Sourceplane Atelier is a four-person research lab studying how AI systems make French businesses visible, vague or absent. Its work sits close to the source record: directories, registries, press mentions, regional naming and the language of the query itself. The lab is built for French SMBs, agencies and trade bodies that need to understand why an answer names a business through one evidence layer while missing another.
The lab began with an irritatingly small discrepancy. The same French business appeared confidently in one AI answer, vanished in another, and then reappeared through an English summary instead of the French source record that actually held the useful detail. It was not a grand theory at first. It was the sort of mismatch that makes a researcher reopen the tabs, rerun the prompt, and ask why the engine trusted that layer rather than the one sitting closer to the business.
Sourceplane Atelier formed around that question. Maël Dubreuil had been mapping commercial citation paths in comparison-style research. Camille Arven brought the bilingual reading habit needed to catch when an English prompt changes the identity of a French business. Noémie Vasseur had worked with administrative datasets and knew how legal names, identifiers and establishments can clarify a case without explaining reputation. Lucien Faroux came from regional market notes and map-based content reviews, where two businesses with similar names can sit in entirely different local realities.
What sets the group apart is its patience with the source layer. They are less interested in whether an answer sounds polished than in whether it can be justified through usable evidence. A directory can lead the answer. A registry record can anchor it. A local article can amplify a framing. A regional detail can disappear, leaving a business described as merely French when the city, department or branch mattered. The lab's position is simple and a little stubborn: AI visibility is a question of which evidence the system can retrieve, connect and treat as safe enough to repeat.
maps citation sources
How AI systems select between directories, registry-style records, press mentions and company-owned pages.
He previously worked on editorial research for commercial comparison projects and local service guides. His work at the lab follows the source path behind a business answer.
tests French retrieval
How French-language queries and English-language queries surface different sources for the same business.
She previously edited bilingual business explainers and prepared structured content for service-sector websites. She reads for the small language shifts that change what an engine retrieves.
reviews registry evidence
How official identifiers, legal names and establishment records are used, ignored or loosely paraphrased in AI answers.
She previously organised administrative datasets for small-business support materials without publishing under a public institutional profile. Her work keeps identity evidence separate from reputation claims.
studies regional representation
How engines distinguish local branches, regional independents, national chains and same-name businesses across French regions.
He previously worked on regional market notes, local-category audits and map-based content reviews. He looks for the local detail that disappears when an answer becomes too generic.
Sourceplane Atelier studies AI visibility as a source-dependency problem. The lab compares answers across engines, languages and location frames, then classifies how those answers depend on directories, registry records, press mentions and regionally specific evidence. The work starts from the concrete artefact: the prompt, the language used, the location frame, the visible citation if one exists, and the answer's wording.
The lab draws a hard line between an observation and a conclusion. One answer is only material. Several related observations are compared across engines, languages or regional variants before the team says that a pattern is likely present. If an engine cites a source, the dependency is recorded directly; if an answer seems to follow the shape of a known directory or registry without naming it, the lab marks that as probable dependency rather than proof.
Repeatability has a practical meaning here: another reviewer should be able to reconstruct the query setup, the business category, the language, the location frame and the comparison logic. Identical wording from every model run is not required — generative systems move. The lab also names its limits and keeps forecasts in a separate box, labelled as uncertainty notes.
A single answer is treated as material, not as a finding. The lab waits for comparison before naming a dependency pattern.
Named citations, probable source traces and source absences are handled separately. That keeps the evidence from becoming smoother than it really is.
French and English prompts are compared because they can retrieve different records for the same business. The difference is part of the result.
A reviewer should be able to rebuild the prompt setup, location frame and comparison logic. Identical model wording is not required.
The lab can describe source conditions that may change visibility. It does not sell prediction as certainty.
For questions about French business visibility in AI systems, Sourceplane Atelier begins with the source trail.
Get in touch