When an AI Confuses Your Company With a Namesake

Some time in the summer of 2026, a query turned up in the Search Console account for psentry.com, in Italian: chatgpt confonde la mia azienda con un'altra. ChatGPT confuses my company with another one. Somebody had asked an assistant about their own firm and been handed, with perfect composure, a description of a different business. The page Google sent them to was one of ours about wrong answers in general, which was near the subject and not on it. This is the piece that should have been there.

The question underneath is mechanical. A name is a string of letters, and more than one organisation can own the same string. Something has to decide which one a question is about. When that decision goes the wrong way, the answer is not vague. It is a confident account of someone else.

A name is not an identifier

A company registry deals in identifiers: a registration number, a VAT number, an address. It never needs the name to tell two firms apart.

A language model has no such column. What it learned during training is which words tend to appear near which other words. If a name appears in a great deal of text alongside a particular sector, city and set of products, those associations are strong. If the same name also appears, less often, alongside a different sector and a different city, those associations exist too, but they are weaker. Nothing in the model marks them as two separate companies. There is a name, and a spread of things that have been said near it.

When a question arrives, the words around the name in the question pull the answer towards one set of associations or the other. A question that says the valve manufacturer in Brescia called X gives the model a good deal to work with. A question that says only what does X do? gives it almost nothing, and the answer falls back on whichever set of associations is strongest. That is the referent that has been written about most.

Why the smaller namesake loses

It follows that the loser in a name collision is rarely the company with the weaker claim to the name. It is the company with less text around it. Size, age, press coverage and a language with more pages on the web all help. Who registered the name first does not.

Three kinds of name are especially exposed, and industrial exporters own a great many of them.

• Surnames. A great many European manufacturers carry the founder's family name. The surname also belongs to other firms, in other sectors, sometimes in the same province, and to people, some of them considerably more famous than the company.
• Ordinary words. A name that means something in Italian or German, a word for strength, or for a mountain, or for a tool, competes with the word itself and with every other business that found it appealing.
• Acronyms. Three letters that stand for a family business in Emilia may stand for a listed company in another country and another sector. One has annual reports and a press office; the other has a website and a trade fair stand.

For the sake of an example, and it is only an invented one, picture a maker of hydraulic fittings near Bergamo trading under a three-letter acronym, and a much larger software firm abroad with the same three letters. Asked in English what the company does, a model has far more material on the software firm. The fittings maker does not appear incorrectly. It does not appear at all, because the question was resolved to someone else before a word of the answer was written.

How retrieval stitches two companies into one

An assistant that searches the web before it answers has a different failure, and in some ways a worse one.

The search step takes the name and brings back a handful of pages. If two companies share the name, the handful can contain pages about both. The model then writes one answer from those extracts. Each extract is a fragment: a paragraph from an about page, a line from a directory, a sentence from a news item. The fragment often does not repeat the company's city or sector, because the page it came from assumed the reader already knew. So the model is holding several true statements, about two different companies, with nothing in the text to say that they belong to different subjects.

The result is a merger that never took place. Your address, their founding year. Your product range, their headcount. Your city, their acquisition. Every individual fact can be traced to a real page, which is what makes the composite so hard to spot and so plausible to a reader who does not already know the truth. It is not a hallucination in the sense of something invented. It is a filing error.

There is a language dimension to this, and exporters should expect it. Ask in Italian, and the Italian pages about the Italian company may dominate what the search step returns. Ask the same question in English or German, and the foreign namesake, with more pages in those languages, may take over. The same name can resolve correctly at home and wrongly in every market you export to.

The signals that separate one entity from another

None of what follows guarantees how any assistant resolves your name. What it does is put disambiguating material on the web, in places where a reader, human or machine, can find it attached to the name.

Say the same three things everywhere. Legal name, city, sector, in the same form, on your site, in directories, on trade association listings and in press releases. The aim is that a fragment lifted from any one of those pages carries enough context to identify you on its own. A sentence that reads X, the Bergamo manufacturer of hydraulic fittings survives being cut out of its page. A sentence that reads we have been doing this since 1974 does not.

Print the identifiers. Many European companies already put the VAT number in the site footer. Put it on the about page as well, with the registered address. A registration number is the one thing that cannot be shared with a namesake, and it costs nothing to publish.

Say what you are not. An about page can state plainly that the company is not connected with the similarly named firm in another sector or country. It feels awkward to write. It is also the only sentence on the web that puts both names on one page with the relationship spelled out, rather than leaving a model to infer it from co-occurrence.

Mark it up. Google's documentation for Organization structured data lists properties such as legalName, vatID, leiCode, iso6523Code and address, and says some of them work behind the scenes to disambiguate organisations. The sameAs property is defined by schema.org as the URL of a reference page that unambiguously indicates the item's identity, for instance a Wikidata entry or an official website. Point it at your registry listing and your profiles elsewhere. Be clear about what this does: it is documented for Google Search. Nothing published says how, or whether, a given assistant reads it when composing an answer.

Use the registers, where you legitimately belong in them. Wikidata's notability policy admits an item that refers to a clearly identifiable entity which can be described using serious and publicly available references. For a company with a public registry entry and some independent coverage, that bar is not high. An item with your registration details and official website is a structured statement that this name, in this place, is this company. Write it accurately, and do not invent references to get over the line.

What measurement shows, and what it does not

The quickest check is to ask an assistant what your company is, with the bare name, and then again with the city and sector attached. If the two answers describe different businesses, the name is not resolving on its own. Do it in each language you sell in, because the result may differ in each.

PSentry runs buying questions across ChatGPT, Claude, Gemini and Perplexity in every language and market you monitor, and shows which other companies the assistants name and which sources they used. When a namesake is being mistaken for you, those two lists are usually where it shows: an unfamiliar company among the names, an unfamiliar website among the sources. It has no dedicated feature that detects confusion between entities, and it will not flag the answer for you. Someone who knows which company is yours still has to read it.