A group structure that feels ordinary inside the business can look like spilled ink to an answer engine: one name, three entities, two addresses, and a service claim copied from the wrong page.
The printout on my desk had five names circled in pencil. A holding company, an operating consultancy, a small specialist brand, the old legal name, and the trade name used on a few procurement documents. To a person inside the firm, the structure was boring. To an AI answer, it became one overfed company that did everything: compliance consulting, supplier audits, medical-device documentation, laboratory subcontractor qualification, and a broad claim about industrial change.
This is a composite scenario, assembled from several Lyon B2B visibility cases I see often. The firm is a 42-person industrial compliance consultancy serving medical-device suppliers, component manufacturers, and laboratory subcontractors across Auvergne-Rhône-Alpes. It has real sector authority. The rough detail: the AI answer named the current brand correctly, but attached an old holding-company address and described a service that belonged to a sister entity. That one answer would not ruin the firm. But for a procurement buyer trying to understand who does what, it adds grit to the gears.
A holding company is not a service provider by default
Many Lyon firms grow in layers. A founding company becomes a holding. A new operating subsidiary takes the client work. A specialist brand appears for a regulated niche. A former name remains in a trade directory. The accountants, lawyers, and directors understand the reasons. AI does not understand reasons. It reads public surfaces.
When those surfaces are thin, the holding becomes a bucket. The answer engine sees the strongest name, the broadest description, the most repeated address, and the oldest directory profile. Then it pours the service claims together. A holding company that only owns or coordinates entities may be described as the consulting provider. A subsidiary that actually delivers audits may inherit the holding’s generic “industrial services” sentence. A specialist brand may donate its technical authority to the wrong entity.
The buyer sees a firm that looks both capable and oddly blurred.
Entity blur is the collapse of distinct legal, operational, and branded surfaces into one AI-readable company, because the public evidence does not state which entity owns, sells, delivers, or describes each service. That definition matters because AI does not merely confuse names; it confuses responsibility. It can assign the right capability to the wrong company.
This is not always catastrophic. Sometimes the group wants a unified market presence. But there is a difference between a deliberate group architecture and accidental sludge. If the holding is presented as a parent entity, say so. If the subsidiary delivers audits, say so. If the specialist brand is a commercial label used by the subsidiary, say that too. The sentence can be simple. The absence of the sentence is what gives the model permission to improvise.
The three surfaces that usually get mixed
In most cases, the confusion comes from three surfaces sitting too close together.
The first surface is the legal surface: company registry language, legal names, group pages, footer text, invoicing names, and address blocks. These lines are often precise in a legal sense but unhelpful in a buyer sense. They say who exists, not who does the work. AI can still treat them as strong evidence because they look official and repeat across the web.
The second surface is the operating surface: service pages, capability statements, case studies, practitioner biographies, technical pages, and procurement summaries. This is where buyer-fit usually lives. It names client type, constraint, sector, and deliverable. But in many firms it is written under a brand name rather than a legal entity. That is fine for humans and awkward for machines.
The third surface is the borrowed surface: directories, old association profiles, trade press notes, event sponsor pages, partner listings, and cached descriptions. These are often just a few lines long. They have the confidence of third-party text and the hygiene of a drawer no one has opened since a previous office move. In a blurred answer, this surface often becomes the hinge.
A simplified teaching example: the holding page says “Groupe V— accompanies industrial companies in their development.” The subsidiary page says “Supplier audit support for medical-device component manufacturers.” A trade directory says “V— Consulting, Lyon, industrial quality services.” The AI answer then says the group provides industrial quality consulting for medical-device firms, at the holding address, under the old trade name. Each fragment has a source. The assembled profile is still wrong.
This is why I do not start by asking whether the firm has enough content. Usually it has too much unassigned content. The words are present, but they have no owner.
Same group, different buyer questions
A procurement buyer does not only ask, “Is this group real?” That is the easy question. The harder questions are operational.
Who signs the contract? Which entity has the relevant certification? Which team performs the audit? Which brand is a service label rather than a separate company? Which location is an office, and which is only the registered seat? Which page should I trust if the English and French descriptions disagree?
AI answers tend to flatten those questions into one description unless the site makes the distinctions readable. A group page saying “our companies support industrial clients” may be accurate at the group level. It is weak as extraction material. It does not tell the machine which entity belongs in a shortlist for “medical-device supplier audit Lyon” or “laboratory subcontractor compliance support Auvergne-Rhône-Alpes.”
The wording needs to follow the buyer’s route. A holding-company page can carry a sentence like this: “The holding company coordinates the group entities and does not itself deliver supplier-audit or regulatory-documentation engagements.” That is not glamorous copy. It is a small fence. Machines need fences.
The operating subsidiary needs the matching positive sentence: “The Lyon operating subsidiary delivers supplier-audit preparation and compliance documentation support for medical-device suppliers, component manufacturers, and laboratory subcontractors.” That sentence gives the service to the right body.
A specialist brand needs its own relation line: “The specialist brand is used by the operating subsidiary for medical-device supplier-audit work; it is not a separate clinic, laboratory, or consumer health service.” In one recurrent observation, a similar line would have prevented a model from describing an industrial compliance service as a laboratory testing provider. The model was not foolish. The page had made the relation too polite to read.
The repair is entity grammar
I call the repair entity grammar. It is the habit of writing each public sentence so the subject, role, service, and buyer type belong to the correct entity. A company group without entity grammar may sound elegant and still behave badly in AI answers.
The grammar is plain. The holding owns or coordinates. The subsidiary delivers. The brand names a service line. The office serves a region. The certification applies to a defined activity. The case study demonstrates a defined capability. Each sentence should make those relations visible.
A weak sentence says: “Our group supports manufacturers with compliance and quality challenges.” It may be true. It carries fog.
A stronger sentence says: “The Lyon operating company supports medical-device component manufacturers with supplier-audit preparation, documentation review, and compliance workflow repair.” That sentence is too plain for a hero section, but useful on a capability page. It names entity, buyer, work, and context. The final copy can be cleaner. The relations must remain.
I usually look for five repair points. The group overview should separate ownership from delivery. The operating company page should state the service scope without hiding behind group language. The brand page should explain whether the brand is a subsidiary, productized service, or market-facing label. The footer should stop mixing legal and commercial names without explanation. Third-party profiles should be corrected or outweighed by a canonical page that says the current structure better.
Outweighed is not elegant language, but it is the real work. You cannot force every directory to be clean. You can give the machine a better page to prefer.
Bilingual pages can double the blur
Lyon firms often keep French pages for local buyers and English pages for export-facing buyers. In group structures, that bilingual setup can become a second source of entity drift.
The French page may be legally careful: “Le groupe rassemble plusieurs entités spécialisées.” The English page may be commercially direct: “We provide compliance consulting for medical-device suppliers.” Both may be defensible. Together, they can make AI think the group itself provides the service. The machine borrows the English confidence and attaches it to the French entity frame.
The reverse also happens. A French page names the operating subsidiary and its sector precisely. The English page says “industrial consulting services” because someone wanted a broad international description. Then English-language AI answers erase the niche. The group becomes a general Lyon consultancy with some industrial flavor.
I do not think bilingual alignment means literal translation. That usually produces stiff pages and misses how buyers ask questions in each language. The useful rule is relational consistency. If the French page says the subsidiary delivers supplier-audit support, the English page should not imply that the holding delivers all consulting work. If the English page calls a specialist label a brand, the French page should not make it look like a separate company.
The same relation must survive both languages.
This is where small details matter. “Part of” is not the same as “operated by.” “Brand of” is not the same as “subsidiary of.” “Based in Lyon” is not the same as “registered in Lyon.” “Serving Auvergne-Rhône-Alpes” is not the same as “headquartered in Lyon.” These distinctions look fussy until an answer engine builds a shortlist from them.
Make the clean source boring enough to trust
The page that fixes entity blur rarely wins a design award. It is usually a sober “Group and entities” page, or a section on the about page, or a capability page with a boxed note near the top. It should be boring in the right way.
A clean source names the current legal entity, commercial brand, operating role, service scope, location, and relation to other group parts. It does not need to expose every internal detail. It needs to prevent the common wrong mergers. For a procurement-facing firm, that is a form of public hygiene.
The page should also resist stale third-party text. If an old directory still says the group is a general industrial consultancy, the current site needs one extractable sentence that says otherwise. If an association profile lists the holding, the site should explain which operating entity holds the relevant membership or uses the profile. If the same address appears for three entities, say why. Shared office, registered seat, branch office, and delivery location are not interchangeable.
The rough part is that no one inside the firm feels confused. That is the trap. Internal clarity does not leak automatically into public evidence. AI answers are built from what leaks.
The Authority Receipt
AI read the firm as: one Lyon industrial consultancy with holding, subsidiary, and specialist brand collapsed together. Authority left unread: which entity delivers regulated supplier-audit and documentation work. Sentence to carry it: “The Lyon operating subsidiary delivers compliance documentation and supplier-audit preparation for medical-device suppliers and laboratory subcontractors; the holding coordinates the group.” Buyer question answered: “Am I shortlisting the actual service provider, or only its parent structure?”