A prompt is a small handle on a large public record. Move the handle by two words, and the answer may pull a different part of the brand into view.
On my desk, prompt notebooks become untidy quickly. One row says “best outdoor equipment retailer in France.” The next says “reliable hiking gear shop with expert advice.” A third asks for “technical outdoor brands for long-distance trekking.” The names move. Sometimes the same brand rises, falls, or vanishes across three questions that a human buyer would consider nearly identical.
A composite scenario from specialist outdoor retail shows the problem well. A French retailer with stores and ecommerce presence appeared strongly when the question emphasized availability and price. It weakened when the prompt asked for reliability. It nearly disappeared when the question shifted toward technical expertise, although its store staff were known for advice. One answer even recommended a younger direct-to-consumer brand for “expert fitting,” which was not exactly its strongest claim. The model had followed the cleanest public wording, not the messiest market reality.
The question is part of the evidence test
People often treat prompt wording as a nuisance. They want the “real” answer. I think that is the wrong starting point. The wording is not outside the test; it is one of the test instruments. A buyer does not ask in one fixed sentence. They ask from mood, urgency, budget, fear, role and language. “Best” is not the same as “reliable.” “Specialist” is not the same as “near me.” “For SMEs” is not the same as “for enterprise procurement.”
The answer order changes because each phrase activates a slightly different evidence trail. A broad prompt may reward famous names. A reliability prompt may reward reviews, long-standing service descriptions or third-party trust signals. A technical prompt may reward guides, detailed product pages and expert language. A French prompt may lean on local evidence. An English prompt may pull from thinner summaries or international sources.
Prompts IA visibilité marque is therefore not a game of finding one lucky sentence. It is the practice of testing how a brand’s prominence changes as buyer language moves across nearby meanings. The nearby part matters. I am not interested in absurd prompts that no buyer would use. I am interested in small, plausible shifts that change who gets recommended.
This is why a single screenshot can be misleading. It may show the brand under one light. Turn the lamp a little, and the object has a different shape.
I use prompt families, not prompt trophies
The trophy screenshot has a strange power in meetings. A brand appears first once, and the room relaxes. Or the brand disappears once, and everyone panics. Neither reaction is useful. One run is a weather report through a keyhole.
A prompt family is a set of buyer questions that belong to the same commercial situation but vary wording, intent, attribute and language. I usually build families around real buyer speech: what a founder would ask, what a procurement person would ask, what a marketing manager would ask after hearing two competitor names, what an English-speaking buyer would ask when exploring the French market.
For the outdoor retailer, one family might include availability prompts: where to buy hiking equipment in France, best outdoor stores with broad stock, reliable retailer for mountain gear. Another family tests expertise: shops with technical advice, best place for boot fitting, retailers for long-distance trekking equipment. Another tests comparison: deciding between specialist retailers and newer direct brands, or finding alternatives to direct-to-consumer outdoor brands. The exact names change by category, but the structure holds.
The rough edges are informative. In one composite run, the retailer gained when the prompt said “store,” lost when it said “brand,” and returned when the prompt said “advice in-store.” That pattern was not random. It showed the public record understood the company as a retailer more than an authority. The business knew it had expertise. The answer did not consistently know where to attach it.
I call this phrase drift: the movement of a brand’s AI-answer position when the buyer changes only a small part of the question. Phrase drift is not noise, because it reveals which pieces of public evidence the model can connect to which buying words.
Some words are heavy
In every category, a few words carry unusual weight. “Reliable” is heavy. So is “specialist.” “Affordable” often pulls one set of sources, while “premium” pulls another. “Local” can help or shrink a brand, depending on the question. “French” can make a national brand visible in one prompt and irrelevant in another if the answer interprets it as manufacturing origin rather than market presence.
The danger is assuming your preferred word is the buyer’s word. A company may describe itself as a “partner,” while buyers ask for a “provider,” “agency,” “integrator,” “retailer,” “supplier,” or “expert.” In French, the gap can be sharper because commercial language, category labels and buyer habits do not always map cleanly into English. “Fiable,” “spécialiste,” “adapté aux PME,” “meilleur,” and “recommandé” do not produce mirror versions of English answers.
For a French brand, this creates two layers of movement. First, the answer moves when the wording changes inside French. Then it moves again when the buyer asks in English. The brand that looks strong for “magasin outdoor France” may be weak for “technical hiking gear retailer France.” The software firm that appears for “intégrateur logiciel retail” may fall under “best retail operations software partner in France” if English evidence is thin.
I try not to overread one movement. Models vary. Retrieval changes. Some answers are simply poor. The pattern matters when the same kind of movement repeats across surfaces and runs. If “reliable” consistently removes you, while “available” restores you, that is a diagnosis. It says the brand has availability evidence but weak reliability evidence, at least in the public record the systems can use.
A prompt is not a magic spell. It is a question that exposes which parts of the entity record are strong enough to answer back.
The answer order can split by buyer intent
A buyer asking “best” may want prestige, consensus or a safe shortlist. A buyer asking “for small teams” wants fit. A buyer asking “with expert advice” wants confidence. A buyer asking “alternatives to X” wants comparison. These intents overlap, but they do not retrieve the same kind of proof.
The outdoor retailer in the composite scenario had good visibility for transactional intent. It was a place to buy. It was a place with stock. It was a known name. But recommendation language for expertise drifted toward brands with cleaner editorial evidence: buying guides, technical positioning, product-use language, and pages written around specific outdoor activities. The retailer had expertise in the building; the competitor had expertise in the answer.
That sentence is uncomfortable, but useful.
Intent splits also explain why a brand may be delighted and worried at the same time. It appears first for “where can I buy,” third for “best,” absent for “expert,” and misdescribed for “premium.” Which result is real? All of them, in their own prompt lane. The business has no single AI position. It has a position map.
For audits, I usually label these lanes by buyer job rather than by keyword. Availability. Trust. Specialism. Comparison. Price. Local fit. Enterprise fit. English-market discovery. The labels are not decorative. They stop the team from trying to fix every prompt with the same page. A weak expertise lane needs different evidence from a weak price lane. A weak English discovery lane may need bilingual proof, not more French content.
Stability comes from repeated public alignment
When a brand wants to “win across phrasings,” the answer is not to create pages for every possible wording. That leads to a thin, repetitive site that smells like it was written for a machine because, well, it was. The better repair is to align real strengths with several natural buyer phrases.
For the retailer, if technical advice is a real strength, it should not appear only in a brand story paragraph. It should appear in buying guides, store service descriptions, product category introductions, staff expertise pages where appropriate, and third-party profiles that describe the company accurately. The phrase does not need to be identical. In fact, identical repetition can feel artificial. The idea must be stable: technical advice, fitting, repair, specialist selection, activity-specific guidance.
The same principle applies to B2B services. If a software integrator wants to appear for “retail operations partner,” “logistics software specialist,” and “mid-market implementation,” its public evidence needs to connect those phrases to real work. One vague “we support digital change” page will not carry the weight. The answer needs multiple, consistent clues.
I sometimes describe this as building a braided rope. One strand is the brand’s own language. One is client or project evidence. One is third-party description. One is category vocabulary. One is language coverage across French and English. Any single strand can fray. The rope holds when the strands twist in the same direction.
The model does not know your internal truth. It reads the braid.
A useful prompt set is narrow enough to repeat
There is a practical temptation to test everything. Dozens of prompts, every platform, every competitor, every language, every buyer type. The table becomes impressive and useless. I prefer a narrower set that can be repeated over time.
A good starting set might include ten to fifteen prompts for one category: broad category, recommendation intent, reliability, specialism, price or value, local or national fit, comparison against named competitors, and language variants where needed. Each prompt should be close enough to a real buyer question that the result matters. Each should be repeated, because one answer is too fragile.
Then I count positions separately. Mention. Recommendation. First place. Omission. Misdescription. Attribute match. I also note source trails where the surface provides them, and I read the pages that seem to support the answer. The result is rarely a clean scoreboard. It is more like a map with muddy footpaths. That is fine. We are trying to see where the brand can travel, not design a poster.
Over time, the prompt set shows whether repairs are working. The brand may not rise everywhere. It may rise first in the lanes where evidence was improved. That is a healthier sign than one spectacular jump in a random prompt. Prominence worth defending is usually boring before it becomes valuable.
The Last Mention Test: if your order changes every time the question moves, the answer is showing where your evidence is narrow. The first-name signal is proof that survives several buyer phrasings, not one lucky prompt. The last-name risk is being strong only for the exact words you already use. Watch the order: a brand’s real AI position is the pattern across nearby questions.