Before an executive ever mentions you to ChatGPT or Claude, these models have probably already assembled a rough identity record on your company, built from sources none of them asked your permission to consult: the commercial register, open knowledge bases, professional directories, and your own site. This record exists, whether you like it or not. Checking it takes an hour; correcting it goes through the sources themselves, not through a conversation with the model.
What exactly is a machine identity record?
I call a "machine identity record" the set of facts a system can retrieve and cross-reference about your company from sources independent of one another: exact legal name, legal form, headquarters, business identification number, founder or executives, sector of activity, founding date. It isn't a profile you deliberately create somewhere; it's a reconstruction, made from scattered fragments, that the model assembles the moment it's asked. This is precisely what I described, in How an AI chooses its sources, as entity consistency: before deciding to cite you, a system must first be sure it knows who you are.
Where does this record come from? The four sources models cross-reference
Four families of sources come up repeatedly, with differing weight depending on the question asked.
Wikidata. An open, structured knowledge base, editable by anyone according to precise notability rules. Generative systems readily use it as an anchor point for entity resolution, because its data is structured and easy to work with. An SME doesn't automatically appear there; many don't appear at all.
Zefix. Switzerland's central register of commercial enterprises, maintained by the Confederation. It's the authoritative source for the exact legal name, legal form, headquarters and registration status — a model that cross-checks your claims against Zefix is, in effect, verifying that you officially exist as you claim to.
Professional directories. local.ch, search.ch and their equivalents index address, phone number, business category, sometimes reviews. They are widely crawled by automated systems, often the first source consulted for a local query — and often also the least kept up to date by the company itself, once the initial listing has been forgotten.
The company's own site. This is the only one of the four sources you fully control. schema.org structured data explicitly states who you are; an llms.txt file, placed at the root of the site, can additionally give models a direct summary, readable without interpretation.
I describe four sources here without ranking them into a definitive hierarchy: their relative weight varies depending on the question asked, and it's their cross-referencing, more than any single one of them, that builds the record.
Why gaps between these sources are costly
A model faced with contradicting sources does not rule in your favor by default: it keeps the most stable version, or it becomes cautious, sometimes to the point of not citing you at all rather than risking a false claim. An old address still listed as active on a directory, a founder's name that no longer matches the current leadership team, a legal name that varies slightly from one page to another on your own site: each of these gaps, taken alone, seems minor. Added together, they form a blurry identity record — and this is exactly the type of inconsistency I detailed, a source of quiet penalty rather than visible error, in Structured data: what AI actually reads on your site.
How do you check your identity record in an hour?
The method requires no specialized tool, just an hour and a few open browser tabs.
- Ask each of the five models — ChatGPT, Claude, Perplexity, Gemini, Mistral — a neutral question: what does it know about your company, in your city.
- For each answer, note the address, the stated activity, the name of the founder or executive, and the founding date if mentioned.
- Compare these elements against the commercial register on Zefix, which is the reference for legal name, legal form and status.
- Check professional directories such as local.ch or search.ch: are the address and phone number still accurate there?
- Note the discrepancies across the five answers rather than looking at just one: it's the gap between models that signals a poorly anchored identity — cross-model consistency is, incidentally, one of the seven themes measured by the Score GEO™.
The outcome of this exercise is neither a score nor a judgment: it's a list of factual gaps, the same raw material I work with in the first stage of a Score GEO™ pre-audit.
How do you correct this record? mcva.ch as a documented example
Correction isn't negotiated with the model: you don't convince ChatGPT it's wrong about your address, you correct the source it consulted. Two levers are entirely within your control. The first is your own site's schema.org markup: on mcva.ch, the Organization schema states the legal name, address, founder and business identification number — CHE-448.625.502 — and links it, via the sameAs property, to the Zefix register listing and to Moneyhouse, two sources a system can consult to verify that the entity speaking on the site is indeed the one officially registered. The second is the llms.txt file, placed at the root of mcva.ch: a direct, structured-text summary of who I am, what the company does, and where to find the pages that matter, without the model having to infer it from the rest of the site. Both levers are part of the work I systematically put in place ahead of any site designed to be cited.
Two other levers are partly out of your hands, and honesty demands saying so. Correcting a directory listing follows its publisher's update cycle, not yours. Obtaining or correcting a Wikidata entry requires respecting the notability and editing rules of the community that maintains it. These are slower efforts, to be pursued alongside the work on your own site rather than in place of it.
This identity record is only the foundation: it makes citation possible, it doesn't trigger it. The underlying work remains that of a site designed to be cited, and the mechanism that then decides what gets kept in an answer is the one I detail in How an AI chooses its sources. To situate this work within the full set of decisions a Swiss SME must make in the face of AI, the complete roadmap remains laid out in What should a Swiss SME do about AI in 2026?
Key takeaways
— Models reconstruct an identity record of your company from four independent sources: Wikidata, Zefix, professional directories and your own site. — Gaps between these sources are costly: a model facing contradictory versions becomes cautious, sometimes to the point of not citing you at all. — Checking takes an hour — ask the five models and compare; correcting goes through the sources you control, schema.org and llms.txt first.
FAQ
Do I necessarily need a Wikidata entry to be cited by AI? No — nothing is mandatory. An entry adds an extra anchor point for entity resolution, useful but not essential, and obtaining one follows the notability rules of the community that maintains the database, not yours.
Does simply changing my address on my site fix what AI knows? It corrects the fastest source to update, your own. Directories and external databases follow their own cycle, often slower. Cross-referencing between sources means a correction takes time to spread everywhere.
How do I know if my company is being confused with another one of the same name? That's exactly what the five-question exercise reveals: if the answers describe an activity, city or founder that aren't yours, the confusion is there. A generic name or a legal name close to a competitor's makes it worse.
Does this identity record replace a full citability measurement? No. It's the foundation, not the measurement itself. The Score GEO™ goes further: it tests whether, once identity is resolved, your company is actually cited, in what context and with what accuracy.
Where do you stand? The Score GEO™ pre-audit measures, free of charge, where your company appears — and doesn't appear — in the answers from ChatGPT, Claude, Perplexity, Gemini and Mistral. A quantified answer within a few days. Request a free pre-audit
Jérôme Deshaie is CEO and founder of MCVA Consulting SA, an AI-augmented agency based in Valais. Fifteen years serving major international brands, now working directly with Swiss SMEs. Background.