A mention can feel like applause when it is actually background noise. The useful question is colder: when the answer has room to cite only a few businesses, whose evidence does it keep choosing?
A composite scenario: a 42-person plumbing and heating network in western France asks whether it is “visible in AI.” It has six local branches, bilingual service pages, decent local profiles in some towns and patchier evidence in others. In one run, an answer names the company for boiler maintenance near Rennes. The marketing manager saves the screenshot. Everyone relaxes for half a meeting.
Then the ledger fills in. Across repeated prompts, the company is named sometimes, cited less often, and placed below two competitors in most commercial-maintenance questions. In emergency-plumbing prompts it appears more often, which is not where the margin is. In nearby towns, it disappears behind single-branch competitors with stronger local pages. The first screenshot was true. It was also not enough to make a decision.
Being named is the weakest useful signal
A business mention has value. I do not dismiss it. If an AI answer never names a business, there is an absence problem to inspect. But a mention alone is the thinnest layer of the measurement. It tells you the model can produce the name under some condition. It does not tell you whether the business is preferred, cited, trusted, well described or present across the prompts that resemble buyer questions.
This matters because AI answers compress choice. A search results page can show ten blue links, a map pack, ads, directories, review stars and several paths out. A generated answer often names only a small handful of options. The difference between being named once and being repeatedly cited in that handful is the difference between a postcard and a seat at the table.
The temptation is understandable. A founder sees the company name in an answer and feels the work is paying off. An agency sees a favourable screenshot and wants to include it in a report. I have done enough ledgers to distrust that feeling. The answer that flatters you once may still cite competitors in the runs that matter, especially when the prompt changes from branded or insider wording to ordinary buyer language.
AI citation share is the proportion of relevant AI answer citations or named-source attributions that point to a business, because visibility only becomes meaningful when measured against the competitors occupying the same answer space.
The definition includes competitors because otherwise the number floats. A company cited in 20 percent of prompts may be doing well in a fragmented category and poorly in a category where one rival holds 60 percent. The share needs a market, a prompt set and a named competitor group around it.
The competitor set keeps the measurement honest
When I start a citation-share ledger, I ask for competitors before I ask for recommendations. Not a list of enemies from a sales deck. A measurement set. The local operator in the next town. The national brand that appears for every broad prompt. The specialist that keeps being cited by trade pages. The directory or marketplace that is not a competitor in the old sense but steals answer space.
For the plumbing and heating network composite, the competitor set would not be identical across every prompt. Emergency repairs, heat-pump installation, commercial maintenance and boiler servicing do not always produce the same rivals. A six-branch business may compete with one kind of company in Rennes and a different kind near Vannes or Saint-Malo. The ledger has to preserve that unevenness instead of sanding it smooth.
There is a rough detail in this kind of work that often surprises teams: a competitor can win citation share with ugly pages. Not always, but often enough to make designers unhappy. The page may be visually tired and still contain clear service boundaries, city names, opening hours, certification details, review traces and old directory reinforcement. Another company may have a better-looking site with weaker extractable proof. The machine does not admire taste in the same way a human does.
This is why competitor names must stay inside the measurement set. Without them, every mention looks promising. With them, you can see whether your business is merely present in the category cloud or actually chosen in the answer.
Citation share is not the same as share of voice
The phrase “share of voice” can be useful, but I separate it from citation share in the ledger. Share of voice can count mentions, positions and answer space. Citation share is narrower. It asks which sources the answer names, cites or uses as support when presenting options. In engines that show citations, this is literal. In answers with weaker citation display, I record named sources, referenced pages, or the absence of visible source support with a caution note.
This distinction prevents a common mistake. A business may be mentioned in a generated paragraph, while the clickable citation points to a competitor, a directory or a trade article that gives another firm stronger authority. The reader sees the mention. The engine’s evidence path may still favour someone else.
In practical terms, I build what I call the citation-share ladder. The bottom rung is presence: was the business named? The next rung is position: where did it appear among named options? The third is citation: was it tied to a source, and which one? The fourth is description accuracy: did the answer describe the business correctly? The top rung is repeated citation share across the agreed prompt set.
A single mention sits on the bottom rung. It should not be treated like the top.
The ladder also shows where repairs belong. If the business is named often but rarely cited, the work may involve strengthening sources. If it is cited but described wrongly, the problem is accuracy. If it is accurate but appears below competitors in high-value prompts, the issue may be category evidence, local proof or competitor strength. Different rung, different repair.
Prompts decide which competitors matter
A bad prompt set makes citation share look cleaner than it is. If every prompt asks for “best plumbing company in western France,” the ledger will favour broad, general evidence. If the buyer actually asks “chauffagiste contrat entretien immeuble Rennes” or “plombier urgence fuite samedi Saint-Brieuc,” the competitor group changes. The citation share changes with it.
For the regional network composite, I would separate prompts by service type, city or catchment area, urgency and customer type. Residential emergency repair is not the same as commercial heating maintenance. A company that wants more maintenance contracts should not be comforted by strong visibility in emergency-only answers. That is the kind of false comfort that arrives wearing a very practical jacket.
The imperfect detail matters here. In one set of teaching runs, imagine the company appears well for “chauffagiste Rennes contrat entretien chaudière,” but the answer describes it as mostly residential. In another prompt, it appears for “plombier urgence Rennes,” but without a citation. In a third, a local competitor is cited from a thin page with a clear city-service title. These are not equal observations. They tell different stories.
Good prompts are not invented in a conference room by people staring at the service menu. They come from sales calls, intake forms, support questions, search-console traces, local vocabulary and the awkward phrases buyers use when they do not know the correct trade term. The ledger should include that awkwardness. AI visibility is measured in buyer language, not in the tidy taxonomy a company wishes buyers used.
The cited page often explains the competitor jump
When a competitor starts appearing more often, the first useful question is not “What did they do to their brand?” I ask, “What source is the answer citing for them?” Sometimes the answer points to their own local page. Sometimes it points to a directory, a review profile, a public procurement mention, a partner listing, a trade article or a page that has not been touched for years but still carries the clearest category evidence.
In the plumbing and heating composite, a competitor might win in one town because its branch page states the service, city and maintenance scope plainly. Another might win because a local directory has a complete entry with opening hours, emergency terms and review snippets. The regional network may be larger, better staffed and more suitable for the job, yet the machine keeps choosing the smaller firm because the smaller firm leaves a cleaner trail.
That is irritating. It is also measurable.
I do not recommend copying the competitor’s page. That creates another weak source in the same pile. Instead I look for the evidence type. Is the cited page category-specific? Local? Fresh enough in its facts? Clear about the customer type? Does it contain the service terms the prompt activated? Does it connect the business to the location without pretending to serve every town in France?
The competitor’s citation is a clue, not a template. The work is to understand what kind of proof the answer found useful and whether your stronger proof exists in a form the engine can repeat.
Measure share by category, not by ego
A company wants to know whether it is visible. That is natural. The better question is visible for which category, against which competitors, in which language, in which location, with what description? Citation share becomes useful only after those boundaries are drawn.
For a multi-branch service business, I would rather show a rough category table than one smooth visibility score. Commercial maintenance in Rennes. Emergency plumbing in Nantes. Heat-pump installation in nearby towns. French prompts. English prompts, if the evidence or customers justify them. In each row, the competitor set may change. The share may be strong in one row and weak in another. That is not messy; that is the market.
The reporting rhythm matters too. Citation share should be tracked over repeated runs, then compared month by month. One answer can wobble. A pattern deserves attention. If a competitor takes more cited positions across several prompts and engines, that is a signal to inspect their sources. If your own cited share improves but accuracy gets worse, do not celebrate too quickly. More citations of the wrong description make the wrong description harder to ignore.
The point is not to turn AI answers into old keyword ranks with a new label. A generated answer behaves differently from a search page. It compresses, paraphrases, cites unevenly and sometimes makes category judgments the business never asked for. The measurement has to respect that shape.
Still, the old discipline remains: define the market, name the competitors, repeat the test, record the evidence, and resist the screenshot that wants to be a conclusion.
What a useful citation-share ledger contains
My basic ledger for this work is not beautiful. It has one row per run. The columns are plain: date, engine, language, prompt, location intent, service category, named businesses, answer order, cited sources, citation target, description accuracy, competitor group and notes. If an answer gives no visible citation, I mark that instead of pretending the source is known. If the engine cites a directory for one competitor and an official page for another, I keep that distinction.
Over time the ledger starts to show the answer space. You see which competitors are always present, which appear only in broad prompts, which are cited by strong sources, and which are named without much evidence. You also see where your business is being flattered by irrelevant visibility. That can sting a little. It is better than spending three months improving the wrong page for the wrong buyer question.
For the western France plumbing and heating network, the useful finding might be something like this: strong presence for emergency prompts in two cities, weak citation share for maintenance prompts, frequent competitor citation from local service pages, and a recurring misdescription that underplays commercial work. That is not a slogan. It is enough to plan.
And if the company appears once in a pretty answer? I still save the screenshot sometimes. It is evidence. I just do not let it sit at the head of the table.
The Measurement Note — Signal: competitors are cited more often than you in the prompts that matter. Distortion: treating any mention as equal visibility. Ledger: record named competitors, answer order, cited page, citation target and description quality for every run. Next Test: choose ten buyer prompts, add three named competitors, run the set twice, and calculate citation share by category instead of counting mentions.