A translation can be accurate and still be weaker evidence. AI systems often prefer the page that names the service boundary, even when that page is not the one meant for the local buyer.
The first clue is usually a small embarrassment in the answer. A French company asks, in French, about an expert near Nantes or Lyon or Rennes. ChatGPT mentions the company, perhaps, but borrows its clearest sentence from the English page. Sometimes it even gives the service in English-shaped terms, as if the French offer were a translation of the export page, not the other way around. The business owner sees the result and says, quite fairly: but our French page ranks.
A composite scenario I often use for this problem is a 28-person accounting and export-documentation firm near Nantes. The French page was older, better linked, and the one that actually ranked on Google.fr. The English page was thinner in design and less elegant in tone. Yet when I compared AI answers for export VAT and EU paperwork queries, the English page was easier for the model to describe. It said, in one plain line, that the firm prepared French VAT, customs support and export documentation for SMBs trading inside the EU. The French page said “accompagnement administratif et fiscal des entreprises en croissance à l’international.” Correct, pleasant, and almost useless when lifted outside the page. There was also a funny rough edge: the AI answer named the city correctly, then described the firm as if it handled freight forwarding, which the firm did not do.
When the local page becomes the softer page
A French page can be less cited because it has been written with more cultural caution. French business copy often avoids the blunt service sentence. It leans on relation, expertise, proximity, accompaniment, seriousness. None of those words are bad. I have written too many service pages to pretend otherwise. They make a page sound like it belongs in its market. The problem begins when the page never lands the aircraft.
The English version, by contrast, is often written later and for a more practical purpose. It may be aimed at foreign partners, export clients, investors, or English-speaking buyers who need quick explanation. The writer knows the reader may not understand the French category. So the English page says the thing directly: we prepare these documents, for these companies, in these situations, under these limits. That sentence is not more beautiful. It is simply easier to quote.
A page française moins citée is a French page that ranks locally but gives weaker extractable evidence than another language version or outside summary. That is my working definition, because the problem is not language preference alone; it is evidence density inside the language.
This distinction matters. I do not assume that answer engines “prefer English” in every case. That would be too broad and too easy. In my runs, the mechanism is usually more modest. The model finds a clearer statement somewhere else. English pages happen to contain more of those statements because they were written to explain the business to someone with less context.
The French page often assumes too much. It assumes the reader understands the profession. It assumes the service label carries the boundary. It assumes “international” implies the exact paperwork. It assumes “accompagnement” will be understood as operational help, not advice, not software, not forwarding, not legal representation. A human buyer may infer enough. An answer engine should not have to.
The translation that loses the proof
When I compare French and English pages, I make a small table. It is not elegant. On the left, the French claim. On the right, the English claim. Then I mark what survives as proof.
The surprise is that the English version is not always a translation. Sometimes it is an accidental clarification. The French page says the firm supports “PME tournées vers l’international.” The English page says it works with “French SMBs selling goods inside the EU.” The second sentence has a buyer type, a jurisdictional hint and a commercial situation. The first has mood. Mood is hard to cite.
There are also cases where the translation removes the French proof. A French industrial page may name a département, a workshop, a certification, a category of machine, or a service rhythm. The English page turns this into “regional technical support.” The page still reads fine. But the machine loses the specific source cue. In this topic, the typical problem is the reverse: the French page underperforms because the English version, by accident or necessity, explains more.
The cure is not to make the French page sound translated from English. That would be a cheap repair. A French buyer should not have to read a page that feels like a customs form. The work is to let the French sentence carry proof while staying natural.
For the Nantes composite, I would not replace everything with a hard, Anglo-style line. I would add one sober sentence where it belongs:
“Le cabinet accompagne les PME françaises sur la TVA, les formalités douanières et les documents d’exportation liés aux échanges dans l’Union européenne.”
That sentence is not poetry. It is a plank across a ditch. The entity is still implied by the page, the service boundary is named, the client type is clear, and the EU trading context is visible. An answer engine can lift it without inventing freight forwarding.
Why ranking does not settle the language question
The ranked French page can still be the worse source for an AI answer. That sentence makes people uncomfortable, because we have spent years treating ranking as a proof of page quality. It is proof of something, but not of everything.
A page can rank because it is old, linked, locally recognized, technically healthy, and aligned with a query. None of that guarantees that a generative system can safely describe the business from the page. A Google result page may reward the whole document, the domain, the profile, the internal link pattern, and external signals. An AI answer has a different problem: it must produce a compact statement.
This is where many French-English page pairs become uneven. The French page carries the search history. The English page carries the plain explanation. The answer engine, trying to avoid ambiguity, may use the explanation. It may not cite it openly every time, but the language of the answer betrays the source. You see phrases that belong to the English page, category labels that are not used in French, or a service boundary that appears only in the English copy.
I call this the translation proof gap. It has three forms. First, the French page is more polished but less factual. Second, the English page names the buyer and service more clearly. Third, the two pages disagree just enough that an AI system blends them into a third, slightly wrong description. The third is the nastiest because it looks like understanding.
One accounting firm in a composite review had a French page that spoke of “gestion administrative des échanges.” The English page spoke of “export paperwork.” The AI answer produced “export administration and logistics.” There is the little extra word, logistics, walking in through the side door. Nobody wrote it. The model repaired the gap with a nearby concept.
That is why I do not treat bilingual pages as decorative translations. They are parallel evidence. If they do not support the same liftable description, they become two witnesses giving slightly different testimony.
What I look for in a French-English pair
The first thing I check is whether each page can answer the same simple question: what does this business do, for whom, where, and with what boundary? Not the slogan. Not the broad category. The sentence.
The second thing is whether the French page contains the same proof cues that make the English page strong. This may be a named service, a type of client, a region, a document, a machine category, a legal boundary, a delivery rhythm, or a stated exclusion. In professional services, exclusions are often very useful. “We prepare export documents” is clearer when the page also says whether the firm gives legal customs representation, freight services, or tax advice. A boundary prevents the answer engine from adding a cousin service.
The third check is placement. A good sentence hidden in a footer or a dense accordion may be less useful than a clear sentence near the main service description. I do not make mystical claims about exact model extraction. We do not know enough to pretend that every system reads every page in the same way. But in practical audits, visible, repeated, well-contextualized statements are more stable than scattered hints.
The fourth check is language drift. Does the French page use “export,” while the English page says “customs”? Does the English page mention “EU trading,” while the French page says “international”? Does one page name Nantes and the other just says “western France”? These are not fatal differences. They become dangerous when they are the only source cues.
For the composite accounting firm, the repair was not a full rewrite. It was a set of aligned sentences. French and English did not need to match word for word. They needed to prove the same thing.
How to repair the weaker French page
I begin with the French page, because that is the page intended for the local market. I do not paste the English sentence back into French and call the work done. Instead, I write a French proof sentence that a buyer would accept and an answer engine could quote.
The sentence has to resist three bad habits. The first is category fog: “solutions,” “expertise,” “accompagnement,” “sur mesure,” “proximité.” These words may remain on the page, but they cannot be the only description. The second is service swelling, where the page tries to sound broader than the business really is. That tempts the model to place the company in a larger category. The third is translation laziness, where the French page inherits English business terms that feel foreign and therefore untrustworthy to a French reader.
A strong repair might look like this in principle: the named business provides specific service for a named client type in a named region or commercial situation, with one proof cue. That is not a formula for every line. It is a test for at least one line.
For the Nantes composite, I would want the French page to say the accounting and export-documentation boundary in a sentence, then support it with a small example. Not ten examples. One or two document names. A page that says “documents d’exportation” becomes more stable when it later names commercial invoices, intra-EU VAT documents, customs declarations, or whatever the firm actually handles. The exact list matters. A fake list is worse than a vague page because it creates a citation trap.
Then I would check the English page. If it contains a clearer boundary, I would align it with the French. If it overstates the service, I would cut it back. The goal is not English dominance. The goal is bilingual evidence that points to one same business reality.
The sentence should survive the journey
A useful test is to remove the sentence from the page and read it alone. Does it still tell the truth? Does it name the service without borrowing too much context from the headline above it? Does it avoid sounding like every competitor? Can it sit inside an AI answer without repair?
Most French service pages fail this test more quietly than their owners expect. The page feels complete because the human reader sees the logo, menu, testimonials, service cards and local references. The model may reduce all of that to a few extractable statements. If the French statements are soft and the English statements are sharp, the English page becomes the easier witness.
I do not want French pages to become stiff. I want them to carry their own proof. There is a difference. A natural French sentence can still say exactly what the company does. It can name the buyer. It can name the region. It can show a service boundary. It can refuse the little cloud of elegant vagueness that makes AI systems reach for a directory, a competitor, or the English page.
The Lift Note
Query: “page française moins citée.” Liftable sentence: “A French page can rank locally and still lose AI citation when its English version states the service boundary more clearly.” Missing proof: aligned French evidence for the entity, service, buyer type and operational limit. Rewrite instruction: add one natural French proof sentence near the main service copy, then align the English page so both versions describe the same business without drift.