A small team watching business facts move
Levertrace Lab is a four-person research group focused on how large language models represent French businesses after public source changes. Their work is slow by design: source tracing, prompt repetition, language-pair checks and discrepancy logs before any claim is made. The page explains who keeps each test narrow, who follows third-party sources, and who watches wording drift across systems and languages.
i. Where the work began
A recurring composite case gave Levertrace Lab its shape: a small French service company corrected its website, fixed a directory entry and added clearer category wording. The public evidence looked tidier afterward. The model answers did not. One repeated the old activity, one picked up the new wording, and one avoided the company so completely that the absence became part of the case. Levertrace Lab formed around that kind of practical contradiction.
The group came together from adjacent work: technical content audits, structured business descriptions, bilingual explanations and qualitative comparison notes. What joined them was a shared irritation with vague AI visibility advice. A business owner can change a title tag, rewrite a French service page, clean up a directory profile and still have no clear sense of what mattered. Levertrace studies the link between the lever and the answer. The team looks at owned pages, third-party sources, structured signals, language versions, interlinking, dated updates and repeated claims as parts of the same evidence trail.
The four roles grew from that question. Maël Durance keeps each tested change narrow enough to interpret. Livia Caron maps which sources hold the model’s wording in place. Noé Vestrin tracks whether a French edit reaches the English answer. Salomé Rivecourt records where facts drift, stick or stay wrong across models and runs.
ii. Why this caution
Its position is cautious, and that caution is part of the product. The lab does not promise that one edit will bend a model into shape. It asks what changed, where it changed, how the phrasing moved, and whether the same movement appeared across languages or systems.
In a market like France, where business identity often lives across local directories, trade pages, regional mentions and bilingual summaries, that kind of close reading is the work itself. The four lever outcomes — adoption, partial echo, source conflict and no visible movement — are descriptions, not promises.
iii. What we do not do
The lab does not promise rankings, force citation, manage review profiles or sell instant AI visibility fixes. It avoids cases where there is too little public evidence to compare, and it does not rewrite business pages as a service. Its work is closer to close reading than to dashboard measurement. Focus: French LLM business descriptions. Method: before-and-after source comparison.
Masthead · four researchers
Runs before-and-after comparison of optimisation edits across owned and third-party sources. He previously worked on technical content audits, search-intent mapping and editorial change logs for service businesses. His role is to keep each tested change narrow enough to interpret.
Studies how directories, trade pages, review platforms and business profiles compete with a company’s own website in LLM answers. She previously edited structured business descriptions and maintained source inventories for local and regional organisations. She watches for the third-party page that quietly outranks the official story.
Checks whether French edits propagate into English answers and whether English summaries distort French facts. He previously worked on bilingual product explanations, terminology alignment and practical documentation rewrites. His work catches cases where a correct French source becomes a crooked English answer.
Follows model-to-model differences, repeated prompt behaviour and the persistence of wrong or outdated facts. She previously produced qualitative research notes, interview summaries and comparison tables for commercial research teams. She keeps the discrepancy logs readable enough to be useful after the first pass.
The lab studies the distance between public evidence and model wording; new cases appear in the index as the work advances.