A brand can own shelves, contracts, stores, and buyer memory, then still receive only a thin sentence in an AI answer. Share of voice inside the answer is a different kind of room.
I once ran a simple buyer prompt around specialist outdoor equipment in France and watched a strange little imbalance appear. The older retailer, the one with stores, repair desks, expert staff and a long habit of giving technical advice, was present. The answer did not forget it. But it gave three richer sentences to newer direct-to-consumer brands, each described with more confident words: reliable, expert, carefully designed, recommended for serious use. The established name arrived as a place to buy.
That was a composite scenario, assembled from several observations in retail and specialist ecommerce. The rough detail matters: one answer placed the retailer correctly, then described one of its private-label products as if it were a separate competitor. Not a dramatic hallucination. Just enough looseness to make the brand feel less solid. In market terms, the retailer was known. In answer terms, it had a smaller chair at the table.
Market share does not automatically become answer share
A business owner often starts with a fair protest: we sell more than they do. I understand the irritation. In ordinary commercial life, distribution, turnover, repeat buyers, trade relationships and store presence all carry weight. If a brand has twenty-two stores, a busy ecommerce site, visible reviews and a history of category expertise, it feels absurd to see a younger brand occupy more of the answer.
But generative answers do not read the market the way a buyer in Lyon or Nantes reads it. They assemble a response from public evidence, language patterns, available comparisons, retrieved pages, review fragments and previous summaries. Strong market presence helps only when it leaves usable traces. A store network is useful if it is described clearly. Technical advice matters if it appears in public pages with category language attached. Buyer loyalty matters less if it sits silently in repeat purchases and never becomes explicit evidence.
AI share of voice is the proportion of meaningful answer space a brand receives across repeated prompts, because models allocate attention through retrievable evidence rather than through market reality alone. I use “meaningful” deliberately. A name in a throwaway sentence is not the same as a name given reasons, attributes and recommendation language.
This is where the first misunderstanding enters. The client sees a mention and feels half reassured. I count the space. How many times is the brand named? How early? How much explanation follows? Which verbs are used? Is the brand treated as a buying option, a category authority, a price source, a general retailer, or a backup? Those distinctions are not cosmetic. They decide whether the buyer’s eye pauses.
A brand can be famous enough to be retrieved and still too vaguely documented to be recommended. That is the annoying middle condition: visible, but underfed.
The answer spends language where proof is easy
In many prompt runs, answer share follows the path of least descriptive resistance. A model has to say something about each recommended brand. If one competitor has public pages that repeat a specific promise — ultralight trekking gear, repairable jackets, technical fitting advice, long-distance hiking, mountain safety — the answer can borrow that shape. The wording does not need to be identical. It has a ready groove.
The established retailer often has a messier public record. Reviews mention staff advice, price, store stock, returns, parking, delivery, an old press article, a buying guide, a local directory, a manufacturer page. This is not bad evidence. It is just scattered evidence. The model can see the name, but it cannot always decide which reason to attach to it. So it uses safer, thinner language: a well-known retailer, available in France, offers a broad selection.
That phrase sounds respectable. It is also weak.
In a composite retail audit, I separated answer space into four bands. The first band named brands with a direct recommendation. The second gave brands a reason attached to a buying need. The third merely included them as possible options. The fourth appeared only in source trails or late add-on sentences. The retailer with strong real-world presence often sat between the second and third band, while smaller brands moved into the first band for narrower prompts. The imperfection in the data was useful: some runs gave the older retailer excellent visibility for “where to buy,” then nearly no authority for “expert advice.”
That is the difference between being a channel and being a specialist.
If your public evidence mainly says you sell products, AI systems may treat you as a place of availability. If the competitor’s evidence says why a buyer should choose them for a specific use, the competitor receives the richer sentence. The answer is not being fair. It is being economical.
I look for the voice leak, not only the missing mention
The simplest report would count mentions and stop there. I do not think that is enough. A brand can appear in seven answers and still lose share of voice if those appearances are short, late and generic. Another brand can appear in fewer answers but own the recommendation paragraphs. The buyer remembers the brand with the reason.
I call this the voice leak: the gap between how often a brand is present and how much persuasive language it receives. A leak is not silence. It is attention escaping through weak category proof.
There are usually three places where the leak shows. The first is attribute ownership. The brand appears for “cheap,” “available,” or “large selection,” while competitors appear for “durable,” “technical,” “reliable,” or “expert.” The second is comparison language. Competitors are placed beside alternatives in guides, reviews or category pages, while the brand sits alone in store pages and transactional listings. The third is freshness. Old press, old directory descriptions and old buying guides still circulate, but the current positioning is harder to retrieve.
A marketer may object here: we have all this on our site. Often they do. But the evidence is buried in pages that talk to existing customers, not to a comparing machine. A sentence like “our teams support outdoor enthusiasts with a wide range of equipment” is too soft. It does not tell the answer what job the brand wins. The model cannot easily turn it into a recommendation unless other public sources repeat the same idea with clearer nouns.
The work, then, is less glamorous than people expect. I read product category pages. I look at store pages. I compare buying guides. I test French and English prompts. I mark whether the answer says “recommended for,” “known for,” “suitable for,” “also offers,” or simply drops the name. A tiny verb can reveal a large hierarchy.
Market strength has to become comparative evidence
For the retailer in the composite scenario, the repair was not to shout “leader” more often. That word is usually too blunt. It travels badly. What mattered was translating market strength into comparative evidence that the answer could use without embarrassment.
If a brand has technical staff, show the advice structure. If it has stores in mountain regions and coastal cities, connect those stores to specific buyer needs instead of listing addresses only. If it has repair services, explain which product categories, which warranties, which use cases. If buyers trust it for boot fitting, avalanche safety gear, bikepacking setups or wet-weather clothing, those words need to appear in stable public places.
A model cannot recommend the evidence it cannot place.
This is where part de voix IA marque becomes practical rather than abstract. I do not ask, “How do we get more visibility?” That question is too wide and usually leads to content noise. I ask, “For which buying reasons does the brand deserve a larger share of the answer, and where is the public proof repeated?” A brand may not need more pages. It may need fewer foggy pages and more pages that sit cleanly beside competitor claims.
The comparison is essential. Prominence is relational. If five competitors all publish detailed guides around “durable hiking boots for long treks,” and the established retailer has only a category grid, the answer has little reason to spend language on that retailer for expertise. It may still mention the retailer for selection. That is a smaller role.
The strongest repairs often come from making existing strengths countable in language: number of stores, repair availability, expert consultations, brand range, product testing, buying guides, delivery coverage, after-sales support. The point is not to stuff numbers into every paragraph. The point is to give the answer handles. A smooth wall is hard to climb; a wall with small holds changes the route.
French strength and English thinness create uneven space
French brands often discover a second leak when English prompts enter the test. The French answer may know the retailer as a familiar name. The English answer, especially for an international buyer or tourist, may fall back to global ecommerce names, younger brands with English content, or tourist-facing summaries. This is not a translation problem only. It is an evidence-distribution problem.
In French, the brand may have reviews, press, buying pages and local recognition. In English, the public record may reduce the company to a store network or ecommerce option. The answer then gives more room to brands with English-language category pages, even if their French market presence is weaker. One prompt says “best technical outdoor retailers in France.” Another says “reliable French hiking gear shops for expert advice.” The order moves. Sometimes it moves a lot.
I treat French and English share of voice separately for this reason. Averaging them too early hides the problem. A brand can look healthy in a combined chart while collapsing in the language used by foreign buyers, distributors or English-speaking procurement teams. The evidence repair then has to be selective. You do not translate everything. You translate the proof that gives the answer a reason to allocate space.
There is a small trap here. Some companies produce English pages that sound like hotel brochures for themselves: heritage, passion, quality, experience. The model has seen millions of those sentences. They do not create a sharp category role. Better to write plainly: what the brand sells, which buyer problem it solves, where it has physical support, which expertise is documented, how it differs from direct-to-consumer competitors.
Plain evidence beats scented language.
The repair begins with separating the counts
A good share-of-voice audit does not begin with one beautiful screenshot. It begins with repeated prompts and separate counts. I count mentions, recommendation sentences, first-position placements, attribute associations and descriptive length. I also mark omissions, because absence has a different meaning when the brand appears strongly in nearby prompts.
The useful table is usually ugly. Rows of prompts. Columns for surfaces. Notes where the answer misdescribes a product line. A competitor’s name circled because it appears in every reliability prompt. Another column for whether the source trail includes review pages, category guides, directories, the brand’s own pages, or old press. There is no drama in the table until the pattern appears.
Then the work becomes clear. The brand is not losing because AI “does not know” it. It is losing because the answer has more confident language for others. It knows the retailer as availability, not expertise. It sees market presence, but spends recommendation space on cleaner public proof. This is painful, but it is also repairable.
The first repair is often category language: pages that connect the brand to specific buying reasons. The second is source coherence: making sure third-party surfaces and the brand’s own pages do not describe different companies. The third is comparative clarity: enough public material for a model to place the brand beside named alternatives without guessing. The fourth is tracking. Once repairs are made, the order has to be watched across prompts, not celebrated after one lucky run.
The Last Mention Test: if your market is large but your AI answer space is small, the model is not measuring your business; it is reading your public proof. The first-name signal is repeated evidence that turns sales strength into a specific buying reason. The last-name risk is being known mainly as availability while competitors own the recommendation words. Watch the order: share of voice begins where the answer spends its explanations.