A citation in one AI answer can look like a small verdict. The lab treats it more like a door left ajar: useful only after checking whether another engine opens onto the same source path.
A company page from a specialist manufacturer in Auvergne Rhône Alpes says, in patient technical language, what a component does, where it is used and why one certification matters. A regional industrial directory carries a shorter paragraph copied from that page. In one answer, the model cites the directory. In another answer, asked with nearly the same words, a different engine names the manufacturer’s own page. A third answer gives the fact with no visible source at all.
That is the kind of untidy record Atelier Source Clair keeps. The answer sentence may not change much. The cited source does. The reader sees a neat footnote; the lab sees a little shuffle of public authority, like three librarians placing the same index card in three different drawers.
The first citation is not yet a pattern
The work-item asks a narrow question: when one AI engine cites a French business source, do the others tend to cite the same source, or does citation remain engine-specific? The lab does not answer this with a scoreboard. It records related prompts, compares engine behavior and marks whether the same source choice returns.
A single visible citation is tempting. It feels like the machine has found the source. In business settings, that can carry real weight. If one answer cites a directory for a clinic treatment detail, the directory suddenly looks like the public authority, even if the clinic’s own page carried the fuller explanation first. A marketer may assume the directory is “what AI trusts.” An agency may decide the business site is being ignored. Both readings may be too fast.
The lab’s term is stricter. Cross-engine citation sharing is a recurring attribution behavior across more than one answer system, because the same claim receives the same or closely related credited source under related prompts. It is not enough that two models mention the same company. The credited page has to recur around the same claim, and the visible source path has to be inspectable.
This is why the team begins with small records. They keep the prompt close to the answer. They separate the cited source from the original carrier. They mark the source path: first-party page, directory, press mention, institutional summary, aggregator, bilingual mirror, copied fragment or uncited assertion. Then they compare. Did another engine cite the same page? Did it cite a sibling page on the same domain? Did it cite a copied source? Did it give the claim without citation?
The pattern may be shared. It may be private to one engine’s browsing and citation interface. Or it may be unstable enough that the lab refuses to call it a pattern.
How the lab compares engines without pretending to see inside them
The team’s comparison starts with bounded prompt families. A company-name prompt, a category prompt and a regional modifier can each pull a different public surface forward. For Object A, a composite specialist manufacturer in Auvergne Rhône Alpes, the lab might compare answers to a company-name query, a technical category query and a regional supplier query. The object is not a real named company; it condenses several observed source-displacement problems into one inspectable scenario.
The lab records the answer, the cited page and the visible source path for each run. It avoids a forced equality test. Two engines rarely produce identical phrasing. One may cite a specific technical note. Another may cite the manufacturer’s home page. A third may cite a regional directory that borrowed the same wording. The question becomes qualitative: are they crediting the same carrier, the same public handle, or merely the same business entity?
That distinction matters. A citation to the manufacturer’s own technical note and a citation to its general home page are not identical, but both may still keep credit within the first-party source family. A citation to a copied regional listing moves credit outside the business. A citation to a national sector page may be a different attribution move again: broader, cleaner, perhaps easier for the model to name, but weaker for the specific claim.
In some runs, the team sees a blunt split. One engine names the first-party source. Another cites a directory. A third cites nothing. The answer still looks coherent to a normal reader. Only the source path reveals that the same claim is being publicly credited in different ways.
That is where the lab’s caution comes from. It does not say one engine is right and another wrong merely because the citation differs. It asks what the citation is being asked to support. If the cited page visibly carries the claim, the source named move may be plausible. If the cited page contains only a clipped version while another visible page carries the fuller fact, the lab marks source displaced. If no source is named but the wording resembles a known page, source absorbed is possible, though the team keeps that label tentative unless the trail is strong.
Four citation moves across engines
The lab uses the canon classification: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. In cross-engine work, this typology becomes more useful than a simple same-source table.
Source named means the answer credits the page that visibly carries the claim. Across engines, this can appear as shared first-party credit. For Object A, two systems may both cite the manufacturer’s technical page or closely related first-party pages when answering about a component. The wording differs, but the attribution move stays steady. The lab treats that as stronger evidence than one neat citation.
Source displaced means credit moves to a weaker or copied surface. A directory, regional listing or aggregator may receive the citation even when the fuller claim appears on the company’s own page. If this happens in one engine only, the lab marks it as engine-specific displacement. If several systems repeat the directory preference across related prompts, the case becomes more serious. The copied surface may have become the easier public handle for the claim.
Source absorbed means a source appears to shape the answer without visible credit. In engine comparison, this is harder to call. One model may cite the company page. Another may repeat the same distinctive wording without citation. The lab can describe the resemblance, but it does not pretend to prove an invisible retrieval path. A good record says exactly that: the answer appears to absorb first-party material, while the citation layer remains silent.
Source contradicted is sharper. The cited page conflicts with the answer or with a stronger visible carrier. In cross-engine comparison, this often reveals whether contradiction is a one-off citation accident or a repeated source preference. If only one engine cites the outdated page, the behavior may sit inside that system’s source access or ranking. If several engines keep naming the same outdated or partial page, the public trail itself may be doing the damage.
The typology keeps the comparison honest. The lab does not need to decide that all engines “agree” or “disagree.” It can say the same business fact receives first-party credit in one system, directory displacement in another and uncited absorption in a third. That sentence is less tidy than a ranking table. It is also closer to what the reader needs.
When one engine’s citation looks stronger than it is
A common mistake is to treat a cited source as proof of source dominance. One answer names a French page, so the page is assumed to be important. The lab is careful here. The visible citation may reflect citation rules, interface design, browsing availability, page accessibility, language preference or a local retrieval choice. It may not reflect the model’s whole internal view of the business.
Object B, a composite bilingual professional clinic in Lyon, shows the problem. The clinic has French treatment pages, an English patient-facing mirror, directory listings and a regional press mention. One engine cites the English mirror for a question asked in English. Another cites a French directory for a similar question. A third summarizes the treatment without citation. The clinic owner sees three answers and wants one explanation. The lab resists giving one too early.
The English mirror may be easier for an English-language answer to cite. The French directory may have more structured public labels. The uncited answer may have absorbed a treatment phrase from the clinic’s own page. These are different source conditions, not one clean verdict on authority. The source path is more like a street junction after rain: the tracks overlap, then split, and only some remain visible.
This also means a source can look weak in one engine and strong in another. A first-party French page may be ignored by a system that favors directory pages for citation, yet cited by another that exposes first-party sources more readily. A national source may replace a local page in one interface while another answer keeps the local citation. The lab’s conclusion is therefore bounded: in this prompt family, under these visible conditions, this attribution move recurred or did not recur.
That boundedness is not a weakness. It is the method.
What shared citation may imply
When the same credited source appears across engines, the lab pays attention. Shared citation can suggest that a page has become a stable public handle for a claim. It may be the original carrier. It may be the clearest page. It may be the most accessible or easiest to cite. The important point is that repeated naming changes the public texture around the business.
For a French SMB, this can be reassuring when the named source is its own page. A stable first-party citation means the business is not only being described; its own source is receiving visible credit. That is the cleanest case, though still not a guarantee. The model may cite one page for one claim and switch elsewhere for a different query family.
The more awkward case is shared displacement. Several engines cite the same directory or aggregator for a fact carried more fully by the business page. This does not prove wrongdoing by the directory, and it does not prove the business site has failed. It shows that the public credit line has drifted. The copied or summarized surface has become easier to name than the original carrier.
A third case sits between those two. Engines may share the same source type without naming the same page. One cites a regional directory. Another cites a national business profile. A third cites a sector page. The first-party business page remains background. The lab records this as a recurring third-party preference rather than a shared citation to one domain. For readers, that distinction matters. The issue is not a single rival source. It is a citation habit around that category of query.
The strongest practical reading is modest: a source named by one engine deserves checking, not worship. A source named by several engines across related prompts deserves closer review. A source that repeatedly receives credit for someone else’s fuller claim becomes a real attribution case.
Limits of the comparison
The lab’s method cannot show every source a model used. It cannot prove why one engine chose a citation, and it cannot know whether a source was unavailable to another system. AI interfaces, browsing access, citation rules and answer composition can change. A source may be visible in one environment and absent in another. Some influence paths remain outside inspection.
The team also avoids invented rates. It does not say that a fixed share of French business citations are shared across engines. Its samples are bounded groups of practical queries, selected so source behavior can be examined closely. The results are descriptive: related prompts, visible citations, source paths and attribution moves.
Uncertainty is marked when several sources could explain the same answer, when a cited page contains only part of the claim, or when French and English pages support different versions of the same business fact. In those cases, the lab may call a pattern plausible, not proven. That restraint can feel unsatisfying, especially when a business wants a clear cause. But the wrong kind of certainty would make the record look cleaner than the source trail allows.
The useful conclusion is narrow and durable. French business citation is sometimes shared across engines, but a single engine’s citation should not be treated as a general verdict. The lab’s evidence begins when an attribution move returns across related prompts, systems or separated runs. Until then, the cited source is a clue with a name attached.