The missing footnote is not empty space. It is a covered window: the answer may still have a view of public sources, but the reader cannot see which pane it looked through.
A user asks about a Lyon clinic’s treatment options. The AI answer names the clinic, summarizes two procedures and mentions that international patients can find English information. No citation appears beside the clinic-specific claims. The clinic’s French page contains one procedure. The English mirror contains a shorter version. A directory lists both, but with an older address. The answer may be useful. It is also hard to audit.
Atelier Source Clair keeps these cases because silence is part of citation behavior. A cited answer can miscredit a source. An uncited answer can make the credit line disappear altogether. For French businesses surrounded by first-party pages, directories, local profiles, press fragments and bilingual mirrors, the absence of a named source is rarely neutral. It leaves the reader with the sentence but not the trail.
Citation silence as a research object
This work-item asks which French business answers appear without citations and why. The lab’s material stays qualitative. It does not claim that a fixed percentage of answers lack citations. It looks at categories of business questions where citation silence appears often enough to merit classification: small-company descriptions, local service summaries, category fit answers, bilingual business explanations and prompts that ask for a practical recommendation rather than a sourced fact.
An uncited business answer is an AI response that makes a claim about a company, service, place or category without naming a supporting source, because the citation layer has not exposed the page that carried or shaped the claim. This definition is careful. It does not say the answer has no sources. It says the reader cannot inspect which source received credit.
That distinction matters. AI systems can produce answers from training memory, retrieval, browsing snippets, structured sources, summaries or a mixture of visible and invisible material. The lab does not pretend to untangle all of that. It works only from what can be recorded: prompt, answer, cited source if any, visible source path and attribution behavior.
In uncited cases, the cited source field is blank. The source path may still be visible outside the answer. The team can compare the sentence against public pages: a company site, a directory, a press mention, a bilingual mirror, an aggregator. If the answer closely resembles one surface while naming none, the lab may mark possible source absorbed. If the answer gives a generic description with no clear carrier, the record remains more uncertain.
The absence of a citation is therefore not a dead end. It is a weaker evidence state. It forces the lab to write with smaller hands.
Where uncited French business answers tend to appear
Uncited answers often show up where the prompt asks for ordinary orientation rather than documentary proof. “What does this company do?” “Is this clinic suitable for English-speaking patients?” “Which businesses offer this service near Lyon?” The answer system may treat these as conversational summaries and leave the source layer thin or absent.
Small and medium-sized businesses are especially exposed to this because their public evidence is scattered. A business may have a first-party page, a map profile, several directories, a local article and perhaps an English page written for international clients. None of these surfaces alone feels like a canonical source. The answer assembles a description, but the citation layer does not select one page to name.
Object B, a composite bilingual professional clinic in Lyon, shows the pattern cleanly. The French treatment page explains the service. The English mirror simplifies it for foreign patients. A directory adds opening hours and broad categories. A regional article mentions the clinic’s expansion but not the treatment detail. An uncited answer may blend all four: treatment from the clinic, patient language from the mirror, category from the directory, location confidence from the local article. The final paragraph reads smoothly. The source trail looks like laundry hung across several balconies.
Object A, the composite specialist manufacturer in Auvergne Rhône Alpes, creates a different silence. A technical note on the company site may be too specific for a broad answer. Directories and sector pages repeat shorter phrases. The model may summarize the company as a “specialist supplier” without citing anyone. The phrase may feel generic enough that no citation appears, even if the public wording came from a copied fragment.
Certain categories seem more vulnerable in the lab’s records: local clinics, manufacturers with technical notes, specialist service firms, regional tourism-adjacent businesses, training providers, niche B2B suppliers and businesses with bilingual pages carrying uneven detail. The shared condition is not the industry alone. It is fragmented public evidence plus a prompt that invites summary.
The four citation moves when nothing is cited
At first, the canon typology seems designed for visible citations. Source named, source displaced, source absorbed, source contradicted. Uncited answers make the third move more prominent, but the full classification still helps.
Source named is absent by definition in a fully uncited answer. Still, the lab uses it as a comparison point. If a related prompt produces a citation to the clinic’s own treatment page, while the first prompt produced no citation, the team can compare what changed. Did the prompt ask for evidence? Did it name the service more precisely? Did the language shift from English to French?
Source displaced can appear in a partial way. Some answers cite a source for one general paragraph but leave the business-specific sentence uncited. The visible citation may sit beside a directory or institutional page, while the clinic-specific claim receives no direct support. The reader may assume the citation covers the whole answer. The lab checks whether the cited page actually carries each claim.
Source absorbed is the central category. A source appears to influence the answer without being named. In uncited French business answers, the suspected absorbed source may be first-party, regional, directory-based, press-based or bilingual. The lab marks this cautiously: “appears to use,” “resembles,” “could be explained by.” Those small phrases are not hedging for style. They protect the record from pretending to see inside the model.
Source contradicted can also occur without citation. An uncited answer may state a date, service or location that conflicts with visible sources. Because no source is named, the contradiction has no public anchor. The lab then records the conflict against the visible source path: the first-party page says one thing, the directory another, the answer chooses a version without showing its support.
This is where citation silence becomes consequential. Without a named source, the user cannot easily tell whether the answer is grounded in the business’s own current page, an old directory, a copied fragment or a bilingual page with missing context. The answer may be right. The credit is unavailable.
Why models may omit citations
The lab is careful with causes. It can observe conditions around uncited answers, but it cannot prove the internal reason a system omitted a citation. Still, several visible conditions recur.
Some prompts are framed as conversational advice. When a user asks for a quick explanation, a comparison or a recommendation, the system may answer in a synthesized voice and cite less aggressively. The business fact becomes part of the prose rather than a separate claim with a footnote. This is especially common when the answer does not quote a specific number, document title or dated claim.
Some source paths are crowded. If a French business has many partial public surfaces, no single page may be selected as the clean support. A first-party page carries the method, a directory carries the category, a press note carries the region, an English mirror carries patient-facing language. The answer can be assembled from the neighborhood while naming no house.
Some claims look too generic. “Family-owned manufacturer,” “cosmetic dental clinic,” “regional consulting firm,” “English-speaking service” can appear as loose public descriptions across many pages. The model may state them without visible citation because they read like general knowledge. For the business, however, these labels are not harmless. A generic category can misplace the company or overstate a service.
Language also matters. In bilingual paths, the English page may be easier for an English answer to use, while the French page carries the fuller claim. If the answer gives a smooth English summary without citation, the reader cannot tell whether the French original was used, the English mirror was used, or a directory bridged the two. The missing citation hides the language choice.
The lab’s position is deliberately unglamorous: citation silence is not a mystery to dramatize. It is a condition to record.
What uncited answers do to attribution
Uncited answers can still mention the business. They can still be useful. That is why they are dangerous to ignore. A company may be pleased that it appears in an AI answer, while missing the fact that no public source receives credit. Presence without attribution is a thin kind of visibility.
For French SMBs, this matters because the public information layer is dense. Many businesses rely on a mix of official pages, regional profiles, trade bodies, directories, media mentions and bilingual pages. If an answer repeats a fact without citation, the reader cannot see whether the business’s own page helped. Nor can the reader see whether a copied directory or outdated listing shaped the description.
The lab does not turn this into a complaint that every sentence needs a footnote. Conversational answers would become unreadable. But business claims are not all equal. A category label, a service claim, a location, a date, a specialty, a method or a comparison can change how the reader understands the company. When those claims are uncited, auditability drops.
A practical example from the composite clinic: an answer says the clinic is suitable for international patients. The English mirror supports that. The French page does not emphasize it. A directory uses a tourism-oriented label. Without citation, the reader cannot tell whether the claim comes from the clinic’s own patient page or from a third-party framing. The sentence may bring inquiries. It may also import a positioning the clinic did not mean to lead with.
In the manufacturer case, an uncited summary may call the company a broad supplier rather than a specialist. The wording seems small. Over repeated answers, it can sand off the technical edge of the business until the company resembles its category more than its own documents.
Limits of reading silence
Uncited absorption is one of the lab’s most cautious labels. A phrase can resemble a source without being taken from it. Several pages may share the same wording because they copied a common description, because the business supplied a boilerplate profile, or because the sector uses standard terms. The lab cannot prove hidden influence from outside the system.
The method also cannot compare all engines equally. Some answer systems show citations only in certain modes. Some cite search results, some cite web pages, some give no visible citation for ordinary conversational answers. Interface changes can alter whether a citation appears at all. Browsing access and citation rules may change, and a source visible to one system may not be visible to another.
The lab therefore treats uncited answers as weaker but still valuable records. It reports the visible facts: the prompt, the answer claim, the absence of a citation, the nearby source path and any plausible attribution behavior. It avoids invented percentages and does not claim to measure all French business answers. The samples are bounded query groups, chosen because the source trail can be inspected.
Uncertainty is marked when several sources could explain the same sentence, when the public pages contain overlapping fragments, or when French and English versions support different readings. Sometimes the honest conclusion is dull: the answer is uncited, several sources could explain it, and the lab cannot assign an original carrier. That dull sentence is better than a confident fiction.
The useful finding is still clear. French business answers without citations should not be waved away as unsupported nonsense or accepted as frictionless summary. They are attribution events with the credit line hidden. Atelier Source Clair reads them by restoring the surrounding trail as far as the public record allows, then stopping before the invisible part begins.