Measuring AI Share of Voice, Step by Step

AI share of voice tells you how often an AI engine names you, compared with the competitors you chose, across the questions you chose. That's all it tells you. The definition and its limits matter more than the number.

Three limits shape everything below:

• Your prompts are a proxy. OpenAI says ChatGPT search typically rewrites a user's question into one or more targeted queries before it searches. Google may fan out across subtopics. You never see the real retrieval queries, so your prompt set samples buyer intent, not engine input.
• Retrieval switches on and off. As of October 2026, ChatGPT searches at the model's discretion, Claude.ai search is a user toggle, and Google AI Overviews often do not trigger. Some answers come from live web results, some from what the model learned before its cutoff. Your number mixes both.
• Brand size dominates. Large brands appear everywhere. A study of AI search citations (arXiv:2507.05301) found the top 20 news sources accounted for 67.3% of citations for OpenAI models. Compare yourself with brands of similar size, or you are mostly measuring who is bigger.

So treat share of voice as a consistent sample you can repeat, not a measurement of the whole market.

The number only means something if the inputs stay fixed between runs. Decide these once and write them down.

The prompt set

Write the questions a buyer asks before they know your name: category questions, comparison questions, problem questions. Leave out questions that contain your brand name, since those test recall of a name you supplied. Save the exact wording. Changing one word changes the answer.

The competitor set

Pick a fixed list of named competitors of comparable size. If a new rival shows up often in answers, note it, but add it only at a planned review, and restart the trend line when you do. Our free way to see which competitors AI recommends helps you build the first list.

The conditions

Run logged out, or with memory and custom instructions off. OpenAI says ChatGPT may use saved memories when it rewrites a search query, so a logged-in run is not reproducible. Use the same location and language each time.

Say you sell project management software and track four competitors across 30 buyer prompts. This is a hypothetical example to show the arithmetic.

1. Pick the engines. List each surface separately: ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews, Google AI Mode, Copilot. Google says AI Overviews and AI Mode may use different models and techniques, so they are two engines, not one.
2. Run every prompt on every engine under your frozen conditions. Save the full answer text and any cited URLs.
3. Record whether the answer showed sources. An answer with citations used live retrieval. An answer without them may come from training data. Tag each one.
4. Count mentions per answer. For each answer, mark which brands from your fixed set are named. Count each brand once per answer, so a long list does not inflate one brand.
5. Calculate per engine. Share of voice = your mentions ÷ (your mentions + all tracked competitor mentions). If ChatGPT named you in 12 answers and your four competitors 48 times in total, your ChatGPT share is 12 ÷ 60 = 20%.
6. Record the presence rate too. Divide the answers that name you by total answers returned. A 20% share can sit next to a low presence rate if answers rarely name anyone.
7. Log the facts stated about you. Note wrong prices, old product names or competitor confusion. A mention that gets your pricing wrong is not a win. See how to measure brand accuracy in AI answers.
8. Store the run with a date and the prompt set version, so the next run compares like with like.

If you want to do this without a tool, the spreadsheet method for free AI share of voice lays out the columns.

A blended score averages engines that behave nothing alike. It hides exactly the problem you need to find.

Otterly's public experiment with two AI-generated sites shows this clearly. Of the 3,934 citations Otterly tracked for one site, 2,718 (69%) came from Microsoft Copilot, while ChatGPT, Gemini, AI Overviews and AI Mode almost never cited it. Google had deindexed the site, and Bing had not. A single blended number would have shown a modest score and hidden a site that was invisible on four engines.

The same test shows a second limit. Microsoft's own AI Performance report credited that site with 42,817 citations, more than ten times what Otterly's prompt tracker logged. The time windows may differ, so it is not an exact ratio. It does show that any prompt sample sees a fraction of real activity.

The next move: report one row per engine. If you must summarise, show the range across engines, not an average. For more on cross-model tracking, read our guide to LLM visibility tracking.

A 30% share does not mean you rank third, and a rising share does not mean you moved up a position. No engine exposes position data for AI answers. As of October 2026, Google's Search Console generative AI report shows impressions only, with no clicks, click-through rate or position.

Counting mentions also differs from counting citations. A mention is your brand named in the answer text. A citation is your URL in the sources. Engines cite differently: Anthropic cites text spans, Perplexity cites whole sources. Our comparison of citation rate vs citation share explains why those numbers do not add up across engines either.

If a vendor reports an AI rank, ask what ground truth it was checked against. The questions in what an AI visibility tool should disclose about its methodology are a good place to start.

Good and bad here describe the measurement and what it tells you to check. No result is a guarantee of the next one.

Signs of a trustworthy result

• The same prompts, competitors and conditions as last time.
• One number per engine, with the retrieval tag recorded.
• Changes you can tie to something real, such as a fixed crawler block or a new review on a site the engine cites.
• Correct facts in the answers that name you.

Signs something needs attention

• Near zero on one engine, healthy on others. Check access first. Is OAI-SearchBot (not GPTBot) allowed in robots.txt? Does your CDN or firewall return a 403 to AI crawlers? Are you indexed in Google and in Bing separately?
• Mentioned only in answers without sources. You may be living on training data. Check whether your key facts appear in the HTML before JavaScript runs.
• High share, wrong facts. Fix the contradictions on your own site first.
• Wild swings between runs with no change on your side. Your sample is probably too small, or your conditions drifted.

When engines cite third-party pages instead of yours, being named on those pages is associated with appearing in the answer. That is an association, not a promised effect.

• Prompt set version saved, wording unchanged
• Competitor list unchanged, or trend line restarted
• Logged out, memory off, same location and language
• Each engine listed separately, AI Overviews and AI Mode split
• Source or no-source tag captured for every answer
• Each brand counted once per answer
• Share and presence rate calculated per engine
• Wrong facts logged with the answer they came from
• Run dated and stored next to the previous one
• Notes on any site changes since the last run (robots.txt, firewall, rendering, pricing page)

Run it monthly on the same frozen set. That's often enough to catch an access break and slow enough that a single odd answer does not drive decisions. Run an extra check after any change that touches crawler access: a CDN migration, a new bot rule, a site rebuild.

Third parties can change your access without you editing anything. From 15 September 2026, Cloudflare set Training and Agent crawlers to blocked by default on ad-displaying pages for new domains and existing free-tier customers. A drop that lines up with a date like that points to access, not content.

Review the prompt and competitor sets once or twice a year, and restart the trend when you change them. For a quick snapshot between full runs, the free checker on this site shows whether engines name you for three buyer questions.