How to Measure AI Share of Voice

To measure AI share of voice, count how often AI assistants mention your brand across a fixed panel of prompts, then divide by the mentions earned by you and your tracked competitors combined. You don't need to buy a platform: a prompt panel, a spreadsheet, and a repeatable routine produce a defensible number, and this guide covers all three.

Key takeaways:

• Mention-based AI share of voice = your mentions ÷ all tracked-brand mentions in AI answers, × 100
• Discovery prompts (no brand named) test category visibility; branded prompts test what AI says once you're named
• Single runs are noise. Fixed panel, repeated runs, read the trend
• Tools earn their fee when your panel outgrows a spreadsheet, not before

What Is AI Share of Voice?

AI share of voice is the percentage of AI-generated answers in your category that mention your brand, measured against a tracked set of competitors. If your brand appears in 12 answers across a panel, and the tracked brands together earn 100 appearances, your share is 12%. Count each brand once per answer, however many times it repeats, or verbose answers warp the maths.

The metric comes in two versions. Mention-based share counts how often your brand name appears in the answer text. Citation-based share counts how often your domain appears as a linked source, and it's the harder one to move because assistants lean on third-party sources more than brand sites.

Track both. A brand that's mentioned but rarely cited is drawing on sources it doesn't control, and those sources can change their mind. If the wider vocabulary is new to you, our plain-English guide to GEO, SEO, and AEO sorts the acronyms first.

Are You Measuring the Right Prompts?

Most teams measure AI share of voice with the wrong prompts. They ask the assistant about their own brand, get a flattering answer, and file the screenshot as evidence. A branded prompt only tests what the assistant says once your name is already in the conversation. It proves roughly as much as looking up your own address to check whether maps work.

Discovery prompts test unprompted category visibility: a buyer describes a problem and names nobody, and the assistant either surfaces you or it doesn't. "Best way to track brand mentions in ChatGPT" is a discovery prompt. "What do you know about [your brand]" isn't. Practitioner citation scans keep finding brands that dominate branded prompts while staying invisible on discovery prompts, and a blended average hides the gap.

So split the KPI. Score branded share and discovery share separately, because they move with different work: branded share follows PR and brand strength, discovery share follows citations and earned coverage. A competitor gap analysis for AI search tells you which competitors own the discovery prompts you're missing.

One large ChatGPT citation analysis, ChatGPT only, offers good clues about what gets extracted:

What the citation data shows:

• 44.2% of ChatGPT citations come from the first 30% of a page's content, documented in an analysis of 18,012 verified citations covered by Search Engine Land
• Cited passages are nearly twice as likely as uncited ones to use clear, definitional language, per the same analysis
• 78.4% of citations tied to questions come from headings, one likely reason question-led structure keeps appearing in cited pages

How Do You Measure AI Share of Voice?

You can measure AI share of voice in an afternoon with five steps and zero procurement meetings.

1. Define and freeze the competitor set. List the 4 to 6 brands a buyer would genuinely weigh against you. This set is the denominator of every future score: you're building tracked competitor share of voice, not a category-wide census.
2. Build a prompt panel. Write 20 to 30 prompts a stranger with your problem would type. Mine sales calls and support tickets for the phrasing. Keep roughly 80% discovery, 20% branded.
3. Run the panel across at least three assistants. ChatGPT, Perplexity, and Gemini frequently disagree with each other, so one engine gives a false picture.
4. Score each answer three ways. Mentioned (your name appears), cited (your domain is a source), recommended (the assistant tells the buyer to pick you). One row per prompt per engine.
5. Compute three numbers, not one. Mention share: your appearances ÷ total tracked-brand appearances. Citation share: same maths on domain citations. Recommendation rate: prompts where you're the pick ÷ all prompts. Log the date, split branded from discovery, repeat.

Here's the kind of panel nobody publishes, for a B2B analytics tool:

• "How do I know which marketing channels drive revenue?"
• "Tools to connect CRM data to ad spend"
• "Best analytics setup for a 50-person SaaS company"
• "Software to prove marketing ROI to the board"

Swap the subject and the shape holds. The panel is the asset; guard it like one.

Why Do Results Change Between Runs?

The same prompt gives different answers on different days, and sometimes within the same hour. Assistants are probabilistic systems: they sample, they personalise, and they refresh their sources. Run your panel twice on a Tuesday and the numbers will differ. This is normal, and few vendor landing pages mention it.

Treat measurement like repeated testing rather than accounting. You can't observe the full population of AI conversations, so a frozen panel is your consistent yardstick: not representative of everything buyers ask, but identical run to run. Run it several times per cycle, average the runs, and report a range.

The skeptics have a point: no tool measures total visibility, only a repeatable signal, and honest measurement admits the difference. Our working rule, borrowed from measurement-minded practitioners: if a score's week-over-week movement is smaller than its run-to-run variance, the score is noise dressed as insight.

Which Tools Track AI Share of Voice?

AI share of voice trackers come in three broad approaches: a DIY spreadsheet panel, SEO-suite add-ons like Ahrefs Brand Radar and the Semrush AI Visibility Toolkit, and dedicated multi-engine platforms. Start with the spreadsheet, and graduate when it starts costing you mornings. The buying criteria that survive practitioner contact: citation data rather than mention counts alone, multi-model coverage, and a rank against competitors rather than a score that can rise for everyone at once.

One caution: vendor scores aren't interchangeable. Each platform has its own prompt universe and maths, Ahrefs scores on estimated impressions for instance, so pick one method and stick to it for trends.

We run our own measurement through Brand Radar on top of a manual panel, and the cited-domain view has been the humbling part: community platforms and third-party sites keep beating brand-owned pages to the citations in our category. That pattern isn't unique to us.

Building the panel is the unglamorous half of the work. Our GEO playbook for B2B companies covers the content patterns we use to earn citations once the baseline exists.

What Counts as a Good Score?

No one has a trustworthy universal benchmark yet, so treat any confident "aim for 25%" advice with suspicion. The numbers in circulation come from vendor content, each computed with different prompts and different maths.

Classic advertising research explains why marketers chase the metric at all. B2B brands whose advertising share of voice runs 10 percentage points above their market share grow about 0.7 percentage points a year, based on IPA effectiveness data analysed for LinkedIn's B2B Institute. That relationship covers advertising, not AI answers. Borrow the framing, not the coefficient.

Two comparisons beat any absolute threshold. First, your trend: a brand at 8% climbing for three straight cycles is in better shape than one at 30% drifting down. Second, your gap to the competitor who owns your category's discovery prompts.

Research backs the humility. A 2025 University of Toronto preprint found assistants systematically favour earned media, third-party reviews and institutional sources, over brand-owned content. Your score is partly a measure of what other people publish about you, which is exactly why it moves slowly and why the trend is the only honest reading.

How Often Should You Measure?

We recommend monthly measurement for most B2B brands. It's frequent enough to catch directional change and infrequent enough to stop you chasing run-to-run noise while topical authority builds in the background.

Move to weekly only during a launch or PR wave, when you need fast feedback. Daily measurement of a monthly-moving metric produces charts, meetings, and nothing else.

Want a second pair of eyes on your baseline? We help B2B teams build measurement panels and the content programmes that move them. No charge for the conversation: talk to Tenpoint Labs.

FAQs

Start this week:

1. Pick the 5 competitors a buyer would weigh against you
2. Write 20 prompts, 16 discovery and 4 branded, from sales-call language
3. Run them through 3 assistants
4. Score mentions and citations in one sheet
5. Diary the identical run for next month

That's a defensible number for the slide, before anyone opens a procurement thread.