Conflicting French sources do not simply give an AI engine a fact problem. They give it a credit problem: the answer must sound coherent while one page, and not the others, becomes the public witness.
The disagreement looked ordinary. A composite Lyon clinic had one date on its French treatment page, a different date in an English patient-facing mirror, and a third wording in a regional article. The differences were small enough for a hurried reader to pass over. One page said the service began after a specialist joined the clinic. Another suggested the service had been available for longer. The regional article placed the change in a broader local health story and rounded the timeline.
When an AI answer summarized the clinic, it chose one version and cited the regional article. The sentence was fluent. The footnote looked respectable. Only after lining up the pages did the conflict show itself. The cited source did not carry the most precise version. It carried the most narratable one.
Conflict is not always a visible contradiction
A reader imagines conflicting facts as two pages fighting in plain language: one says yes, the other says no. French business source trails are rarely that clean. More often the conflict hides inside category labels, dates, service scope, branch names, old English descriptions, copied fragments and short profiles that have lost context. The answer can combine them without sounding wrong at first.
The lab calls this work discrepancy analysis because the object is the gap between sources, answer and citation. An observation still has to contain the prompt, answer, cited source, visible source path and attribution behavior. A conclusion requires recurrence. A single contradiction is useful, but it is not enough to say that an engine has a stable habit.
Conflict in this material means that two or more visible sources support different versions of the same business fact. The difference can be explicit, as when a directory lists an old address while the company page gives a new one. It can be interpretive, as when a first-party page describes a narrow technical service and a sector profile turns it into a general category. It can be linguistic, as when the French page and English mirror carry similar wording but not the same claim. The lab treats each version as a source position before judging which one the answer elevated.
A working definition helps keep the analysis from becoming a fact-checking sprawl: a source conflict is an inspectable disagreement between public carriers of the same claim, because the answer must either choose, merge, ignore or obscure the difference. The citation then shows which version receives visible authority.
How the lab sets up the comparison
The team begins with a practical prompt rather than an abstract dispute. A user might ask what a company does, whether a clinic offers a treatment, which manufacturer is known for a process, or how two regional businesses compare. The lab records the answer and the citation. Then it builds the visible source path around the disputed claim: first-party page, directory, press mention, institutional page, aggregator, bilingual mirror or copied fragment.
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. It is useful for conflict because the copied versions often simplify the technical category. The company page might describe a process as suitable for a narrow material condition. A sector listing may reduce that into a broad capability. A regional economic profile may present it as evidence of local expertise. The facts are related. They are not interchangeable.
Object B is a composite scenario: a bilingual professional clinic in Lyon with French treatment pages, an English patient-facing mirror, directory listings and occasional regional press mentions. It produces a different kind of conflict. The French page may distinguish consultation, diagnosis and treatment. The English mirror may compress these into a patient-friendly service label. A directory may list the service as a category. A regional article may mention the clinic in passing and date the expansion loosely. In an AI answer, those layers can become one confident sentence.
The lab does not assume the first-party page always wins. That would be too neat. A first-party page can be outdated, vague or promotional. A directory can sometimes preserve an older factual trail that the business later softened. A press mention can contain a date that no business page displays. The question is not which source deserves moral loyalty. The question is which public source visibly carries the claim the answer made, and which source the answer credited.
This is where the records become slow. The team separates the claim into parts: the entity, the category, the date, the service, the location, the comparison, the supporting detail. A citation may support one part and fail another. The answer may cite a directory for a business category while silently using the company page for the technical explanation. Calling the citation simply “wrong” can flatten a more useful observation.
The four citation moves in a conflict
The canon’s anchor becomes especially useful under disagreement: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. In conflict analysis, these moves describe what happens when the answer has more than one plausible public witness.
Source named is the rare clean case. The answer chooses the version that a visible source carries, and the citation points to that same source. If a company page says a branch moved to a new district and the answer repeats that, citing the updated branch page, the attribution is aligned. The lab still checks whether other sources disagree, but the credit line has not wandered.
Source displaced appears when the answer seems to use the fuller or more precise source while citing a weaker one. In Object A, the manufacturer’s technical note may explain a process in detail, while a sector directory carries a shortened copy. If the answer uses the detail but cites the directory, the directory becomes the named authority for knowledge it did not really carry. The conflict may stay invisible to the reader because both sources point in the same general direction.
Source absorbed appears when one source’s wording or claim seems to shape the answer without receiving citation. In Object B, the French treatment page may contain a careful distinction between consultation and treatment. The answer may reproduce that distinction while citing an English mirror that lacks it. The lab marks this cautiously. It can point to distinctive wording or a unique claim, but it cannot prove hidden influence with certainty.
Source contradicted is the sharpest pattern. The cited source conflicts with the answer or with a stronger visible carrier of the same claim. A regional article may give an older date, while the answer states a newer date and cites that article anyway. A directory may list a service category the clinic no longer uses, while the answer cites it to support a current service description. Here the footnote does not merely under-credit. It misleads the reader about where the sentence stands.
This typology keeps the lab from treating conflict as a binary pass or fail. A citation can be partially useful, displaced, absorbed or contradicted around different parts of the same answer. The material has to say which claim is being inspected.
Which source tends to become the visible winner
In bounded observations, the source that wins the citation often has one of several practical advantages. It may be easier to name. It may match the prompt language. It may carry a broader category label. It may sit on an institutional or press surface that looks like a public explanation rather than a company claim. It may be the only page whose title directly resembles the answer sentence.
That does not mean the engine has judged the source more accurate. The lab is careful with that inference. A citation layer can reward legibility. A national or regional page with a neat summary can be easier to attach than a first-party page where the fact is embedded in a long technical note. A directory with a clean category field can look more citable than a company page that explains the category across three paragraphs. The winning citation may be a handle, not the origin.
Language adds another pressure. A French page may carry the precise fact, while an English mirror gives the answer system a safer-looking match for an English prompt. If the two versions conflict, even slightly, the citation can move toward the language of the query. In Object B, an English patient page may receive credit for a claim that the French treatment page states more carefully. The result is not necessarily a mistranslation. It is a source choice made under bilingual pressure.
Source type also matters. Press and institutional pages can pull citation toward themselves when the prompt asks for a broader context. A business page says what the firm does. A regional article says what the firm means in a local industry story. If the answer frames the business as part of regional expertise, the article may become the citation even when the factual detail started elsewhere. That is source displacement with a narrative coat on.
The lab sometimes sees the opposite. Exact company-name prompts can bring the citation back to the first-party page, especially when the question asks “according to the company” or asks for a service detail. This is why the method uses related prompts. The winning source in a conflict is not a permanent champion. It may win only under one query shape.
What this means for French business information
The public memory of a business can shift through small citation choices. A copied profile receives credit for a technical method. A regional article becomes the witness for a timeline it only summarized. An English mirror becomes the cited source for a fact that was more precise in French. None of these cases requires the AI answer to be absurd. The danger is quieter: the reader is shown the wrong supporting page and learns to trust that page as the fact’s home.
For French SMBs, marketers and researchers, this means conflict should be inspected before asking whether an answer is good or bad. The first question is narrower: which claim is being supported by which citation? A business description can be broadly correct while one date is unsupported. A comparison can be useful while the cited page only supports the competitor’s category. A treatment summary can be readable while flattening a distinction the French page made carefully.
The lab’s preferred record for a conflict includes the disputed claim in a short quote or paraphrase, the cited source, the stronger visible carrier if one exists, and the type of attribution move. This makes the observation reusable. It also prevents the team from turning every discrepancy into an accusation. Some conflicts are caused by outdated pages. Some by bilingual simplification. Some by copied listings. Some by the ordinary compression that happens when a long page becomes a short answer.
A good discrepancy note has a small grain of discomfort in it. It does not say, “the model is wrong” and move on. It says, for example: the answer’s service claim matches the clinic’s French page, the citation points to an English mirror that omits the limitation, and a directory uses a broader category that may have encouraged the answer’s label. That sentence is less dramatic. It is also more useful.
Limits of deciding which source won
The lab can observe visible attribution. It cannot prove every influence path. Several sources may contain overlapping language because one copied another, because both used a common sector phrase, or because the business itself reused wording across pages. A cited page may contain part of the claim and miss another part. French and English pages may differ because they address different audiences, not because one is wrong.
Conflict records also age. A business may update a page after an observation. A directory may change its listing. A press page may remain static while the underlying company fact moves. The lab therefore avoids treating old screenshots as permanent verdicts. Each material should preserve the date of observation in the underlying record, even when the public article explains the pattern without turning it into a time-bound claim.
The method does not rank source types by truthfulness. First-party pages can be wrong. Directories can be useful. Press mentions can preserve context. Institutional summaries can give a broader frame. The lab asks a more inspectable question: when these sources disagree, which one did the AI answer name, and did that citation actually support the claim being made?
That modest question is enough. In citation behavior, the winning source is not always the best source. It is the source the answer made visible.