How citation behavior becomes evidence

In the lab's founding source-mismatch case, a French manufacturer publishes a technical note on its own site. A regional directory copies one shortened fact from that note. An AI answer later repeats the technical detail, yet the visible citation points to the directory. The copied page has the easier public label, but it does not carry the full explanation. The claim may be accurate. The credit may still be wrong.

The lab calls an observation a recorded prompt, answer, cited source, visible source path and attribution behavior. The record keeps the wording of the query close to the surface, because a small change from a company-name prompt to a category prompt can change which source gets named. A conclusion requires more than one memorable case. The team looks for the same citation move returning across related prompts, across engines, or in time-separated runs: source named, source displaced, source absorbed, source contradicted.

Samples are built around practical French business and topic questions. They include company names, regional modifiers, sector categories, comparison prompts, bilingual variants and questions where a reader would reasonably expect a source to be credited. The point is bounded inspection rather than a measurement of the whole French web. The team creates answer situations where attribution can be examined while the outside machinery remains partly hidden.

The lab separates citation from retrieval and synthesis. A cited page may be only the easiest surface to name. Another page may have carried the original fact, the fuller method, the older wording or the missing context. Sometimes the answer seems to absorb a phrase from a first-party page while citing a directory that copied it. Sometimes an institutional summary receives credit for a claim that began on a business page. Sometimes a French page and an English mirror support slightly different versions of the same company fact, and the answer quietly chooses one.

Repeatability in this work does not mean identical sentences. AI answers change their phrasing too easily for that to be a useful standard. Instead, the lab watches for stable source choices and stable attribution shifts: the same directory favored over the same company page, the same English page cited for a French topic, the same regional article used to support a broader business description. A pattern that survives variation earns more attention than a single neat screenshot.

The limits stay in the record. Interfaces change. Browsing access changes. Citation rules change. One engine may see a source another engine cannot reach, and some influence paths remain unprovable from the outside. Forecasts are therefore written as conditional notes: if a pattern persists, it may imply that a source has become the model's preferred public handle for a claim. Uncertainty is marked when several sources could explain the answer, when the citation contains only part of the claim, or when French and English pages point to different versions of the same fact.

The lab works from a small shared vocabulary. An observation is a recorded prompt, answer, cited source, visible source path and attribution behavior around one specific claim — its basic unit of evidence. The cited source is the page named by the answer as support; it may be the original carrier, a partial carrier, a copied surface or only the easiest public page to name. The original carrier is the source that appears to have published the claim, method, fact or context in its fuller or earlier form, marked cautiously when the public trail supports that reading. A source path is the visible trail connecting first-party pages, directories, press mentions, institutional summaries, aggregators, bilingual mirrors and copied fragments around the same claim. An attribution shift is a change in credit from the source that appears to carry the claim to another source named or implied by the answer. Uncited absorption is a case where the answer seems to use material from a source without naming it in the citation layer.

The classification anchor is one typology: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. It is a qualitative description of attribution behavior, not a metric, score or scale. Source named means the answer credits the page that visibly carries the claim. Source displaced means credit moves to a weaker or copied surface. Source absorbed means a source appears to influence the answer without being cited. Source contradicted means the cited source conflicts with the answer or with a stronger visible carrier.

Working principles

1. Observation before conclusion

   A single answer can be useful evidence, but it is not treated as a settled pattern. The lab waits for recurrence across related prompts, engines or separated runs.

2. Citation is its own layer

   The lab reviews attribution separately from retrieval and answer wording. A source can influence an answer without being named, and a named source can be a weak carrier of the claim.

3. Source types are marked

   First-party pages, directories, press mentions, institutional sources, aggregators, bilingual mirrors and uncited assertions are recorded qualitatively. The label helps explain how credit moved.

4. Uncertainty stays visible

   When several paths could explain the same sentence, the material says so. The lab does not fill gaps with certainty for the sake of a cleaner story.

5. Forecasts remain conditional

   Forward-looking notes describe what a recurring pattern may imply if it persists. They are kept separate from what the current observation directly shows.