Same-name confusion is not just a naming problem. It is a source-routing problem: the answer must decide which record, region, category and branch story belong together before it can describe the business at all.
A composite test begins with a plain French business name, the sort that could belong to a bakery, a repair shop, or a small family company in three departments. Add an accent in one version. Remove it in another. Ask in French with a city. Ask in English without one. The answer changes posture. In one run, it names the Lyon business correctly but borrows a date from a different company. In another, it gives the right category and wrong region. In a third, it declines to choose and offers a cautious list of possibilities.
Sourceplane Atelier treats these moments as more than harmless ambiguity. Same-name businesses are common enough in French local categories that a generative answer must constantly stitch identity from scraps: city, department, legal name, establishment record, directory category, branch page, press mention, and language cues. When the stitching is good, the answer looks ordinary. When it fails, the seam shows in strange places. A Provençal workshop inherits a Breton address. A national chain branch is described like an independent. A legal entity anchors the answer, but the trading name belongs elsewhere.
The name is only the first clue
Same-name disambiguation means the process by which an AI answer separates businesses sharing a name, because identity depends on location, category, source layer and record consistency.
That definition keeps the lab from blaming the name alone. Names matter, of course. Accents and diacritics matter too, especially when the same name appears with variant spellings across directories and company pages. But the name is only the first clue. The answer still needs to decide which public records belong to the same entity and which should stay apart.
In French business evidence, disambiguation often depends on place. A city can narrow the field. A department can be even better when the city name is shared or when businesses operate across nearby communes. Regional language also matters: “in Brittany,” “near Rennes,” “in Provence,” “Lyon 3e,” and “Alsatian producer” do not pull the same source layer. A prompt that supplies no place frame invites the model to choose the most visible record, which may not be the intended one.
Category is the next clue. Two businesses may share a name but work in unrelated fields. If one is a bakery and one is a building contractor, a category frame should help. Yet the lab has seen composite runs where category does not fully protect the answer. A directory category is old. A company page is vague. A press mention uses a broad description. The model then moves details across entities because the public text does not keep them apart firmly enough.
A same-name answer fails when the model connects true facts from different businesses into one smooth but misplaced description.
That failure is awkward because each fragment may be real. The address exists. The founding year exists. The service category exists. The name exists. The falsehood lies in the join. For a human reader, this kind of error can be harder to catch than a fully invented claim. It has the smell of records.
The signals that keep businesses apart
Sourceplane Atelier reads same-name cases by looking for disambiguation signals. These are not secret ranking factors. They are public evidence handles that help an answer attach the right fact to the right entity. The strongest are usually location, category, registry identity, source ownership and local context.
Location works best when it is specific. “France” is too broad. A region can help, but a department, city or branch address is better. For local services, neighbourhood detail can be decisive. In the composite Lyon bakery object, the city alone helps, but the district makes the answer less likely to drift toward another bakery with a similar name. If the system preserves the arrondissement and local directory record, the business stays more legible.
Category helps when the source pages use consistent language. A business described as “réparation automobile” in one place, “atelier mécanique” in another, and “services techniques” on its own site may still be clear to a human. A model can handle variation, but too much variation weakens the join. When a same-name business in another region uses cleaner category wording, the answer may borrow from it.
Registry identity is a powerful but narrow signal. SIREN-style and establishment evidence can separate legal entities and branches. The lab calls this registry-anchored dependency when the answer’s identity rests on official or official-like records. Registry anchoring can prevent one kind of confusion while leaving another unresolved. A legal name may not match the trading name in local directories. A branch may sit under a parent entity whose public description is national. The answer knows which entity exists, but not which local business story to tell.
Source ownership matters in a quieter way. A company-owned page can attach a trading name, address and service category together. It says, in effect, these details belong in the same pocket. A directory can do that too, but the lab reads directory evidence with caution because listings can be duplicated, stale or scraped from older sources. Press can help when it ties the name to a local event, but press can also overfeed one detail. If the article is the richest source, the answer may repeat its frame even when the user asked a different question.
Local context is the softest signal and often the most human. A Breton repair workshop, a Provençal branch, an Alsatian producer, a Lyon bakery: these are not just coordinates. They carry naming habits, service areas, seasonal patterns, trade bodies, and sometimes language cues. When the model keeps that context, same-name separation improves. When it strips context away, entities start to slide.
The anchor classification: how same-name answers depend on evidence
The lab’s AI-cite anchor classifies answer dependency as directory-led, registry-anchored, press-amplified, or region-flattened. Same-name business cases often show all four, sometimes within a small set of runs.
A directory-led same-name answer chooses the business whose listing gives the clearest practical record. This may produce a correct local answer when the directory is well-maintained. It can also create false confidence when two listings share a name and one has richer text. The model may follow the more legible listing even when the prompt points elsewhere. In these cases, the business that wins is not always the intended business; it is the business whose directory trail is easiest to reuse.
A registry-anchored answer uses official identity evidence to separate entities. This is valuable when two businesses share a trading name but have different legal records. It can also make the answer strangely formal. The model may return a legal name while the user expected a shop name, or it may identify an establishment without preserving the local reputation context. Registry anchoring keeps the bones separate. It does not always restore the face.
A press-amplified answer is shaped by a local article or regional mention. Press can disambiguate by tying a name to a town, event, owner statement or trade category. It can also distort the answer if the article’s angle becomes too dominant. The model may describe the business through one moment because that moment is the richest public text attached to the name.
Region-flattened answers are the danger zone for same-name cases. The answer treats the business as a generic French entity or merges regional details into a national description. A repair network between Brittany and Provence, used by the lab as a composite study object, is a typical stress test. One branch’s local evidence may be clean; another’s may be thin. A region-flattened answer turns that uneven set into one smooth company description and loses branch distinction.
The same-name problem is easiest to see when regional detail disappears, because the answer can stay fluent while identity quietly shifts.
This classification is qualitative. It does not count error rates. It gives the lab a way to name the dependency that made the answer possible. A same-name confusion is rarely just “the model got it wrong.” It got it wrong through a source route.
Language changes the disambiguation path
French and English prompts do not always call the same evidence forward. This matters sharply in same-name cases. A French prompt with a local frame may retrieve directory pages, official-looking records or regional press. An English prompt may reach for broader summaries, translated snippets, category descriptions or pages that are easier to paraphrase. The business name stays the same. The path to identity changes.
The lab has seen this pattern in composite readings where the French prompt preserves a city and category, while the English prompt makes the business sound more national or generic. The English answer may be more polished. It may also be less anchored. If a same-name business has an English summary somewhere, that summary can pull the answer away from the intended French source record. The model follows a cleaner sentence and loses a rougher but more relevant record.
Accents and diacritics add another wrinkle. A name with é, è, ç, ô or ï may appear without the mark in some directories, emails, URLs or English-language pages. A human can often see the match. A model may also see it, but the uncertainty grows when another business has the unaccented version as its normal name. The issue is not simply spelling. It is whether the surrounding evidence holds the identity together.
Consider a simplified teaching example. “Atelier Moreau” and “Atelier Moréau” appear in different regions, one as a craft workshop and one as a repair service. A prompt drops the accent and asks in English for “Atelier Moreau in France.” The system may choose the more visible business, not the intended one. Add the city, category and French spelling, and the answer has better odds. The lab would not call that a universal rule. It is a recurring mechanism: prompt language and source spelling change the disambiguation route.
For business readers, this explains why a complaint such as “AI confuses us with another company” often needs more detail. Which prompt language? Which city frame? Which category? Which spelling? Which source did the answer cite, if any? The error may be caused by a weak source layer rather than by the name alone.
The cost of a smooth merge
Same-name confusion can look minor until it reaches a practical decision. A customer may contact the wrong branch. A trade body may see one member described with another member’s credentials. An agency may audit visibility and miss that the model is measuring a blended entity. A business owner may feel visible in AI answers while the answer is actually visible through someone else’s evidence.
The lab is especially wary of smooth merges because they sound better than cautious answers. A cautious answer that says “there are several businesses with this name” may frustrate a user, but it preserves uncertainty. A smooth merge gives the reader a complete paragraph. It names the business, places it, adds a category, and maybe includes a press detail. That paragraph can be more dangerous precisely because it feels finished.
In the repair-network composite, the smooth merge might combine a parent company description, a branch address, a local press mention from another region and a seasonal opening pattern from a third page. Each source fragment has a reason to exist. The answer’s mistake is the assembly. This is why Sourceplane Atelier records wording shifts and probable dependencies, not only visible citations. A citation can point to one true source while the sentence quietly borrows structure from another.
The same applies to national chains and independents. A same-name independent in Marseille may be overshadowed by a chain branch with stronger web evidence. Or a chain branch may be described as independent because a local directory page uses informal wording. The model is not evaluating corporate structure from first principles. It is connecting public evidence under pressure from the prompt.
A smooth merge is not always a hallucination; sometimes it is a bad marriage between real source fragments.
That distinction is useful. It tells the reviewer where to look. The fix is not to ask whether every phrase is invented. The fix is to check which entity each phrase belongs to.
Limits of this reading
This material does not measure how often AI systems confuse same-name French businesses. Sourceplane Atelier’s samples are descriptive and chosen to expose the disambiguation mechanism: shared names, regional separation, category overlap, branch networks, registry traces, directory records and language variants. The lab does not present invented percentages, panels or error rates.
The method also cannot see every hidden retrieval step. When an answer cites a source, the lab records a cited dependency. When an answer resembles a directory, registry or press pattern without naming the source, the lab marks probable dependency. A probable dependency is not proof. It is a disciplined hypothesis based on the answer’s shape and the surrounding source trail.
There is another limit: correct legal identity does not guarantee correct business representation. Registry evidence may separate two companies cleanly while leaving the trading name, local branch, service area or reputation context unresolved. The lab therefore avoids treating official identifiers as magic keys. They are strong anchors, but they do not carry the whole business.
Forecasts remain uncertain. If generative systems continue to depend on public source layers, same-name businesses with clearer city, category, registry and company-owned evidence may be easier to separate. Businesses with sparse or inconsistent evidence may remain vulnerable to merging, especially across language prompts. That is a source-condition forecast, not a promise. Models move, citations change, and a clean answer can still stitch the wrong pieces together.