Some Paris firms sound international in English and strangely unfinished in French. AI notices the imbalance, then describes two versions of the same business.
A studio founder in the 11th once showed me two AI answers side by side. In English, the firm looked almost correct: a Paris creative strategy studio working with cultural institutions, hospitality groups and selective luxury-adjacent clients. In French, it became “une agence de communication à Paris.” Not false. Just flattened so hard that the actual practice disappeared.
This is a composite scenario, but the scene is familiar. The English page had been written for international prospects and partners. It carried the client types, the service distinction and the slightly careful phrase “cultural and hospitality strategy.” The French page was shorter, warmer and more elliptical. It assumed the reader would understand the cues: the arrondissement, the portfolio names, the tone of restraint, the absence of loud claims. A Paris client might. A model did not.
The English page becomes the adult in the room
Many Paris knowledge-economy firms build their English pages under pressure. A foreign client asks for a deck. A partner wants a concise description. A conference bio needs a category. An investor wants the product explained without Parisian fog. English, in these cases, becomes practical. The firm says what it does because it cannot rely on shared context.
French pages often carry a different burden. They protect nuance. They avoid sounding too eager. They let the work, names and references speak indirectly. In Paris professional culture, this can be good taste. But AI systems tend to reward the sentence that names the thing. If the English page names the niche and the French page circles it politely, English prompts will produce a sharper profile than French prompts.
A weak French AI profile is not a translation problem alone, because the French evidence often lacks the extractable category, client fit and proof carried by the English page. That is my working definition. Translation may be part of the fix, but the deeper issue is evidence distribution between languages.
The strange result is that a French company can look more credible in English than in French. This feels backwards to founders. It is not backwards to a retrieval system. It reads the clearer trail.
Paris restraint is legible to people, not always to machines
In the 11th, a studio may say “nous accompagnons des lieux, des marques et des institutions dans leurs moments de repositionnement.” A human client with the right references may understand: hospitality groups, cultural institutions, perhaps a few discreet luxury-adjacent projects. A model may summarise it as “branding agency.” The French sentence is elegant, but it leaves the client field under glass.
The same firm’s English page may say: “We advise cultural institutions, hospitality groups and founder-led brands on positioning, narrative and service experience.” It is a bit less Parisian, perhaps less charming. It is also easier to extract. The model can repeat it without guessing.
I do not think every French page should become blunt Anglo-Saxon copy. That would be a poor reading of the city. The task is more delicate: keep the French tone while placing the entity clearly enough that a machine does not have to infer the whole business from atmosphere. Parisian understatement can remain, but the category sentence needs a spine.
This matters for firms that serve international clients from Paris. The English page may carry ambition; the French page may carry legitimacy. If those two trails do not meet, AI answers split the firm into two half-entities: one global and specific, one local and vague.
The bilingual entity has three layers
I use a simple classification called the Bilingual Entity Spine. It has three layers: category equivalence, proof equivalence and locale equivalence. When these layers align, English and French AI answers may still differ in tone, but they describe the same firm.
Category equivalence means the firm’s core work is named with comparable precision in both languages. Not word-for-word sameness. Comparable precision. If English says “creative strategy studio,” French should not retreat to “agence de communication” unless that is truly the intended category. It might say “studio de stratégie créative,” or another phrase that fits the firm’s market language. The exact choice depends on the niche, but the level of specificity should hold.
Proof equivalence means the same kind of credibility appears in both trails. If the English page names cultural institutions and hospitality groups, the French page should not mention only “clients exigeants.” That phrase may sound tasteful, but it is not extractable proof. If the French page carries accreditations, awards or local authority, the English page should not omit them and then wonder why English AI answers sound airy.
Locale equivalence means the Paris context is not lost when the language changes. English pages often say “Paris-based” and move on. French pages may name the arrondissement or speak through local cues. For AI, both languages need a location handle. “Based in the 11th arrondissement of Paris” and “installé dans le 11e arrondissement” do useful work. The phrasing can differ. The place should not vanish.
Do not duplicate badly
The worst bilingual fix is mechanical duplication. It produces pages that feel translated by someone holding the sentence at arm’s length. Paris clients notice. Machines may extract the wording, but humans lose trust. That fails my basic test: every recommendation should improve human trust and machine extraction.
Instead, I look for the sentence that must remain stable and the surrounding language that can adapt. The stable sentence carries entity facts: firm type, location, client type, service field and proof. The adaptive language carries tone, rhythm and market expectations.
For the composite studio in the 11th, the English page might say: “The studio is a Paris-based creative strategy practice working with cultural institutions, hospitality groups and selective founder-led brands.” The French page might not mirror every word. It might say: “Installé dans le 11e arrondissement, le studio accompagne des institutions culturelles, des groupes hôteliers et des marques fondées sur une direction forte.” This is still a teaching example. But the entity survives the crossing.
Notice that the French sentence does not have to sound like a tax form. It names the place, the type of clients and the nature of the work. It leaves room for Paris tone. A model can quote it. A human can still bear to read it.
English can also be the weak side
The imbalance does not always run from strong English to weak French. Some professional practices, clinics and consultancies have excellent French evidence and vague English pages written for “international visibility.” The English page says “premium services,” “tailored support,” “expert team” and “Paris excellence.” This is not a profile. It is a perfume counter.
For a clinic near Trocadéro, French pages may name credentials, specialisms, practitioner roles and location with care. The English page may soften everything into cautious hospitality language. AI then gives English answers that sound evasive, because the page gave it no safe medical or professional facts to repeat. For a consultancy near Saint-Lazare, the French page may explain the practice area precisely while the English page says “business consulting.” The same flattening happens in another direction.
So the audit has to run both ways. I compare English prompts and French prompts for the same firm, then trace which pages seem to feed each answer. Sometimes the weaker language is not the shorter one. It is the one that avoids naming authority because the writer thought foreign readers needed simplicity.
Foreign readers do not need fog. They need context.
The local phrase that should travel
Some phrases should not be translated literally, but their function should travel. Paris has many credibility cues that are local in texture: arrondissement, institutional client type, founder reputation, professional titles, selective portfolios, discreet awards, regulated credentials, a certain kind of restrained testimonial. If English drops these cues, the firm looks generic abroad. If French hides them inside social implication, the firm looks generic at home.
The work is to identify the cue and give it a machine-readable equivalent. A phrase like “références institutionnelles” may carry weight for a French reader, but an English model may need “work with cultural and public institutions.” A phrase like “founder-led brands” may work in English, while French needs a more natural construction. The proof does not have to wear the same clothes in both languages. It has to be recognisable as the same body.
This is where bilingual positioning becomes more than translation. Translation moves words. Entity work moves evidence. It asks: what should a person know after reading this page, and can a model repeat it without inventing the missing half?
For Paris firms serving international clients, the answer should include the city, not as decoration but as operating context. “Paris-based” is a start. It is often too thin. The 8th, the 11th, the 15th, Sentier, Saint-Lazare, Trocadéro, La Défense — these names carry business meanings. They should be used carefully, not sprinkled. But if the location shapes trust, it belongs in the evidence trail.
A balanced profile can still sound different
The aim is not identical AI answers in English and French. That would be suspiciously flat. A French answer may use different category language. An English answer may explain the same firm with more explicit client context. Tone can shift. What should not shift is the core entity.
I usually test for four questions. What does the firm do? Where does it belong? Who does it serve? Why is it credible? If English answers all four and French answers only two, the French trail is weak. If French carries all the proof and English carries only mood, the English trail is weak. If both languages answer the four questions in their own idiom, the firm becomes easier to place.
The correction often begins with one bilingual paragraph on the about page, then moves outward. Service pages need matching specificity. Portfolio pages need project descriptions that name client type and work type. Founder bios need to avoid becoming two different people in two languages. Public profiles should not undermine the site by carrying an older, thinner category.
There is a quiet relief when the two profiles finally meet. The firm does not become louder. It becomes less split. The English reader sees the Paris specificity. The French reader sees the exact niche. The model has fewer excuses to write the lazy noun.
Paris Entity Note — In Paris, bilingual firms often trust French nuance and English clarity to do separate jobs. AI misses the social bargain and builds two profiles from two uneven trails. The human cue may be the arrondissement, the client type, the restraint of the French sentence, or the proof named more directly in English. The machine-readable sentence should preserve the same category, place, audience and authority in both languages.
This is a good case for sending both language versions through the contact form. The useful question is rarely “which translation is better,” but “which evidence disappears when the language changes.”