables, direct definitions, numbered steps. A page that ranks on raw authority but reads like a brochure gets skipped.

The practical consequence: you cannot infer AI share of voice from a rank tracker. You have to measure it directly, with its own instrument.

How do you build an AI share of voice scorecard? (The formula)

This is the part most guides skip. They tell you to "count mentions and divide," which collapses the moment a brand is named first in one answer and buried at the bottom of another. A raw count treats those as equal. They aren't.

Here's the scorecard formula we use. It has three inputs - prompt set × engine × position weight - and it rewards prominent mentions over mere presence.

Step 1 - Assign a position weight to every mention.

Step 2 - Score each prompt, per engine. For one prompt on one engine, your prompt score is the sum of your brand's weighted placements. Run each prompt 3 times and average, because generative answers vary between runs.

Step 3 - Compute weighted share of voice. Across the full prompt set on one engine:

AI SoV (engine) = (sum of your weighted mentions) ÷ (sum of all brands' weighted mentions) × 100

Step 4 - Report per engine, then blended. Keep ChatGPT, Perplexity, and Google AI Overviews as separate columns, because a single blended number hides the engine where you're losing. A worked example from one vendor shows the simple version - 300 of 1,000 total mentions equals 30% share (texta.ai) - but the unweighted count would have scored a first-place mention and a throwaway aside identically. The weighting is what makes the number decision-useful.

Why bother with weights? Because the output isn't a vanity figure - it's a work queue. A brand with 35% unweighted share but 18% weighted share is getting mentioned but not recommended, and that's a completely different fix than being absent.

What does a 20-prompt benchmark look like? (Sample dataset)

A benchmark is only as good as its prompt set. The trap is loading it with branded and head-term queries you already win. A real benchmark is built from the unbranded, buyer-intent prompts where the category is genuinely contested. Below is a sample 20-prompt structure, organized into five categories so you can read gaps by type of query - the diagnostic layer almost everyone misses.

Run all 20 across three engines, 3 runs each - 180 data points. Then read the scorecard by category, not just the top-line number. The pattern we see most often in audits looks like this:

1. Strong on trust prompts, weak on discovery. The engine speaks well of you once it knows the name - but you're absent when buyers ask open-ended. Fix: unbranded comparison and "best-of" content, plus third-party coverage.
2. Strong on discovery, weak on comparison. You get listed but lose the head-to-head. Fix: build the comparison and "vs" pages AI engines quote.
3. Strong on ChatGPT, absent on Perplexity. Perplexity leans harder on freshly cited, linkable sources. Fix: earn citations and keep pages current.

Reading gaps by category is the whole point. A single "22% share of voice" tells you nothing you can act on. "22% overall, but 4% on discovery prompts and 40% on trust prompts" tells you exactly where to spend the next quarter. You can automate the per-engine, per-prompt tracking with SEO Magics' AI citation tracker so you're not running 180 prompts by hand every month.

How many prompts do you need, and how often should you run them?

Short answer: start with 20-30 and expand only where the data is noisy. Vendor guidance is all over the map - averi.ai-your-category-in-chatgpt-perplexity-and-google-ai-mode) recommends 20-50 buyer-intent queries run on a steady cadence, while pepper.inc argues for 150-300. Both can be right; the correct size depends on how varied your category is.

The variable that actually matters is consistency, not volume. Because generative answers drift, a benchmark is only comparable to its past self if you hold the prompt set, the run count, and the cadence fixed. Change the prompts mid-stream and you've reset the baseline. We generally run a core set every two weeks and review the trend monthly - frequent enough to catch movement, stable enough to trust the line. For building a durable query set, our note on prompt-based rank tracking goes deeper on construction.

How do you turn the scorecard into action?

A benchmark that doesn't change what you do next week is a dashboard, not a tool. Here's the sequence we run once the scorecard is in:

1. Find the lowest-scoring category. That's your constraint. Everything else is noise until it's fixed.
2. Pull the actual answers. Read what the engine said about the winners. What source is it citing? A review site? A comparison page? Their own docs?
3. Match the source type. If discovery prompts cite roundup articles you're absent from, your job is earning placement in those roundups - not rewriting your homepage.
4. Fix structure where you're present but buried. A 0.3-weighted mention usually means your content exists but isn't shaped to be lifted. Add the direct answer, the table, the comparison.
5. Re-benchmark in 30 days. Hold the prompt set fixed and watch the weighted score, not the raw count.

This is the same diagnosis-first logic Semrush describes for AI search strategy - you don't rewrite content blind, you fix what the engine can't reach or keeps getting wrong. If the engine is describing your brand incorrectly, that's a separate repair job covered in our brand correction playbook.

How We Assessed This

The scorecard formula and the 20-prompt structure in this article come from how we actually run AI visibility audits for growth-stage clients, not from a published study. For each engagement we build a prompt set from the client's real unbranded buyer language, run it across ChatGPT, Perplexity, and Google AI Overviews on a fixed cadence, and log every brand named plus whether the client's own domain was cited as a source. We average multiple runs per prompt because generative answers vary between sessions, and we apply position weighting so a top recommendation outscores a passing mention. The external coverage figures here are sourced to Semrush's AI SEO tracking and the vendor measurement guides linked throughout - we cite the trend and name the source rather than inventing precision, because several of these numbers conflict across methodologies. Where we reference audit patterns ("most sites we see"), those are qualitative observations from retainer work across 12-month optimization cycles, not a sampled dataset. Tooling on our side spans Ahrefs and Semrush for the traditional-search baseline and our own citation tracking for the AI layer.

Frequently asked questions

What is a good AI share of voice score?

There's no universal benchmark, because it's relative to your category and prompt set. The useful read is directional: are you gaining on competitors over successive runs, and is your weighted score climbing faster than your raw count? A brand mentioned prominently in a quarter of relevant answers in a crowded category is in strong shape; the same number in a two-player market is weak.

How is AI share of voice different from traditional share of voice?

Traditional share of voice measures your slice of paid impressions or organic SERP visibility - deterministic and query-level. AI share of voice measures brand mentions inside generative answers, which are probabilistic, synthesized from many sources, and vary run to run. It's a category-level metric read across a prompt set, not a position on one keyword.

Can you measure AI share of voice for free?

Yes, manually. Build a 20-prompt set, run it across ChatGPT, Perplexity, and Google AI Overviews three times each, and log the results in a spreadsheet using the position weights above. It's tedious but costs nothing. Paid tools exist to automate the runs and tracking at cadence, with listed prices ranging widely (averi.ai-your-category-in-chatgpt-perplexity-and-google-ai-mode)).

Why does my brand appear in ChatGPT but not Perplexity?

Perplexity leans more heavily on freshly cited, linkable web sources, while ChatGPT draws on a broader training-plus-retrieval mix. A brand with strong name recognition but thin recent citations will often score well on one and poorly on the other - which is exactly why you report per engine instead of blending into a single number.

How often should I benchmark AI share of voice?

Run a fixed core prompt set every two weeks and review the trend monthly. Frequency matters less than holding the prompt set, run count, and cadence constant, because changing the inputs resets your baseline and makes the trend line meaningless.

Does getting cited as a source matter more than being mentioned?

They're different wins. A mention puts your brand in the consideration set; a cited source link both signals authority to the engine and can send referral traffic. In the scorecard, a citation is an additive bonus on top of the mention weight - our guide on tracking AI citation share over time covers the referral side.

Benchmark your category, then close the gap

If you only take one thing from this: measure AI share of voice directly, weight it by prominence, and read it by prompt category - because that's the version of the number that tells you what to build next. A blended "we're at 20%" is a vanity metric; "we're invisible on discovery prompts and winning on trust" is a roadmap.

Want the benchmark run for you? Start with our free SEO audit tool to pressure-test the technical foundation AI engines need to reach your pages, or if AI visibility is the priority, our AI SEO service is built specifically to get growth-stage brands cited inside ChatGPT, Perplexity, and Google AI Overviews - not just ranked on blue links. When you're ready to map your category and close the gaps, book a strategy call and we'll build the prompt set with you.