A keyword rank tells you where a page sat on a search results list. A generative answer tells you which names were allowed onto the buyer’s small shelf.
The first thing I noticed was the silence around the old champion keyword. A B2B software integrator near Lyon still held decent search positions for several implementation phrases. The team had not lost the classic SEO story. Yet when the prompts sounded like a factory manager looking for help with an industrial system rollout, AI answers kept naming two competitors, a vendor directory and one consultancy that barely appeared in the keyword report.
This is a composite scenario, built from several B2B service and software cases. The firm had French case studies, English vendor references and a few trade-media mentions. The awkward detail: one AI answer named the company correctly, then described it as a reseller. That was not a small wording issue. For an implementation specialist, being read as a reseller moves the buyer into the wrong mental shelf. The keyword rank did not show that. The generative answer did.
Rank measures a page; answers ration attention
Keyword rank comes from a list. It has its own complications, but the shape is familiar: query, page, position, volume, maybe location. A founder or marketer can argue about whether position three is good enough. At least the object is clear. The page has a place.
Generative answers behave differently. The engine chooses a handful of names, arranges them in prose, cites some sources, skips others and often compresses the category into a short explanation. The buyer may see five recommended firms, or three, or one neat paragraph. The business does not merely “rank.” It is either included in the answer space or left outside the frame.
That is why share of voice matters more than keyword rank for this kind of measurement. I still respect keyword data. It shows demand language and page visibility in search. But it cannot tell me how often the business is named when an answer engine writes the buyer’s shortlist.
Generative share of voice is the proportion of repeated AI answer space a business receives in a defined category, because visibility now depends on being named, cited and described among alternatives.
This definition has three brakes built into it. “Repeated” prevents the single-screenshot mistake. “Defined category” prevents vague brand monitoring. “Among alternatives” keeps competitors in view. Without those brakes, share of voice becomes a soft phrase for feeling visible.
The metaphor I use in audits is an answer shelf. The shelf is small. The engine stocks it with names. Some names get eye-level placement, some sit on the lower plank, some are not stocked at all. Keyword rank tells me how your own page performs in a search aisle. Generative share of voice tells me how often your name reaches the shelf the buyer actually reads.
Count appearances, but do not stop there
The simplest version of share of voice is mention share. You run a set of prompts several times and count how often each business appears. If there are ten runs and your company appears in four, the mention share is visible enough to discuss. If a competitor appears in nine, the gap is hard to ignore.
That first count is useful. It is also incomplete. In AI answers, a mention can be shallow. The business may appear at the end of a list without a cited source. It may be named in a paragraph that gives no reason to consider it. It may be described through the wrong service category. A raw mention count gives every appearance the same weight, and that is rarely how a buyer reads.
So I split the measurement. Presence is one field. Position is another. Citation is another. Description accuracy is another. I do not combine them too early. The separate fields make the result slower to read, but they stop the team from celebrating a weak mention.
In the composite software integrator case, the company appeared in several answers, so a presence-only score would have looked tolerable. Once I separated the fields, the weakness was obvious. Competitors were cited through category pages, trade articles or specific implementation references. The integrator was often cited through vendor directories. The generated description leaned toward resale and support, not implementation strategy. The name was present; the commercial meaning was thin.
That is why I sometimes use weighted reading after the raw count. A first-position mention with a relevant cited source and accurate description deserves more attention than a final-position mention with no source or a confused description. I do not pretend there is one universal formula. The weighting should fit the decision. A sales team may care most about inclusion and description. A brand team may care more about citation share and category association. A founder looking at investor perception may care about position across broad category prompts.
The error is not choosing a simple formula. The error is hiding the fields that explain the formula.
The category boundary decides the result
A share-of-voice number is meaningless until the category is named. “AI visibility for our company” is too wide. The software integrator may compete against local IT consultancies in one prompt, vendor-certified partners in another, global consultancies in a third and niche industrial specialists in a fourth. Each category produces a different answer shelf.
I define the category from buyer intent, not from the company’s preferred label. If a French industrial SME asks for help implementing a certain class of software, the prompt should use that buyer wording. If the buyer asks for a partner near Lyon that can handle integration, training and support, that becomes another category cell. English vendor language may deserve its own test if buyers search that way or if the vendor ecosystem is largely documented in English.
This is where many reports become too flattering. They choose a category where the company is unusually visible and treat it as the whole market. I prefer a rougher map. One cell for the core service. One for the higher-margin work. One for the problem the buyer thinks they have before they know the service name. One for location. One for competitor comparison. The map should contain the uncomfortable prompts as well as the obvious ones.
I call this category shelf mapping: defining the small answer spaces where a buyer expects alternatives, then measuring which names occupy those spaces repeatedly.
For the integrator, “software reseller near Lyon” was not the desired category, but it mattered because the model kept pulling the company toward that shelf. The business wanted to be understood as an implementation specialist. The share-of-voice ledger showed that the answer engines had not fully accepted that distinction. This is not a branding mood. It is a measurable category drift.
Competitors are not decoration in the ledger
If the ledger records only your business, share of voice cannot exist. It becomes presence tracking with a nicer label. Competitors have to stay inside the measurement set because they define the lost answer space.
I record named competitors even when the client dislikes them. Especially then. The engine is showing who it treats as relevant. Some names will be expected. Others will be odd: a directory, a vendor marketplace, a regional agency with stronger local proof, a large consultancy that appears because its pages explain the category better. Each odd name is a clue.
In the composite integrator scenario, one competitor appeared across French prompts because a trade publication had described a case study in plain language. The company’s own case studies were more detailed, but they buried the business role under internal terms and vendor vocabulary. Another competitor appeared in English prompts through vendor partner pages. The client’s vendor page existed too, but it was thin and gave no useful implementation description.
A competitor jump should not immediately lead to copying their content. I read their cited sources first. What kind of source is feeding the answer? Is it their site, a directory, a review page, a partner record, a trade article? Is the source current, clear and specific? Does it explain the service in the same language as the prompt?
This reading keeps the work from becoming nervous imitation. The goal is not to sound like the competitor. The goal is to understand why the engine had enough evidence to place them on the shelf.
A competitor can win because it is better known, better described, better cited or simply less ambiguous. Those are different conditions. The ledger should make them visible.
How I compute it without making a false science
I do not sell a mystical score. I build a table that can be rerun. A practical generative share-of-voice audit begins with a fixed prompt set, a defined engine set, repeated runs and a competitor roster. Then each answer is scored for the fields that matter: presence, position, cited source, description accuracy and category fit.
The raw mention share is the first view. If twenty answer observations produce one hundred named business slots, and your company occupies twelve of those slots, that is a rough presence share. If a competitor occupies thirty, the difference is visible. But I also look at citation share: how often the engine cites a source that supports your business compared with sources supporting competitors. Then I examine top-position share, because buyers often notice early names more than later names. Finally, I read description accuracy, because a wrong category description can turn a mention into damage.
I keep the arithmetic simple enough that a manager can inspect it. A score no one can audit becomes another decorative dashboard. The best measurement tables are slightly plain. Date. Engine. Prompt. Language. Named businesses. Order. Cited source. Description quality. Notes. The plainness is the virtue.
There are limits. Answer engines vary. Runs vary. Interfaces change. Some answers cite sources, some do not. Google AI Overviews behave differently from a chat answer. I do not flatten these differences into one confident number. I show separate readings first and use the combined view only as a management summary.
The number should provoke the right question. Why are we absent from industrial prompts but present in vendor prompts? Why does a competitor own English citations? Why do we appear but get described as a reseller? A good share-of-voice metric opens the cabinet; it does not pretend the cabinet is the whole factory.
The fix starts with the source that earned the voice
Once the share-of-voice pattern is clear, recommendations become less theatrical. If the company is absent across all relevant category prompts, the issue may be insufficient entity evidence. If it appears but loses citation share, the better sources may be weak, buried or less trusted by the engine. If it appears with the wrong description, the cited sources may be teaching the wrong category.
For the software integrator, the correction would not start with chasing every keyword rank. It would start with the sources that shaped the answer shelf. French case studies need to state the implementation role plainly. Vendor pages need a description that separates implementation from resale. Service pages need buyer-language headings, not only internal practice names. Third-party profiles should not be allowed to carry the clearest description if they are vague or old.
Then the same prompt set runs again. I want to know whether the business gains answer slots, whether cited sources improve, whether the description moves toward the right category, and whether competitor share changes. Some movement will be slow. Some will not move at all. That is useful too. It tells us whether the correction reached the evidence path that the engine uses.
Keyword rank still belongs in a search conversation. I use it to understand demand and page performance. But when the question is “are we visible in AI answers against the competitors buyers see,” keyword rank is no longer the main instrument. It is one old gauge on a panel that now has other needles.
The Measurement Note — Signal: the business occupies repeated answer slots for a defined buyer category. Distortion: treating keyword rank as proof that AI answers will name the company. Ledger: record prompt, engine, language, named competitors, answer order, cited source and description accuracy. Next Test: choose one category question, run it across three engines, and count both your mentions and competitor mentions before reading any ranking report.