How to Measure AI Share of Voice

ChatGPT and Perplexity can name a SaaS vendor inside an answer without sending a click to its site. Search impressions and click-through rate therefore cannot describe the full competitive exposure in those answers. AI Share of Voice measures a brand's share of the tracked brand mentions across a defined panel of commercial prompts.

Define the denominator

Mention-level Share of Voice is the percentage of tracked category brand mentions assigned to one brand. Count each tracked brand at most once per answer observation, even if the answer repeats its name. If an answer names two tracked vendors, each receives one mention and the denominator grows by two.

The category needs a fixed competitor set. Include the brand being measured and vendors that buyers consider direct alternatives; document every inclusion before collecting answers. Changing the set from five brands to ten can lower SOV without any change in the brand's actual appearance rate.

The prompt panel also needs a fixed boundary. Use commercial prompts drawn from sales calls, comparison searches, and buyer objections. Record the exact wording, market, language, engine, and observation date so the next run measures the same task.

An answer with no tracked brand has zero mentions in the SOV denominator. If every answer in a period has zero tracked mentions, SOV is undefined; report "no tracked brand mentions" instead of 0%. A zero would imply that competitors received mentions while the measured brand received none.

Calculate share and reach separately

Assume a weekly panel produces 300 answer observations across ChatGPT, Gemini, and Perplexity. The measured brand appears in 42 answers, Competitor A in 88, and Competitor B in 70. Those counts produce 200 tracked brand mentions, including answers that name more than one vendor.

Your brand's SOV is 42 ÷ 200 × 100 = 21%. Its mention rate is 42 ÷ 300 × 100 = 14%. Mention rate measures reach across observed answers; SOV measures competitive share among the vendors included in the denominator.

The distinction changes executive interpretation. A brand can increase its mention rate while losing SOV if competitors gain mentions faster. It can also gain SOV during a week when the engine names fewer vendors overall, even though its own mention count stays flat.

Build a repeatable observation unit

One observation is one answer from one engine for one prompt at one recorded time. Store the prompt ID, exact answer, detected brands, cited URLs, engine, model where exposed, location settings, and timestamp. Retain the raw answer so an analyst can audit entity matching errors.

Aliases need an approved dictionary. Map a product name to its parent brand only when the reporting question calls for parent-level SOV; keep product-level counts available for direct comparisons. Exclude ordinary words that match a brand name in another context.

Run repeated observations if the budget allows. AI answers can change between identical requests, so one answer per prompt creates a noisy weekly series. Report the number of repeats and keep it constant across periods.

Prompt selection can distort the metric more than answer variability. A panel dominated by the company's own brand queries will inflate its share; a panel filled with a rival's name will favor that rival. Separate unbranded category prompts from direct comparison prompts and show each segment to executives.

Keep engine reports distinct

ChatGPT, Gemini, and Perplexity have different retrieval and answer behavior. Calculate SOV for each engine before producing a combined figure. A pooled percentage gives more influence to any engine with more observations or more brands named per answer.

If the executive report needs one cross-engine number, fix engine weights in advance. Equal weights mean averaging the three engine-level SOV percentages, regardless of how many brands each engine names. Weights based on measured buyer usage are defensible when that usage data exists and the method remains stable.

Citation share is a separate metric. A response can mention a brand while citing a review of its competitor, and a citation to the brand's domain can appear without a recommendation. Track owned-domain citations and third-party citations beside SOV rather than adding them to the mention denominator.

Referral traffic answers another question. OpenAI says publishers can track visits from ChatGPT through analytics, but a visit records a click rather than every exposure inside an answer. Google's Search Console now reports impressions for its generative AI Search features; those first-party Google reports should sit beside observed SOV instead of being described as unavailable across all AI products. OpenAI publisher guidance · Google Search Console reporting.

Report the visibility gap

A visibility gap is the difference between a competitor's SOV and the measured brand's SOV under the same observation rules. In the example, Competitor A leads by 44% - 21% = 23 percentage points. State the unit as percentage points; calling it a 23% gap changes the arithmetic.

Break that gap down by prompt cluster. A deficit on "best software" prompts may call for different evidence than a deficit on integration comparisons. Attach the answers and cited pages that produced each segment so the report points to work a team can assign.

Weekly collection across hundreds of prompts and three engines creates thousands of records once repeats enter the design. A GEOall Pro tracking dashboard can calculate SOV, mention rate, and visibility gaps from those records while preserving the prompt panel and competitor dictionary. Its output still needs sampling checks for missed aliases, false brand matches, and answer changes caused by collection settings.

Version the panel whenever the team adds a competitor or rewrites a prompt. Keep the old series intact and start a new baseline under the revised definition. That version boundary prevents a change in measurement design from appearing as a gain or loss in AI Share of Voice.