A translated page can behave like a clean window and a distorting pane at the same time. It lets the answer system see the business quickly, while quietly deciding which French details deserve to travel.
A Lyon clinic publishes a French page explaining a treatment in careful, patient-facing language. The English version is shorter. It keeps the appointment wording, removes a few clinical distinctions, and changes one phrase into a broader international term. A user asks an English question about the clinic. The answer gives the treatment detail and cites the English page. The French page, where the fuller explanation sits, remains unnamed.
In another run, the prompt is written in French. The answer cites a directory. A third prompt asks for clinics in Lyon offering the same procedure for foreign patients, and the English mirror returns as the credited source. The lab does not read this as a simple language preference. It reads it as a source-path problem: when two pages carry related facts in different languages, the citation may choose the page that speaks the prompt’s language, even if the original carrier is elsewhere.
The mirror is not always the original carrier
English mirror citation — in this material — is the visible credit given to an English version of a French business page when the answer uses a claim whose fuller, earlier or more precise carrier appears to be the French source. The point is not that English pages are bad sources. Some are carefully written, complete and intended as primary pages. The problem begins when the mirror receives credit for work it did not fully carry.
Object B is a composite scenario: a bilingual professional clinic in Lyon with French treatment pages, an English patient-facing mirror, directory listings and regional press mentions. The lab uses it because the source trail is ordinary and messy in a believable way. The French page explains procedure scope, appointment conditions and practitioner roles. The English mirror is useful for foreign patients but trims the distinctions. A medical directory repeats one phrase. A local article gives a date but not the method. The AI answer may touch all of these surfaces while naming only one.
The lab’s canon separates the cited source from the original carrier. That distinction is decisive here. The cited source is the page named by the answer. The original carrier is the source that appears to have published the claim, method, fact or context in fuller or earlier form. In bilingual business pages, those roles can split. The mirror can be cited while the French page carries the claim with more care.
A mirror can also be the right citation. If the English page is not a shortened translation but a dedicated primary page with its own complete explanation, source named may be the proper classification. The lab avoids assuming that French always comes first. France’s bilingual public web is less tidy than that. Some companies maintain parallel pages. Some update English first for international buyers. Some let an old English page linger with facts that no longer match the French page.
That is why the lab looks at paths, not language loyalty.
How a prompt pulls the citation across languages
The simplest mechanism is prompt language. An English prompt often invites an English answer, and an English answer may prefer an English citation when one is available. That preference can feel natural to a user. The answer is easier to read, the cited page is easier to inspect, and the wording aligns. But ease can hide a transfer of credit.
A category term can intensify the pull. French business pages often use specific terms rooted in local practice, regulation or sector habit. English mirrors may replace them with broader terms that travel better across models. A clinic page that distinguishes between a precise procedure and a more general cosmetic category may be mirrored in English under the general category. An answer system asked in English may then cite the mirror and repeat the broader framing. The business becomes more legible and less exact at the same time.
Regional wording adds another wrinkle. A French page may use commune, arrondissement, département or regional phrasing in a way that makes sense locally. The English mirror may flatten the geography into “Lyon area” or “near central Lyon.” If a user asks an English regional question, the mirror may look like the cleanest match. The citation then supports a geography that the French page would have expressed differently.
In the lab’s records, bilingual variants are therefore not translation checks. They are pressure tests. The same business is queried through French wording, English wording, company-name prompts, category prompts and comparison prompts. If the English mirror receives credit only when the question is English, that is one behavior. If it also receives credit for French prompts, the lab marks a stronger source choice. If neither page is cited but the answer uses terms from both, uncited absorption becomes the more plausible move.
The four citation moves across a bilingual trail
The lab’s classification anchor applies cleanly to bilingual pages: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. A bilingual trail can show all four without leaving one business.
Source named appears when the answer cites the page that visibly carries the claim in the form used. If the English page fully explains the service and the answer credits it, the citation layer is behaving plainly. The lab may still compare it with the French page, but there is no need to force an error.
Source displaced appears when the answer cites the English mirror while relying on the fuller French page. This is the classic mirror problem. The mirror may contain a short version, but the detail in the answer points back to the French source. The visible credit has moved to the translated surface. The reader sees a citation and assumes the mirror carried the load.
Source absorbed appears when the answer borrows language from either page without naming it. The answer may combine a French technical distinction with an English patient-facing phrase, then cite a directory or provide no citation. This move is harder to prove. The lab marks it cautiously, especially when several nearby sources contain similar wording.
Source contradicted appears when the cited page and the stronger visible carrier support different versions of a fact. A French page may give one practitioner list, while an English mirror keeps an older one. A directory may copy the English version. The AI answer cites the mirror and states the old fact. In that case, the problem is not only language choice. The citation supports a version that conflicts with a stronger current trail.
The typology prevents a common overreaction. The lab is not saying that English mirrors should never receive credit. It is asking which citation move occurred. A mirror can be correct, displaced, silently absorbed or contradicted. The answer lies in the path around the claim.
Why English mirrors become attractive surfaces
English mirrors often offer answer systems a cleaner sentence. They reduce local terminology, shorten context and place the business inside internationally familiar categories. For a reader arriving from outside France, that can be helpful. For attribution, it can be risky.
A French manufacturer may write a technical note using sector terms that require background knowledge. Its English page may say the same thing in broad buyer language. The AI answer, especially in English, may cite the broad page because it fits the answer’s phrasing. Object A is a composite scenario: a specialist manufacturer in Auvergne Rhône Alpes whose first-party technical notes are copied by regional directories and sector pages. If it also maintains an English mirror, the mirror can become one more surface that competes for credit against the original technical note.
The mirror’s structure matters. Pages with compact headings, explicit categories, short service descriptions and clear company labels are easier citation candidates than dense pages with embedded PDFs, image-heavy sections or local idioms. That does not mean the structured page is more authoritative. It means it may be easier for the answer system to name.
There is also a reader-side pressure. AI engines often compose answers that feel helpful to the user’s language. Citing a French page under an English answer can be less convenient, especially if the English mirror appears to cover the same topic. The citation layer may optimize for user inspection. The original carrier may lose credit because it is less convenient to the imagined reader.
The lab treats this as a normal behavior to observe, not a moral failure by the system. A bilingual web creates duplicate and near-duplicate surfaces. Some citation choices will be reasonable. Others will expose where translation has become a filter. The work is to tell them apart.
What businesses and researchers can learn from the split
A bilingual citation split can show which version of a business the answer system finds easiest to handle. If the English mirror is always cited for category claims, it may be acting as the public handle for international understanding. If the French page is cited for detailed procedural claims, the split may be healthy. If the mirror is cited for details it does not carry, source displacement is more likely.
The lab’s material does not turn this into a checklist of page fixes. Its purpose is research. Still, the implication is practical enough: businesses should know which language version receives public credit for which claims. A French page may be the source of truth internally while an English page becomes the cited source externally. That gap can matter when facts diverge.
For marketers and agencies, the useful move is to record exact prompts and citations before making claims about “AI visibility.” A business may be visible in both languages but credited differently. It may be correctly described in French and loosely categorized in English. It may receive no citation in French while an English mirror becomes the named source. Presence alone misses the attribution layer.
For researchers, bilingual cases are valuable because they make hidden assumptions visible. The model’s language choice, the user’s query language, the page’s structure and the claim’s original carrier all come into view. A single answer rarely settles the matter. Related runs give the pattern its shape.
The lab is especially cautious with bilingual contradictions. If the French page and English mirror differ, the answer may not be wrong in a simple sense. It may be correct relative to the page it cited and outdated relative to another page. The better question is which source should carry public authority for the claim. That is an attribution question, and it is sharper than asking whether the answer sounds plausible.
Limits of bilingual source review
The lab cannot prove every hidden influence path between French and English pages. AI systems may access one page and not another. They may use snippets, cached representations, search summaries or browsing layers that are not visible from outside. A cited English page may be named even when the French page influenced the answer. The inverse may also occur.
Timing complicates the record. A business may update the French page and leave the English mirror unchanged, or revise the English page for international users while the French page remains older. Without a clear public trail, the lab avoids declaring which page came first. It marks the original carrier only when the evidence supports that reading.
The method also does not say that English citation is weaker by default. Some English pages are primary surfaces for export-facing companies, clinics serving foreign patients or institutions publishing in two languages. The lab’s claim is narrower: when an English mirror receives credit for a claim better carried by a French page, the citation layer has shifted attribution across language. That shift deserves inspection before anyone treats the footnote as neutral.