Aizii Learn · AI Search
← Back to Learn

Share of voice in AI search: how to measure it

Share of voice in AI search is a working measurement, not a vendor metric. No search engine and no AI assistant publishes a share of voice number for your brand, and no vendor publishes a formula for one. What this guide gives you is a method you can run yourself: ask a fixed set of prompts, count how often each brand is named or cited, and divide your count by the total count. The result is a share of the answers you tested, for the prompts you chose, in the assistants you tested. Say that plainly every time you report the number, because a share without its prompt list is a number without a meaning.

The method works because AI answers are countable. An answer either names your brand or it does not. It either cites your page or it does not. It either recommends your product or it recommends a competitor's. Count those three things separately, because they measure different kinds of presence. A brand can be named often and cited rarely, which means shoppers hear the name but cannot reach the site. A brand can be cited often and recommended rarely, which means the pages are found but the product loses the comparison.

Where it goes wrong

Share of voice numbers go wrong when the method behind them is hidden, or when the count measures the wrong thing.

  • The prompt list is never stated. A share of 40 percent means nothing if nobody knows which questions were asked. A list of brand name prompts gives a high share, and a list of category prompts gives a low one, and both can be reported as the same number.
  • The wording changes between rounds. Each round uses different phrasing, so a change in the share cannot be tied to a change in the brand. The prompts must stay fixed, or the trend is noise.
  • Only one assistant is counted. A share measured in ChatGPT alone is a ChatGPT share, not an AI search share. Each assistant reads different sources, so the share differs by engine, and the difference is itself useful information.
  • Mentions and recommendations are blended. A brand named in a list of five counts the same as a brand recommended as the top pick, and the number flatters the brand that is often listed and rarely chosen. Count naming, citing, and recommending as three separate shares.
  • Training answers and search answers are mixed in one count. A training based answer that cannot move yet is counted next to a search based answer that moved after a recrawl, and the trend reads as flat when the search share actually improved. Keep the two types in separate columns.
  • The total is the wrong total. The share is divided by the number of prompts, when it should be divided by the number of brand mentions across all brands in those answers. A prompt with no brand named at all is not a share for anyone, and counting it in the total shrinks every brand's share for no reason.
  • A tool's number is reported as a vendor number. A visibility tool such as Profound reports its own share of voice metric, calculated from its own prompts and its own method, as documented in its own analytics pages. That is a tool metric, not a vendor metric, and it should be reported with the tool's name attached. Do not present any tool's share as the share of voice in AI search.

The steps

Step 1. Fix the prompt list and the assistants. Write twenty to thirty prompts that cover the questions shoppers ask: your brand name, your main product, the category without any brand name, a comparison question, and a price question. Test ChatGPT, Perplexity, Gemini, and Claude, since each reads different sources. Keep the list fixed for the next round, and change it only when you add a product or retire one. Write the list down next to every number you report, so the number can be checked.

Step 2. Run the prompts the same way every time. Ask each prompt with the same wording, in a fresh conversation, and save the full answer and every citation. Do not edit a prompt mid round to get a better answer. A changed prompt is a new test, not a retest. Run the full list on a schedule you can keep, such as once a month, and more often around a launch or a price change.

Step 3. Count three things in every answer. For each answer, record whether your brand was named, whether your page was cited, and whether your product was recommended as a pick. Record the same three things for each competitor you track, using the same definitions. A brand named in a list counts as named, not as recommended. A page cited counts as cited only when the citation points at that brand's own page. A product recommended counts only when the answer picks it, not when it merely lists it.

Step 4. Calculate the share from the counts. For each of the three measures, divide the brand's count by the total count across all tracked brands in the same answers. That gives three shares: share of namings, share of citations, and share of recommendations. Report all three, with the prompt count, the assistant list, and the round, so the number can be checked. A share that moves in citations but not in recommendations tells you the pages are found but the product loses the comparison, and the fix is the page, not the access.

Step 5. Compare the share by assistant and by question type. Break the counts down by assistant, and by brand name prompts versus category prompts. A brand can hold a high share in brand name prompts and a low share in category prompts, which means shoppers who already know the brand find it, and shoppers who do not know it never hear of it. A brand can hold a high share in one assistant and a low share in another, which points at a crawler or index gap for that engine. The breakdown is where the fix lives, and a single blended number hides it.

Step 6. Track the trend, not the single round. One round gives a snapshot. Three rounds give a trend, and the trend is what tells you whether a fix worked. After a fix, such as a corrected page or an access change, run the same prompts again and compare the same three shares. A citation share that rises after a page fix is evidence the fix worked. A recommendation share that stays flat after the same fix is evidence the product still loses the comparison, and the next fix is the page content, not the access.

Step 7. Act on the largest gap first. Pick the largest gap among the three shares and fix that one. A low citation share usually means access or a missing page, so check the crawlers and the sitemap. A low recommendation share with a high citation share usually means the page is found but the product loses on price, range, or clarity, so improve the page. A low naming share in category prompts usually means the brand is absent from the comparison pages and roundups the assistants read, so publish the comparison content those answers need. Note the fix in the log next to the round, so the next trend can be tied to it.

The machine-readable piece

This log format is your own tracking sheet, not a vendor format. No vendor defines how you should count share of voice, so this file is a working template you can copy into a spreadsheet. The columns are the three counts from step 3, plus the citation that lets you check a count. Keep one row per prompt per assistant per round, and calculate the shares from the rows.

prompt,assistant,brand_named,page_cited,product_recommended,citation_url
"best running shoes for flat feet",ChatGPT,yes,yes,yes,https://[yourdomain.com]/products/example-running-shoes
"best running shoes for flat feet",ChatGPT,yes,no,no,https://[competitor.example]/running-shoes
"best running shoes for flat feet",Perplexity,no,no,no,https://[competitor.example]/running-shoes

This formula is the working calculation from step 4, written as plain arithmetic you can use in a spreadsheet. It is a working measurement, not a vendor metric, and no vendor publishes it. The total is the count of all tracked brands named in the same answers, not the count of prompts.

share of namings = your brand namings / all tracked brand namings
share of citations = your page citations / all tracked brand citations
share of recommendations = your product recommendations / all tracked brand recommendations

Use the same formula for each assistant separately before you blend the assistants together. A blended share across assistants hides the engine where you are weakest, and the weakest engine is usually the one with the fixable access gap.

What it looks like when it works

On the page, nothing looks different to a human reader. The change is in the log: the same prompts run on the same schedule, the same three counts are recorded every round, and the trend shows whether each fix moved the share it was meant to move. A citation share that rises after a page fix is visible in the next round. A recommendation share that stays flat after the same fix is visible too, and the next fix is the page content, not another access change. The number you report always carries its prompt count, its assistant list, and its round, so anyone can check it.

Say you run a running shoe store, and you track your brand against two competitors across twenty five prompts in ChatGPT, Perplexity, and Gemini. In the first round, your share of namings is high on brand name prompts and low on category prompts, your share of citations is low everywhere, and your share of recommendations is lower still. The breakdown points at access: the search crawlers cannot read your product pages reliably. You fix the robots file and the firewall, confirm the crawlers return success responses, and request indexing of the key pages in Search Console. In the next round, your share of citations rises in every assistant, and your share of namings rises in category prompts. Your share of recommendations rises only slightly, so you compare your product page with the competitor page that keeps winning, and you improve the price clarity and the size range on your page. In the round after that, the recommendation share moves.

The fundamentals

These mechanics are shared across every engine and covered in full in their own guides:

What this guide does not cover

This guide covers measuring share of voice in AI search as a working method, with a fixed prompt list, three separate counts, and a trend you can check. It does not cover a vendor share of voice metric, because no vendor publishes one, and this guide does not present its method as one. It does not cover correcting a wrong answer, which is a different job with its own guide, or monitoring mentions, which has its own guide too. It does not cover the full technical audit that checks every crawler and every page, or the transaction layer that lets an agent complete a purchase. Those are separate jobs, and this guide is the measurement, not the whole program.

Check your own store now

Run the free scan. One click, no signup, and it shows whether AI crawlers can read your store today: aizii.ai/aeo-scan

Run the free scan

Sources

Note: links marked specific point to the page that supports the claim in this guide. Links marked general docs point to the vendor's general docs section for that topic, not to a page that proves every step. Check the vendor docs for the current path before relying on a menu name or click path in this guide.