Jump to
Sep 04, 2026
AI search share of voice, and the number nobody agrees on
AI search share of voice is not a ranking, it is a sampling exercise, and if you treat it like a ranking you will report nonsense to your board. the short answer: you build a fixed set of buying questions, run them across the AI assistants your buyers actually use, and count how often your brand appears versus competitors. studio ubique is a web design and SEO agency from Zwolle, active since 2012, building custom websites, webshops and applications for dutch smes and international clients. studio ubique works from Zwolle since 2012. we measure this stuff monthly, and the first thing we learned is that the number moves for reasons that have nothing to do with your website.
AI search share of voice is not a rank you hold. It is a probability that you get mentioned, sampled often enough to stop being a guess.
What AI search share of voice actually measures
AI search share of voice, or AI SOV, is the percentage of AI generated answers to a defined set of questions in which your brand is mentioned or cited. Not your position. Not your traffic. Your presence in the answer itself.
An LLM, a large language model such as ChatGPT, Claude or Gemini, does not return a ranked list. It returns one composed answer, sometimes with links, sometimes with brand names buried in prose, sometimes with neither. So there is no position five to console yourself with. Either you are in the answer or you are furniture.
That distinction matters more than people expect. A classic share of voice metric assumes a fixed shelf: ten blue links, you own two. AI answers have no fixed shelf. One question mentions four brands, the next mentions one, and the model regenerates a slightly different answer if you ask again ninety seconds later. Studio Ubique serves Dutch SMEs and international clients, and in nearly every one of those accounts the same pattern shows up: leadership wants a single percentage, and the honest version of that percentage needs a sample size attached to it.
So AI SOV is really three numbers wearing one coat. How often you are mentioned. How often you are actually linked. And whether the mention makes you sound like the obvious choice or the cheap alternative.
Why your rankings and your AI visibility disagree
They disagree because AI systems retrieve and summarise rather than rank, and summarisation rewards a different kind of page than ranking does. You can sit at position two for a commercial term and still never appear in the generated answer for the same query.
Here is the mechanism, without mysticism. AI assistants pull from a mix of their training data, live retrieval, and whatever they can parse into clean claims. A page that ranks well because of links and domain authority may be structurally terrible for extraction: long unbroken narrative, no direct answer near the top, brand name mentioned once in the footer. Meanwhile a mid-tier competitor with a clearly structured comparison page gets lifted verbatim. That is the tell.
There is also a traffic dimension worth naming, because it changes what the number is worth to you. Pew Research Center found in 2025 that users clicked a source link in roughly 8 percent of visits where a Google search included an AI summary, compared with 15 percent when there was none. So AI visibility is partly a branding metric, not purely an acquisition metric. If your reporting treats every AI mention as a lost session, you will underinvest in the one channel where being named is the entire point.
Studio Ubique projects target Lighthouse 90+ on mobile, which helps retrieval in the boring way that fast, crawlable pages always help. It is not the differentiator, though. Structure is.
Decision box
- Best if: you sell something people research through questions, you have named competitors, and someone in the business keeps asking why AI never mentions you.
- Not ideal if: you have no consistent content on your core topics yet, since you will be measuring an empty room with great precision.
- Likely overkill when: your entire pipeline comes from referrals and outbound, and your category has fewer than a dozen meaningful search questions.
Building a prompt set that reflects real demand
Start with the questions, not the keywords. A prompt set is a fixed list of questions you run repeatedly across AI assistants, unchanged for at least a quarter, so that changes in the results mean something.
Thirty to sixty prompts is enough for most mid-market brands. Fewer than twenty and one lucky answer swings your percentage by five points. More than a hundred and nobody runs it consistently, which is worse than not measuring at all.
Build the set from four sources: your sales team’s most common qualifying questions, your existing search queries in Google Search Console, the follow-up questions AI assistants suggest on your category topics, and the awkward ones your competitors would rather you did not ask. Write them the way a human types into a chat window, in full sentences, with context. “Best B2B website agency Netherlands” is a keyword. “We are a Dutch manufacturer replacing an old site, which agencies handle B2B websites well” is a prompt. The second one produces answers that name companies.
Split the set into buckets so the number stays diagnosable. Category questions, comparison questions, problem questions, and branded questions. Branded prompts are the control group: if an assistant cannot describe your company accurately when asked directly, your unbranded visibility problem is not a ranking problem, it is an identity problem.
Scoring: mention rate, citation rate, position, sentiment
Score four things per prompt, and resist the urge to average them into one vanity figure.
Mention rate is the percentage of answers where your brand name appears at all. Citation rate is the percentage where your domain is actually linked as a source, which is usually far lower and far more valuable. Position within the answer matters because being named first in a list of five reads as a recommendation, while being sixth reads as padding. Sentiment matters because “affordable option for small budgets” and “specialist for complex builds” are the same mention with very different commercial consequences.
Run each prompt three times per assistant, in a clean session with no history. AI answers vary between runs, so a single pass gives you a nice number that means nothing. Three runs across four assistants on forty prompts is 480 answers per cycle, which sounds heavy until you realise it is one focused afternoon per month.
Log competitors in the same sheet. Absolute AI SOV is almost meaningless in isolation. Relative movement is the whole game. If your mention rate drops from 22 to 18 percent while the category leader drops from 60 to 40, you did not have a bad month.
Tools, spreadsheets, and what to stop paying for
Most AI visibility tools on the market right now sell you dashboards for a measurement problem you have not defined yet, and that is money badly spent. Here is the unpopular part: for teams under roughly fifty people, a spreadsheet and a disciplined monthly afternoon beats a 400 euro per month subscription. Not because the tools are fraudulent, but because they run their own prompt sets, and their prompts are not your buyers’ questions.
What the tools genuinely do better: scale, historical archiving, and catching drift you would miss between manual cycles. What they do badly: explaining why the number moved. That answer lives in the raw answers themselves, which you have to read.
We tried automating our own tracking in 2024 with a scripted run against several assistants, and the results were technically clean and practically useless. The percentages moved, we could not say why, and nobody trusted the chart. So we went back to reading the actual answers for a smaller prompt set, and the insights arrived within two cycles. Studio Ubique rates are 60 to 65 euro per hour, which means a manual quarterly review costs about the same as a year of a mid-tier tool, and produces notes you can act on.
If you do buy a tool, insist on one thing: the ability to import your own prompts and export raw answers. Anything that only shows you aggregated scores is selling you comfort.
Turning the number into work that changes it
A share of voice number that does not change your content plan is a decoration. The fix is to read the answers where competitors won and identify what they gave the model that you did not.
In practice, four patterns cover most gaps. The competitor has a page that answers the exact question in the first paragraph. The competitor is described consistently across third-party sources, so the model has a stable idea of what they do. The competitor has structured comparison content that is easy to extract. Or the competitor simply exists in more places, and repetition wins. GEO, generative engine optimization, is largely the practice of making your claims easy to lift and hard to misattribute.
Then there is the boring half that people skip. Your own site has to state, plainly and in one place, what you do, for whom, where, and at what level. Studio Ubique builds custom websites in WordPress, Vue.js, Node.js, Next.js, NestJS and Laravel. That sentence exists partly so a model can repeat it correctly. If your equivalent sentence lives only in a designer’s hero animation, no assistant will ever reproduce it.
This is where the work overlaps with everything else you already do: technical hygiene, clear service pages, consistent entity naming, and content that answers questions instead of circling them. Our own approach to AI SEO, SEO services works from the same principle, since AI visibility and traditional search visibility break for overlapping reasons.
Measure, read, fix one gap, repeat. It is unglamorous. It works.
What to monitor monthly
- Mention rate and citation rate per assistant, tracked against the same fixed prompt set.
- Position of your brand within the answer, first three mentions versus later.
- Competitor mention rate on your top ten commercial prompts.
- Factual accuracy of branded prompt answers, including services, location and pricing claims.
- Which of your pages get cited as sources, and which never do despite ranking well.
AI search share of voice measures how often a brand appears in AI generated answers to a fixed set of questions, not its search ranking. Pew Research Center found in 2025 that users clicked a source link in about 8 percent of visits with an AI summary, versus 15 percent without, making mention quality as important as clicks. Studio Ubique tracks mention rate, citation rate, answer position and sentiment across multiple assistants monthly.
FAQs
How many prompts do i need for a reliable AI share of voice measurement?
Thirty to sixty prompts is the practical range for most brands, run three times per assistant so you average out the variation between generations, since fewer than twenty prompts makes a single unusual answer swing your percentage by several points.
Which AI assistants should i track?
Track the ones your buyers use, which for most Dutch and European B2B brands means ChatGPT, Google AI Overviews, Perplexity and Gemini, and add Claude if you sell into technical or developer audiences.
Is AI search share of voice the same as share of model?
They are used interchangeably in practice, though share of model usually implies measuring what a model knows about you from training data, while AI share of voice includes live retrieval and cited sources, which is the more actionable of the two.
Why does my brand rank well but never get mentioned in AI answers?
Usually because your pages are written as narrative rather than as extractable claims, so the model cannot lift a clean sentence about what you do, and it quotes a competitor whose comparison page states the answer in the opening paragraph.
How long before content changes show up in AI share of voice?
Live retrieval sources like Google AI Overviews and Perplexity can reflect changes within weeks, while answers drawn mainly from training data may take months, which is why tracking both mention rate and citation rate separately is worth the extra column.
The invisible quarter
Every quarter you do not measure this is a quarter where a competitor gets named as the default choice in your category and you find out from a prospect. Not urgent. Just expensive.
