What an AI share of voice tool actually measures
An AI share of voice tool asks assistants a set of questions on a schedule and reports how often your brand is named compared with other brands. Everything else is packaging around that loop.
Share of voice is an old marketing idea: of all the mentions in a space, what fraction are yours? In AI search, the "space" is a list of questions, and a "mention" is your brand appearing in an assistant's answer. If you ask a hundred questions and your brand is named in twenty answers, while a competitor appears in forty, your share is lower. Our glossary defines the measure in share of voice in AI search.
The number is only as good as the questions and the method. A tool that asks the wrong questions, or asks each question once, will give you a confident figure that means little. That is why this guide is organized around the choices that determine whether the number can be trusted, not around feature counts.
We name no tools here and make no rankings. The market changes quickly, and the useful skill is knowing what to ask any vendor. If you want a broader overview of the category first, see AI visibility tools. The checklist below is the one to use on a demo call.
The criteria in order of importance
Put the questions and the engines first, honest measurement second, context third and what happens afterward last. The diagram shows that order.
Why this order. If the tool does not let you control the questions or does not cover the assistants your buyers use, nothing else matters, because you are measuring the wrong thing. If it measures the right thing but does not handle variation honestly, the result is noise dressed as a trend. Context, such as which sources were cited and what competitors do, turns a number into something you can act on. And whether the vendor also does the work is a deciding factor mostly after the first three are satisfied.
A common mistake is to choose on the dashboard. Polished charts are the easiest part to build and the least informative. Ask to see the raw answers behind any chart. If they cannot show you the text of each response, you cannot verify anything.
Each tier has its own section below with concrete questions to ask. Take the list to every demo and write down the answers. When two tools tie on the top tiers, then choose on price, export options and support.
Which engines it covers
The tool should cover the assistants your own customers use, and it should report each one separately instead of blending them. Different assistants name different businesses.
At a minimum, ask whether it covers ChatGPT, Google's AI Overviews and AI Mode, Perplexity, Claude and Gemini, and Microsoft Copilot if your buyers use it. Then ask how it asks. Some tools query an assistant's public product the way a person would. Others call a model through a developer interface, which can behave differently from the consumer product, especially on whether web search is on. A figure from a model with no search tells you about what it learned in training, which is not what your customers see when the assistant searches live.
Ask about location. Many questions are local, and answers vary by place. Can the tool run a question as if asked from your service area?
Ask whether results are shown per engine. A blended score can hide that you are named often in one assistant and never in another, which is the most actionable finding. Our page on why AI answers change explains why the engines diverge so much.
Finally, ask how quickly a new engine is added. The set of assistants will keep changing, and a tool stuck on last year's list will drift out of date.
Who writes the question set
You should be able to write, edit and fix the questions, because the question set is what the tool is really measuring. A generated set reflects your marketing language, not your buyer's uncertainty.
Some tools generate questions automatically from your website or category. That is convenient and fine as a starting point, since you may not know what to ask. But a set generated from your own site tends to use your terms, which are not always the words customers use. The result flatters you, because the assistant is being asked about things your site already says.
Ask: can I import my own questions? Can I mark which are the same question in different wording? Can I add questions later without resetting the history? Can I see the full list at any time?
Two further cautions. The set should include questions where you expect not to be named, such as early-stage and comparison questions, or you are measuring only your comfort zone. And it should be fixed. If the vendor changes the set silently, you cannot compare month to month. Our guide to building an AI visibility question set shows how to write one worth testing against.
A vendor who resists letting you see or control the questions is telling you how they expect to be judged.
How often it runs and how it handles variation
Because the same question can produce a different answer each time, a trustworthy tool asks each question several times and reports a rate, not a single yes or no. This is the biggest technical difference between tools.
If you ask an assistant the same question ten times, you may be named in three answers and not in seven. A tool that ran it once and got a miss would say you are absent. A tool that ran it once and got a hit would say you are present. Neither is true. The truth is you appear about thirty percent of the time, and that is the number that can move.
Ask: how many times is each question run per cycle? Is the result shown as a rate or a count? Does it show a range of uncertainty, or at least tell me when a change is too small to trust? How often does it run, daily, weekly or monthly? More frequent runs smooth the noise but cost more, so match frequency to how fast you can act on findings.
Ask also about conditions. Does it run each question in a fresh session, so earlier answers do not bias later ones? Does it record the date and the model version? Do search and location settings stay constant?
If the answers are vague, the figure will bounce around for reasons unrelated to your work, and you will chase noise.
Whether it records sources and exact wording
A tool should save the full text of every answer and the sources cited, because the sources tell you what to fix and the wording tells you how you were described. A count of mentions alone is a shallow reading.
Cited sources show you where the assistant looked. If an outside directory, a forum or a competitor's article keeps appearing in answers for your category, that is a place you can act. Ask whether the tool lists the cited pages, groups them by domain and shows how often each appears. Our glossary entry on share of voice checklist for AI search lists what to capture per run.
Exact wording matters because being named is not always good. An assistant can name you as "a budget option" or "limited selection," and a simple tally will count that as a win. Ask whether the sentence about you is stored and searchable, and whether the tool tags sentiment or leaves you to read them. Even if it tags, read a sample yourself.
Ask also whether it records position. Being first in a list of three may matter differently from being third, though assistants do not rank as search engines do, so treat position as a weak signal.
If the tool only shows a percentage, you will know how you are doing and have no idea why.
Competitor share, exports and ownership of the data
A useful tool shows every business named, not only you, and lets you take your data with you. Both are about being able to act and being able to leave.
Competitor share. The businesses named alongside you are your real competitive set in the assistant's mind, and it often differs from who you consider a rival. Ask whether the tool captures every brand named in each answer or only a list you typed in. The first approach finds surprises. The second only confirms your assumptions.
Segmentation. Can you slice results by question type, such as local, comparison and early research? Can you slice by engine and by location? Without it, you cannot tell which part of your presence is weak.
Exports. Can you download the raw answers, sources and scores as a spreadsheet or by a programmatic interface? Your history is an asset. If you cancel and lose it, you lose the baseline that makes future measurement meaningful.
Retention. How long is history kept? Does it keep every run or only summaries?
Alerts and reporting. Nice to have, and the last thing to decide on. Weekly emails are fine, but they should link to the evidence behind any change.
For the broader process of reading the results once you have them, see our guide on getting recommended by AI, which includes the monthly check.
Does it only measure, or also do the work
Measuring tells you where you are absent, and someone still has to write the pages and earn the mentions that fix it. Decide who that someone is before you buy.
Most share of voice tools are measurement products. They are good at telling you which questions you are missing from. The work that follows is content, technical fixes and outside evidence, and it falls to you or to a service. A tool that reports forty gaps a month produces a backlog. Budget for it, or the subscription buys a recurring reminder.
Some vendors pair the measurement with the work. That can be efficient, since the findings flow straight into writing, but check the same criteria above for the measurement half. Ask what the work includes, how it is delivered and how it is tied to the questions you are tracking.
If you only buy a tool, plan the response. Who writes? Who publishes? Who earns mentions? How long will it take to act on each finding? A realistic answer might be that you can fix five gaps a month, which tells you not to pay for a tool that reports five hundred.
Finally, check the claim of results. No honest vendor can guarantee that you will be named. What they can promise is measurement and work, both of which you can verify.
For what to do with the findings across every engine, see our cross-engine guide to showing up in AI search.
A script for the demo call
Bring ten questions to any demo and ask the vendor to answer them live, with your own brand and your own questions. A prepared demo shows the best case. A live one shows the product.
Before the call, write down five of your real buyer questions, including two where you suspect you are not named. Bring them.
Ask these in order. Can I enter my own questions and see them listed? Show me the raw answer text for one of them. Show me the sources it cited. How many times did you run it, and why that number? What happens if I run it again tomorrow, and how much will the result change? Which assistants are covered, and is search on? Can you show every brand named, not just mine? Can I export all of it today? What do I keep if I cancel? What does your service do after the report, and what is left to me?
Listen for specifics. A vendor with a sound method answers quickly and shows data. A vendor with a weak one talks about proprietary scoring and avoids showing raw answers.
Afterward, score each vendor against the pyramid: top tier first. If a vendor fails a top-tier question, do not let a good price or a nice dashboard move you down the list. Ask for a short trial on your own questions before paying for a year.
How RunOctopus handles this
RunOctopus is a done-for-you service. Its AI, Ollie, researches a business and then writes and publishes pages directly on the business's own website, with a team that plans, writes, reviews and publishes alongside it. It works with Shopify, WordPress, WooCommerce and Wix.
It is a service that does the work as well as measuring it. It reports on where ChatGPT, Perplexity, Claude and Gemini name your business, and the pages that respond to the gaps are written and published for you, with nothing to install, approve, upload or schedule. The first week brings 8 long-form articles, 6 pages and a custom interactive tool built for your niche, and after that 175+ pieces of content a month, written, published and linked.
The 90-day promise is more visibility on Google and in AI answers within 90 days, measured against the 90 days before you joined, or every dollar back.
If you only want a measurement tool, the checklist above is the one to hold any vendor to, including us. To see how your site reads today, use the free site read at app.runoctopus.com/start. Paste your address and Ollie reads the site and shows the plan. No account, no call.