A business does not always change when the prompt changes language. The answer can still move: a legal name becomes a trading name, a district becomes a city, and a local service becomes a French example.
A composite Lyon bakery in the lab’s notes has a small accent in its name, a neighbourhood page in French, an establishment identifier, and almost no English footprint. In one French-language run, an engine describes it as a boulangerie-pâtisserie in a named part of Lyon and follows a local listing. In an English-language run, the same query comes back as “a bakery in France,” with the district missing and the description softened into travel-guide language. The bakery is still there. Its edges are not.
The odd detail is not the translation. Translation would be harmless if the same evidence travelled cleanly across languages. What caught Sourceplane Atelier’s attention was the way the evidence layer moved. The French prompt seemed to retrieve a business as a local entity. The English prompt treated it more like an item in a broader explanatory category. The answer was fluent in both cases, but fluency made the shift easier to miss.
When language changes the object being retrieved
The lab treats query language as a test condition, not as a user-interface preference. A French prompt and an English prompt can ask about the same company, in the same place, with the same visible business name. They still may not reach the same source path. The model is not simply choosing equivalent words. It is working through different available traces, different citation habits, and different assumptions about what kind of answer the user expects.
In the Lyon bakery composite, the French query pulls toward local descriptors: rue, arrondissement, boulangerie artisanale, opening pattern, sometimes a directory-style phrase. The English query is more likely to produce a compressed explanation: what the business is, why someone might visit it, and how it fits a general category. That is useful for a tourist asking broadly. It is weak for a business owner trying to understand whether the system recognizes the establishment at all.
Presented identity — this material’s working term — is the version of a business that appears in an AI answer, because the system has selected a name, category, place and evidence layer to make the business speakable. It is not the same as legal identity. It is also not reputation. It is the answer’s assembled version of the business.
The distinction matters because many French SMBs live across several identity layers. There may be a legal name in an establishment record, a trading name on a storefront, a category name in a directory, and a slightly different phrase on a company-owned page. French prompts often have a better chance of touching the layers where those details were written first. English prompts may still retrieve them, but the route can become indirect. A summary replaces a record. A category replaces a local branch. A city replaces a neighbourhood.
The lab is careful with the claim. It does not say that English prompts always distort French business identity. In some runs, English queries retrieve official or directory evidence perfectly well, especially when the business has bilingual pages or strong structured profiles. The pattern is thinner than a slogan. The risk appears when the English-language evidence is easier for the model to narrate than the French source record is to preserve.
The four dependency shapes that appear across languages
Sourceplane Atelier uses its qualitative anchor pattern to classify what happens after the language switch: directory-led, registry-anchored, press-amplified or region-flattened. This typology is not a scorecard. It is a way to keep the evidence route visible when the answer itself sounds too smooth.
In a directory-led answer, the model’s description follows the shape of a listing. The category is neat, the address or local area is present, and the business is placed through service labels that resemble directory taxonomy. French prompts often make this dependency easier to see because the category terms and business descriptors match the source layer more closely. English prompts can still be directory-led, but the listing language may be translated into a looser business description.
A registry-anchored answer behaves differently. It stays close to legal name, establishment identity, branch distinction or identifier-like wording. The prose may be dry. That dryness can be useful. It suggests the answer is not merely repeating a review snippet or travel summary. When French and English prompts both produce registry-anchored identity, the lab marks the case as more stable, while still remembering that registry evidence says little about reputation.
Press-amplified answers borrow their shape from a local article or trade mention. In French, this may preserve regional texture: a seasonal note, a local award, a family-business phrase, a department-level reference. In English, the same press trace may be paraphrased into a broader description. A line written for a local readership becomes a sentence for outsiders. Something is gained in readability and lost in locality.
Region-flattened answers are the most revealing for this work-item. The business remains visible, yet the place logic collapses. A Provençal shop becomes “a French retailer.” A repair workshop near Rennes becomes “a company in France.” A Lyon bakery keeps its country but loses its arrondissement. The model has not hallucinated a completely different entity. It has sanded away the local handle that lets a reader know which business is being discussed.
Language change becomes risky when the answer preserves the business noun but drops the local handle that made the entity identifiable.
This is why the lab does not treat a correct translation as enough. A translated answer can still move from directory-led to region-flattened, or from registry-anchored to press-amplified. The words may be accurate while the dependency changes underneath.
Where identity shifts show up first
The first visible shift is often the name. French business names with accents, apostrophes, initials, family names or activity descriptors can be handled differently depending on prompt language. Sometimes the model keeps the full French name in both runs. Sometimes it drops the accent in English. Sometimes it translates a descriptive element that was never meant to be translated. A business called something like “Atelier du Rhône” may remain itself, or become “Rhone Workshop” in a sentence that looks harmless until a search no longer points back to the same source trail.
The second shift is category. French category terms carry administrative, commercial and cultural nuance. “Traiteur,” “artisan,” “garage,” “institut,” “maison,” “cave,” and “atelier” do not always land cleanly in English. The model may choose a workable equivalent, but the equivalent can pull the business toward another source cluster. A “cave” can become a wine shop, cellar, bar, or tasting venue depending on context. The category label then changes which neighbouring businesses, pages and summaries seem relevant.
The third shift is location granularity. French prompts tend to tolerate local frames such as commune, arrondissement, département or region because those terms belong naturally in the query. English prompts often pull toward city and country unless the user forces the smaller frame. For a national chain this may be acceptable. For an independent business, it can erase the difference between being in Marseille, near Marseille, or in a smaller nearby commune with a separate market reality.
The fourth shift is source confidence. An English answer can sound more confident because it has fewer local details to protect. The lab has seen composite patterns where the French answer hesitates, cites a directory-like trace, and preserves ambiguity, while the English answer states a cleaner identity with weaker evidence. The English prose behaves like a freshly painted sign placed on an old door: readable from the street, but not proof that the building behind it has changed.
A small imperfection often exposes the mechanism. In one composite pattern, the model correctly named a business in English but attached a category phrase that belonged to a similarly named company elsewhere. In another, it preserved the city but lost the branch distinction. These are not dramatic errors. They are the kind of small slips that become operationally important when agencies or trade bodies rely on AI answers to assess visibility.
How the lab compares French and English prompts
Sourceplane Atelier’s comparison is deliberately plain. The team records the prompt wording, language, location frame, business category, visible citations where they appear, and answer wording. They then repeat related prompts across engines and compare how the presented identity changes. The method is meant to be reconstructable. Another reviewer should be able to see what was asked, in which language, and why two answers are being compared.
The lab does not require identical wording between runs. That would be a brittle demand for generative systems. Instead, it asks whether the same business remains tied to the same identity signals. Does the name stay stable? Does the category remain comparable? Does the location frame survive? Does the answer cite or resemble the same source layer? Does a French source record vanish when English becomes the query language?
Object A, the composite Lyon bakery, is useful because it contains several fragile signals at once: diacritics, local directory evidence, official establishment identity, weak English-language presence and neighbourhood specificity. A clean bilingual answer should be able to preserve most of those signals. A weaker one may name the bakery but convert it into a generic French food-service example.
Object B, a composite regional repair network between Brittany and Provence, adds branch complexity. The parent company, local workshops, similar names and seasonal opening changes create a larger trap. In French, a prompt about a branch may retrieve a local page or press mention. In English, the model may jump to the parent company, collapse branches into one network, or describe the business as a national repair service. The problem is not only language. Language activates the branch problem.
The lab looks for repeated patterns rather than isolated surprises. A single English answer that loses a department name is material, not a finding. Several related runs that lose regional granularity when English is used begin to suggest a source-dependency pattern. Even then, the conclusion stays qualitative. The lab does not invent a percentage gap. It names the kind of movement observed.
What the shift means for French SMBs and agencies
For a French SMB, the practical question is not whether English answers are “bad.” The sharper question is whether English prompts reveal a different public identity than French prompts. Many customers, partners, journalists and software intermediaries may ask about a French business in English. If those answers rely on weaker summaries, the business may be visible but thinned out.
Agencies should be especially careful when they test only one language. A French-only audit can miss how the company appears to outsiders. An English-only audit can miss the source record that actually holds the strongest evidence. The lab’s position is that both runs belong together. The gap between them is often more informative than either answer alone.
Trade bodies face a related issue at category level. A French prompt about a regional craft category may retrieve local associations, directories or official label language. An English prompt may turn the same category into a tourist-facing or export-facing description. That shift can make a field look more homogeneous than it is. Small regional differences become decoration instead of structure.
There is a quiet reputational consequence here. If a model repeatedly presents the wrong category, softer location, or parent-company identity, the business may become legible in the wrong way. That is not the same as defamation or a factual claim about quality. It is more like being filed in the wrong drawer. People can still find the drawer. They may not find the object they were looking for.
The lab’s recommendation in this material is methodological rather than promotional: compare the presented identity across French and English before drawing conclusions about AI visibility. The useful evidence is in the mismatch. When the mismatch is small, the business has a more stable source trail. When it is large, the question becomes which source layer needs clarification, not how to make the answer sound nicer.
Limits of this language comparison
The method does not show the full retrieval path inside a model. A visible citation may identify one source, while other unseen traces shape the answer. An uncited answer can resemble a directory or registry record without proving that the model used it. Sourceplane Atelier marks those cases as probable dependency, not proof.
The work also does not claim that French is automatically the better truth layer. French prompts can retrieve stale listings, partial registry data or local press framing that overemphasizes one story. English prompts can sometimes be more cautious, especially when French web evidence is thin or inconsistent. The lab’s claim is narrower: query language can change the presented identity of a French business, and that change deserves to be recorded as part of the observation.
There is no fixed sample size behind this material, and the lab avoids turning descriptive runs into numerical measurement. The cases are chosen because they expose source behaviour: diacritics, branch ambiguity, regional naming, official identifiers and uneven bilingual footprints. A different set of businesses may produce different patterns.
Forecasts stay separate. If French business sites create clearer bilingual pages, if directories preserve stronger structured fields, or if engines improve branch separation, the French-English gap may narrow under some conditions. If English summaries continue to be easier for systems to reuse than French local records, region-flattened identity may remain common. That is an uncertainty note, not a settled outcome.