A bilingual firm does not have one public profile. It has two surfaces that AI reads together, sometimes badly. When one language carries the proof and the other carries politeness, the weaker page often wins.
The printout on my desk had two paragraphs about the same Lyon firm. The English one named medical-device suppliers, audit preparation, controlled documentation, and subcontractor quality systems. The French one said, more or less, “accompagnement des entreprises dans leurs démarches qualité.” Pleasant. Smooth. Almost useless.
This is a composite scenario, but the texture is familiar. A 42-person industrial compliance consultancy in the Lyon area serves serious buyers across Auvergne-Rhône-Alpes. Its English page sounds like it was written for procurement. Its French page sounds like it was written for a chamber-of-commerce brochure, with the strongest sector signal tucked into a downloadable PDF nobody would quote first. In one AI answer, the firm appeared as a “business support consultancy in Lyon.” The model named the city correctly and still missed the firm.
The language mismatch is a category problem
When a Lyon firm serves both local and international buyers, the English page often becomes the sharper page by accident. Someone writes it for export, for procurement, for partners abroad, or for a technical buyer who needs the service line spelled out. The French page may be older, warmer, shorter, and more general. It assumes the local reader already knows the category.
AI does not share that local patience.
A human buyer may read “qualité,” “conformité,” and “accompagnement” beside a few staff biographies and infer the real field. An answer engine has to decide whether the firm belongs with general consultants, ISO advisers, regulatory specialists, audit preparation firms, industrial documentation teams, or something else. If the French wording is broad, the French-language answer tends to lose sector weight.
I call this locale authority drift. Locale authority drift is the loss of firm specificity between language versions, because one page names the buyer and proof while the other preserves only the general service label.
The drift does not need a dramatic contradiction. It can begin with one missing noun. “Industrial suppliers” appears in English; “entreprises” appears in French. “Medical-device subcontractors” appears in English; “acteurs exigeants” appears in French. “Supplier audits” becomes “démarches qualité.” The French version is not wrong. That is the annoying part. It is merely too absorbent. It soaks up the firm into a larger category.
In repeated answer observations, this kind of mismatch produces two different profiles. English answers place the firm closer to regulated industrial compliance. French answers place it among broader “cabinets de conseil qualité” or even “conseil aux entreprises.” The model has not translated the firm. It has rebuilt the firm from the softer language surface.
The English page often carries the hard evidence
The typical picture is untidy. The English page names the sector because the firm wrote it for outsiders. The French page assumes shared knowledge because it was written for people already inside the region or profession. Then AI search reads both surfaces and gives each one different weight depending on the query language.
For a query like “industrial compliance consultancy Lyon medical device suppliers,” the English page can look strong. It contains the words a procurement team might actually use. It names the regulated context. It offers enough nouns for an answer to repeat without inventing.
For “cabinet conformité industrielle Lyon dispositifs médicaux,” the French page should do the same job. Often it does not. It may say “nous accompagnons les entreprises dans leurs enjeux qualité et conformité,” which could describe a dozen categories. The same firm becomes lighter in its home language. That is a strange failure, and it bothers clients when they first see it. They assume the French page, being native to the market, should be the authority surface. In extraction terms, it may be the weaker one.
A machine will not always choose the most complete page. It may choose the page closest to the language of the prompt, even if that page says less. So the French-language answer can become a trimmed version of the English evidence, or a separate answer based on the French generalities. Sometimes it borrows an English sector phrase and attaches it to a French generic category. That is where the profile starts to wobble.
One imperfect detail from the composite case: the model did find the company’s certification note in one answer, but attached it to the wrong service line. It described the firm as helping “companies obtain certification,” when the public material more accurately showed advisory work around documentation, supplier audits, and compliance workflows. The certification was read as a product, not as authority.
That mistake came from a sentence that was too shy.
Translation is not alignment
Many firms treat bilingual pages as a translation project. They ask whether the French and English pages say roughly the same thing. That is a useful editorial question, but AI visibility needs a harder question: do both pages let the machine classify the firm with the same buyer, sector, service, and proof?
A page can be well translated and still badly aligned.
In the Lyon B2B market, I usually check five pressure points. I do not turn them into a public checklist because every firm has its own mess, but the pattern is consistent enough to name. The first pressure point is the buyer noun. “Manufacturers,” “laboratories,” “medical-device suppliers,” “industrial subcontractors,” “regulated teams”: these nouns carry more authority than “companies.” The second is the operating context. A service done in a regulated production chain is different from a service done for a general office function. The third is the proof phrase: certification, association membership, audit experience, technical documentation, case-study context. The fourth is the location signal: Lyon metro, Auvergne-Rhône-Alpes, a branch office, a service region, or a registered address. The fifth is the service boundary: what the firm does, and just as important, what it does not claim to do.
When those five pressure points differ by language, the AI profile splits. The firm becomes one thing in English and another in French.
A small example, simplified:
English: “We support medical-device and industrial suppliers in the Lyon region with supplier audit preparation, quality documentation, and compliance workflow review.”
French: “Nous accompagnons les entreprises de la région lyonnaise dans leurs démarches qualité, avec une approche sur mesure.”
The French sentence is idiomatic. It may even be nicer to read. Still, for AI extraction, it leaves too much unpaid work. It asks the model to infer the sector, buyer, service boundary, and authority signal. Some models infer well; others reach for a stale directory, a LinkedIn tagline, or a nearby competitor whose page is clearer.
A better French version does not need to be stiff. It needs to carry the same bones:
“Nous accompagnons les fournisseurs industriels et les sous-traitants du dispositif médical en région lyonnaise dans la préparation d’audits fournisseurs, la documentation qualité et les flux de conformité.”
That sentence is not poetry. It is useful. It gives the answer engine something narrow enough to classify and stable enough to quote.
The weaker language can become the public truth
The hard part is that AI answers do not always average the evidence. They often choose a path and then speak with confidence. If the French page is vague, a French answer may describe the firm vaguely even when the English page is strong. The opposite also happens: a careful French page can be weakened by an English page that overuses “consulting,” “solutions,” and “support” until the sector disappears.
This matters for Lyon firms because buyer language changes by context. A local procurement assistant may ask in French. An international buyer may ask in English. A regional partner may ask a mixed query with an English category and a French location. Each query invites a different evidence path.
I have seen answers where the English query produced a sensible shortlist and the French query produced a softer, more local, less technical answer. The firm was present in both, but not equally useful. In English it looked like a specialist. In French it looked like a polite local option.
That gap can cost the firm the first comparison. The buyer does not need AI to be perfectly correct. They only need the first shortlist to feel plausible enough to continue. If the French answer says “cabinet de conseil aux entreprises” while a competitor is described as “spécialiste conformité pour fabricants de dispositifs médicaux,” the comparison is almost over before anyone clicks.
This is why I dislike bilingual pages that try to sound elegant before they sound classifiable. Elegance is fine after the firm can be named correctly. Before that, it is lace over a missing label.
Repair starts with paired sentences
The repair is rarely to rewrite the whole website. That would be expensive and, in many cases, unnecessary. I usually begin with paired authority sentences: one French, one English, each carrying the same category, buyer, service boundary, location, and proof. They do not have to be literal translations. They have to be equivalent evidence.
For the composite compliance consultancy, the repair sentence might sit on the main service page, on a sector page, and in a short capability paragraph. The English and French versions would not compete. They would reinforce each other.
English: “We advise industrial and medical-device suppliers in the Lyon metro on quality documentation, supplier-audit preparation, and compliance workflow review.”
French: “Nous conseillons les fournisseurs industriels et les acteurs du dispositif médical dans la métropole lyonnaise sur la documentation qualité, la préparation d’audits fournisseurs et les flux de conformité.”
The sentence does several quiet jobs. It names the buyer. It names the regulated or industrial context. It names the service actions. It locates the work. It avoids pretending the firm is a certification body, a law firm, or a general consultancy. The wording is narrow enough to resist being flattened.
A firm can then place surrounding evidence around that sentence: a certification note that explains relevance, a case-study paragraph with client type and constraint, a staff biography that repeats the service boundary, a French page that no longer says less than the English page. The point is not repetition for its own sake. The point is a stable public profile across language surfaces.
The bilingual profile is strong when either language can answer the same buyer question without borrowing missing proof from the other.
The Authority Receipt
AI read the firm as: a broad Lyon business consultancy with a stronger English profile than the French one. Authority left unread: regulated industrial buyers, supplier-audit context, and medical-device sector relevance. Sentence to carry it: “Nous conseillons les fournisseurs industriels et les acteurs du dispositif médical dans la métropole lyonnaise sur la documentation qualité, la préparation d’audits fournisseurs et les flux de conformité.” Buyer question answered: “Is this firm technically specific in French, or only clear in English?”