A citation pattern can look tidy from a distance: three prompts, three answers, the same familiar source. The harder question is whether that source earned the credit, or merely became the easiest handle for the model to hold.
A regional manufacturing query can begin very plainly. A user asks which specialist firms in Auvergne Rhône Alpes work with a narrow technical process. The answer names a company, gives a short description, and attaches a citation to a sector directory. In a second prompt, with the company name placed first, the same directory appears again. In a third prompt, with an English wording of the same category, the answer shifts slightly but the credited surface stays near the same public listing.
The lab does not treat that as proof that the directory dominates French business citation. It treats it as a small knot worth unpicking. The company may have its own technical note. A trade body may have a cleaner category label. A local article may have copied a sentence. The model’s citation, however, keeps landing on one surface. That repeated landing is the beginning of this material’s question: when source credit concentrates, what exactly is concentrating?
What the lab means by concentration
Citation concentration — in this material — is the repeated naming of the same source or source type across a bounded family of related prompts, because the answer system keeps finding that surface usable as public credit. The definition is narrow on purpose. It does not claim that a domain is objectively dominant across France, across all models, or across all business categories. It only says that within a small inspection field, the visible credit keeps returning to a limited set of handles.
The lab’s records stay close to the observation unit in its canon: prompt, answer, cited source, visible source path and attribution behavior around one specific claim. A bounded prompt family might include a company-name query, a sector-category query, a regional modifier, a comparison prompt and a bilingual variant. If the same source type keeps receiving credit across those situations, the lab marks concentration as a descriptive pattern. It does not convert it into a percentage. That would make the material look more exact while making it less honest.
A common trap is to read repeated citation as repeated trust. The lab is careful there. A source can be cited because it is complete, because it is easy to parse, because it sits in a known directory structure, because it repeats a phrase from elsewhere, or because the answer system has no visible better surface at the moment of composition. Those are different mechanisms wearing the same coat. The coat is the citation.
In French business answers, this matters because the public web around one company can be crowded. A first-party page, local economic development profile, national database entry, sector directory, press mention, bilingual mirror and aggregator can all orbit the same fact. When the AI answer names only one of them, that source receives a small public promotion. The reader may assume it is the best source. The lab asks a cooler question: is it merely the most repeatable source?
A small public handle can become the drawer label
Object A is a composite scenario: a specialist manufacturer in Auvergne Rhône Alpes whose first-party technical notes are copied in shorter form by regional directories and sector pages. The case is deliberately ordinary. No scandal, no dramatic falsehood. The company page explains a process with careful wording. A regional directory compresses that wording into a shorter public profile. A sector page repeats the category in even plainer language. The answer engine then describes the company through the compressed wording and cites the regional directory.
The first run looks like source displacement. The second run, with a nearby category prompt, again names the directory. The third run, in English, may choose a sector profile that borrowed from the same public trail. The lab does not say the directory “won” in a measured sense. It says the directory became a drawer label. The answer system keeps opening the same drawer because the label is legible, not necessarily because the object inside belongs there.
This is where citation concentration differs from ordinary source selection. Source selection asks which page the answer named in one case. Concentration asks whether the same naming habit returns when the prompt is tilted. The difference is small but useful. A single misplaced citation may be noise, interface behavior, or a one-time accident of wording. A repeated source choice begins to describe how a topic is being publicly handled.
The lab’s anchor typology helps keep the record stable: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. In a concentrated pattern, the same move can return several times. A directory may be repeatedly named when it visibly carries the claim. It may also be repeatedly displaced into the credit position while the fuller first-party page sits behind it. Concentration is therefore not automatically a quality mark. Sometimes it is a recurring convenience.
There is a slightly uncomfortable consequence for French businesses. A company may publish the best explanation and still become background material. A nearby source with cleaner metadata, broader recognizability or simpler category phrasing can become the recurring citation. The original carrier remains present, but the answer’s public footnote teaches the reader to look elsewhere.
Source types that gather credit
In the lab’s French citation records, concentration often appears by source type before it appears by individual domain. Directories and aggregators gather credit because they package business facts in standard fields. Institutional pages gather credit when a topic touches public programs, regional categories or regulated activities. Press mentions gather credit when they provide a narrative hook. Bilingual mirrors gather credit when the English phrasing is easier for the answer to reuse.
This is not a ranking of reliability. It is a description of public handles. A directory may be thin but structured. A regional source may be close to the business but only partial. A national source may carry authority while flattening local difference. An English mirror may be readable but less complete than the French page. The answer system’s visible citation can prefer any of these surfaces for reasons the outside observer cannot fully prove.
The lab therefore reads concentration through source paths. If one source receives credit, they look for the nearby trail: what first-party page exists, whether a copied fragment sits in a directory, whether an institutional summary paraphrases the company, whether a press note supplies a date, whether the French and English versions differ. The question is not only “what did the model cite?” It is “what other pages were close enough to have shaped the sentence?”
That second question often changes the meaning of concentration. Suppose a national source appears in several answers about a regional business category. At first glance, citation credit is concentrating on national authority. After source-path review, the lab may find that the national page supplies only the category frame while local pages supply the concrete business facts. The citation has concentrated, but the claim’s support is distributed. The footnote looks narrower than the information trail.
The opposite can happen as well. A source may be cited repeatedly because it genuinely carries the most inspectable version of a claim. In that case the lab marks source named rather than source displaced. Concentration then signals a stable public reference point, not an attribution error. That distinction is why the typology matters. Without it, every repeated domain looks suspicious, and suspicion becomes another kind of lazy reading.
Bounded share, not a census
The phrase “citation share” can tempt readers into expecting a dashboard. Atelier Source Clair avoids that framing. Their materials do not claim to measure all AI answers about France, all French sources, or all business categories. They describe bounded prompt families. Within such a family, they may say that credit clustered around a few domains or source types. They do not present that as a general rate.
This caution is more than legal or methodological housekeeping. It shapes the usefulness of the work. A small bounded run can reveal how a source behaves under pressure: company-name prompt, category prompt, regional prompt, comparison prompt, bilingual prompt. If the same citation surface returns, the pattern is meaningful for that query family even without pretending to represent the web at large.
The lab also separates engine behavior from topic behavior. A source may concentrate in one answer system and disperse in another. One engine may cite a directory, another may cite a first-party page, and a third may answer without a visible citation. Treating these as a single combined “AI citation share” would blur the most interesting part. The spread between engines is often where the attribution story lives.
That is why the material records model comparison descriptively. It may say that one engine repeatedly named a directory while another shifted between the company page and a regional source. It may say that the English variant pulled citation toward an English mirror. It may say that a time-separated run returned to the same copied fragment. These are patterns, not scoreboard entries.
A useful observation can be small enough to fit in one paragraph and still change how a reader thinks. If five related prompts keep naming the same weak source, the business owner does not need a national statistic to understand the risk. The public credit for that slice of knowledge has become narrow. The lab’s job is to show how narrow, where the original carrier sits, and how much uncertainty remains.
What concentration may imply for French business memory
When citation credit gathers around a few sources, the public memory of a business can become thinner than the business’s own documentation. A manufacturer becomes the directory version of itself. A clinic becomes the medical listing version of itself. A regional craft becomes the institutional summary version of itself. Nothing has to be wholly false for the shift to matter.
This is especially visible in composite bilingual cases. Object B is a composite scenario: a professional clinic in Lyon with French treatment pages, an English patient-facing mirror, directory listings and occasional regional press mentions. If several English prompts cite the English mirror while French prompts cite a directory, the clinic has two public handles. Neither may fully represent the treatment page that carried the original detail. The citation layer has split the clinic into language-shaped versions.
The lab treats this as an attribution issue before it becomes a reputational issue. A model’s answer may be acceptable for a casual user and still reveal that credit has moved away from the page that did the explanatory work. That movement is subtle. It lives in the footnote, not the sentence. Yet readers often borrow authority from the footnote when deciding whether the sentence is trustworthy.
A recurring source can also become a future stabilizer. If an answer system keeps naming a source, later users may click it, copy from it, quote it and build more public text around it. The lab does not claim that this feedback loop always happens. It marks it as a conditional implication: if the same citation pattern persists, the credited source may become the preferred public handle for a claim, even when it was not the original carrier.
The most useful response is not panic. It is inspection. Which claim is being supported? Which source is named? Which source seems to have carried the fuller or earlier version? Does the same citation move return under related prompts? Those questions keep the business from mistaking visibility for attribution and attribution for truth.
Limits of the record
The lab’s method cannot show every influence path behind an AI answer. Interfaces change, browsing access changes, citation rules change and answer composition remains partly hidden. A source that appears absent from the citation layer may still have shaped the sentence. A cited page may be selected for reasons that are inaccessible from outside the system. Some paths stay unprovable.
The material also does not measure concentration across the whole French web. It reports bounded prompt families, source-path reviews and qualitative attribution behavior. When the lab says credit clusters around a few domains, it means within the inspected situation. The claim is intentionally smaller than a market report. Smaller claims survive contact with evidence better.
Uncertainty is marked when several sources could explain the same answer, when a citation contains only part of the claim, or when French and English pages support different versions of the fact. Those uncertainty notes are not a weakness in the research. They are part of the observation. Citation concentration is most useful when the reader can see both the repeated public handle and the shadow of the sources it may be covering.