Share of Voice

Search dashboards have trained us to measure visibility through rankings and clicks. Those same signals miss a growing touchpoint where buyers form impressions: conversational AI. When a buyer asks an assistant for recommendations and your competitor is the one named, traditional SOV numbers look safe while your actual visibility erodes. Measuring brand presence inside generated responses gives you a clearer read on influence in moments that often precede a click.

AI Share of Voice measures how often a brand is mentioned in assistant responses within your target topic set. Call it AI SOV, or just SOV after this, but treat it as a different signal from SERP share. Below I outline the definition, a practical way to measure it, how it differs from legacy SOV, and how to use the metric to set priorities and goals.

What is AI Share of Voice?

AI Share of Voice captures the proportion of brand mentions in generated answers for a defined set of queries. At its simplest the formula reads: Share of Voice = Your brand mentions / Total category mentions in AI responses. If you query assistants 100 times about "project management for remote teams" and your brand appears in 25 of those responses, your SOV is 25%.

Mentions can be explicit brand names, product names, or clear aliases people use when referring to your product. Count citations separately if you want a signal for perceived authority. For many teams, two buckets are useful: direct mentions inside the assistant text, and mentions that appear as citation links or suggested resources. Track both to understand raw presence and attributable referral potential.

How to Measure AI SOV

Measurement has to be pragmatic because there is no universal standard yet. The core workflow follows four ordered steps:

1. Define the query set. Start with 30-200 seed queries across intent categories: awareness, comparison, purchase. Use SERP query logs, keyword research, and top customer questions from support as inputs.
2. Run queries across platforms. Include the major assistants you care about, different model settings if available, and repeat queries to capture session variance. Capture the full response, model version, prompt template, and any citations.
3. Detect mentions. Use exact matching, normalized aliases, and named-entity recognition with fuzzy matching for paraphrases. Flag whether a mention is in the answer body, a citation, or a suggested step.
4. Calculate SOV. Aggregate counts by topic and platform, then compute percentage share. Report per-platform SOV and a weighted aggregate if some sources drive more traffic for you.

Practical guidance on sampling: aim for at least 200 queries per topic when variance is high, or 500+ when you need confidence across multiple assistants. Control prompts: use a stable user persona and temperature where available. Record date, time, cookies, and any chat history. For noisy signals, bootstrap with multiple runs and report confidence intervals rather than a single point estimate.

AI SOV vs Traditional SOV

The table highlights why you cannot treat SOV numbers from assistants like a straight swap for SERP share. Assistants can produce a concise recommendation without sending a click. That creates influence without measurable referral. At the same time, assistants often synthesize multiple sources and surface names that were rarely visible in traditional SERPs. Volatility matters: model updates can change who gets named overnight, so track trends not one-off snapshots.

Using AI SOV for Strategy

Treat SOV as an early-warning and prioritization signal. Low share for a high-intent topic signals a content or positioning gap, while a decline in your SOV for an established category is a competitive alert. Aim for two operational outcomes: protect brand presence where you lead, and build presence where you want to win.

• Benchmark and cadence: Establish a baseline by assistant and topic, then measure monthly. Report both raw SOV and a rolling 90-day trend to smooth model noise.
• Content and prompt focus: Create concise, answer-first content that maps to the exact question language assistants use. Include unambiguous product names, short value statements, and example use cases so models can pull the correct token sequences.
• Signal plumbing: Make sure your public docs, FAQ, schema, and canonical pages are clear, factual, and easy to cite. Where possible, publish short-form reference pages that answer the single question an assistant might be asked.
• Product positioning: If assistants miss your product because the phrasing differs, introduce common aliases in headings and lead lines. Update messaging in a way that reads naturally to humans and matches query patterns.
• Experimentation: Run prompt conditioning tests and controlled content pushes for 6-12 weeks, track SOV movement, and tie changes back to downstream metrics like assisted conversions.
• Risk management: Monitor model updates and spikes in competitor mentions. Keep a watchlist of high-risk topics where an assistant often recommends a competitor, and assign owners to respond with targeted content or PR.

Start with a small, high-value topic set and build process muscle. Over time you’ll refine which queries predict business outcomes and which assistants matter most for your buyers. Use SOV not as a vanity number, but as a signal to change what you publish and where you focus product storytelling.