A national citation can make an answer look settled. The regional source often holds the burrs: the district, the trade context, the local use case, the wording that made the claim specific.
A regional article mentions a manufacturer outside Lyon because the company supplied a narrow industrial component for a local project. The business’s own note explains the technical reason. A national sector page later describes the same component in broader terms. When an AI answer replies to a user asking about French suppliers in that category, the citation lands on the national page. The regional trail is still nearby, but the named source has moved upward.
That upward move is easy to miss. National sources look tidy. They carry institutional tone, broader category labels and cleaner public authority. Regional sources are messier: chamber pages, local press fragments, trade-cluster notes, municipal economic profiles, copied listings. Yet in French business questions, the regional layer often contains the first useful context. It says where the company sits, what local market uses it, which phrase was first attached to the work.
Why broad sources become easy to cite
This work-item asks how AI engines decide between citing a national French source and a local or regional one when both can support the same answer. The lab does not treat the answer as a legal decision about authority. It treats the citation as a visible behavior: which page received credit for a claim that may have moved through several public surfaces?
National-over-regional citation shift is an attribution move where an AI answer names a broad French source while a local or regional source appears to carry the more specific business context, because the broader source offers a cleaner public handle. That definition matters because the national source may not be wrong. It may contain a true summary. The issue is whether it is the best credited carrier for the claim being used.
In the lab’s records, national sources can win for ordinary reasons. They may have clearer titles. They may describe a sector in standardized language. They may be easier to cite for a category question because the prompt itself uses broad wording. A user asks for “French companies in medical device manufacturing,” and the system reaches for a national industry page. A user asks about a regional supplier by name, and the answer may return to a more local source.
But the boundary is thin. A national source can support the category while failing to support the specific claim. It might say that a sector exists, while the regional page gives the company’s actual involvement. It might define a treatment area, while a Lyon clinic page gives the procedure detail. A normal reader sees a credible citation. The lab asks whether that citation carried the sentence it was asked to carry.
The regional layer is not just geography
Regional sources in France are not only maps with smaller boundaries. They can be trade bodies, local directories, regional media, municipal pages, cluster profiles, chamber notes and event pages. They often explain how a business is socially and commercially placed. That placement can be crucial in AI answers, especially when a company is small enough that national coverage is thin.
Object A, a composite specialist manufacturer in Auvergne Rhône Alpes, is useful here. Its first-party technical notes are copied in shorter form by regional directories and sector pages. A national source may describe the industrial category, while a regional listing repeats a clipped company fact, and the manufacturer’s own page carries the full explanation. If an answer cites the national source for a sentence about the company’s specific method, the lab treats the citation with caution.
The national source may be a good background source. It may not be the original carrier. In a clean record, the lab would separate those roles. Background support belongs to one page; claim support belongs to another. AI answers often collapse that distinction. The footnote sits at the end of the paragraph, like a single label on a box full of mixed screws.
Object B, a composite bilingual professional clinic in Lyon, shows a different version. A regional article may mention the clinic opening a service in a particular district. The clinic’s French page explains the treatment. A national health information page explains the procedure in general. When a user asks whether the clinic offers the treatment, a national citation can make the answer look medically grounded while leaving the clinic-specific claim under-supported.
The lab does not punish national sources for being broad. It asks what part of the answer they actually support. If the answer explains a general procedure, a national source may be suitable. If it attributes that procedure to a named clinic, the credited source needs to carry the clinic link as well.
Prompt wording pulls the source upward or downward
Small wording changes can change the citation level. A company-name prompt often invites first-party or directory sources. A category prompt invites broader sector pages. A regional modifier may pull local sources forward, though not always. A comparison prompt can invite national overview pages because the answer needs a frame larger than one business.
The lab therefore builds related prompt sets instead of isolated screenshots. A query about a company in “Auvergne Rhône Alpes” is not the same as a query about “French suppliers” or “manufacturers in France.” The same business fact may travel through all three. Watching where the citation lands tells the team whether national sources appear only when the prompt is broad, or whether they keep receiving credit even when the question is local.
This is where attribution behavior becomes evidence. Suppose a model cites a national page for the broad category prompt. That may be expected. Then it cites the same national page for the regional prompt, while ignoring a regional source that contains the local detail. Then it cites the national page again for the company-name prompt, even though the first-party page carries the claim. The recurring upward pull becomes more interesting than any one citation.
The opposite can happen too. A model may cite a local source for a regional query, then switch to a national source for a category query. That is not necessarily displacement. It may be a reasonable source choice. The lab’s work is to keep the difference visible: a source can be appropriate for one query family and weak for another.
The reader should notice the grain of the prompt. A regional word is not decoration. “Lyon,” “Bretagne,” “Nouvelle-Aquitaine,” or “Auvergne Rhône Alpes” can alter what source deserves credit. If the answer keeps citing national sources despite regional wording, the lab marks that as a source-choice pattern worth checking.
Four citation moves in national and regional cases
The canon typology fits these cases cleanly: source named, source displaced, source absorbed, source contradicted. The same four moves can be applied to the national-regional split without turning it into a score.
Source named appears when the answer credits the page that visibly carries the relevant claim. A regional source can be the named source if it contains the local business fact. A national source can also be named appropriately when the claim is general enough. The lab does not prefer regional sources on principle. It prefers the source that carries the claim.
Source displaced appears when the answer credits the broader national surface while a regional or first-party page carries the sharper context. This is the central pattern for this work-item. The displacement may be gentle: the national source contains a partial summary. Or it may be severe: the national page explains a category but not the named business fact at all.
Source absorbed appears when a regional phrase or first-party detail seems to enter the answer without receiving citation. The answer may describe a manufacturer in exactly the local category used by a regional cluster page, then cite a national overview. The lab marks this cautiously. It can describe the visible resemblance, but it cannot prove hidden influence from the outside.
Source contradicted appears when the cited national page conflicts with the regional or first-party evidence. A regional article may give one date, a national page another. A clinic’s current page may list a service differently from an older national profile. When the model cites the conflicting source and chooses its version, the answer does more than move credit. It changes the public fact.
This typology prevents a lazy conclusion. The question is not “national bad, regional good.” The question is whether the named source, at the level it occupies, actually supports the answer’s claim.
What the shift means for French SMBs and researchers
For a business, national citation can feel flattering. The answer is anchored in a larger source, perhaps one with institutional tone. But if the cited page does not carry the business-specific detail, the public credit may have moved away from the source that did the explanatory work. The company remains described, yet its own evidence and its local context become less visible.
For regional institutions, the issue is different. Their pages may seed facts that later receive national citation elsewhere. A local economic profile, a cluster note or a regional trade article can make a company legible, but the visible credit may go to a broad source that repackages the category. That does not erase the regional source. It does make its role harder for a reader to see.
For researchers, the national-over-regional move is a useful stress test of citation. Broad sources are often plausible enough to hide weak attribution. They rarely look absurd. The answer is grammatically confident, the citation is respectable, and the business fact may be mostly right. The mismatch sits in the distance between general support and specific origin.
The lab’s internal rule is practical: inspect the sentence, not only the domain. A national domain can support a national claim. A regional source can support a regional claim. A first-party page can support a company claim. Trouble begins when one citation is asked to carry all three jobs at once.
That is also why the lab avoids turning this work into optimisation advice. It does not say “get national coverage” or “build regional citations.” The research question is about attribution behavior. Which source gets named when several public surfaces surround the same claim? What does that named source actually contain? What falls out of view when the answer climbs to a broader citation?
Limits and uncertainty in source-level comparison
The lab cannot see every retrieval path behind an answer. A national source may have been cited because the system could access it more easily, because the interface prefers certain source types, because the prompt used broad category language, or because the regional page was unavailable at the time of the run. These reasons cannot always be separated from outside inspection.
The samples are bounded. They are practical French business and topic queries, not a measurement of the whole French web. The lab does not assign a numerical freshness effect, authority score or source hierarchy weight. It records the visible source path and classifies the attribution move. When several sources could explain the same answer, the record says so.
Uncertainty also increases in bilingual cases. A national English-language page about a French topic may compete with a fuller French regional page. The answer may cite the English source for a question asked in English, even when the French page carries richer context. That may be a language-path effect rather than a pure national-over-regional preference. Lise Varnier’s bilingual review keeps those cases from being squeezed into the wrong drawer.
The strongest conclusion is therefore modest. National French sources often make cleaner citations, but cleaner is not the same as closer to the claim. When an AI answer names a national source for a local business fact, the lab follows the trail back down: first to the sentence, then to the regional evidence, then to the original carrier if the public record allows it.