Why ChatGPT gets your company name wrong, and how to fix it
AI assistants describe companies from what the web says about them, and the web rarely agrees with itself. The five usual causes of a mangled name or a wrong description, and the fixes, from cheapest to slowest.
Mauro Miotti4 min read
An AI assistant gets your company name or description wrong when the sources it can find disagree: an old name on a directory, a misspelling in a press release, a similarly named firm in another canton, a homepage that never states plainly what you do. The fix is to make the correct version the easiest one to find, in the places machines read first: your own pages, your structured data, the registries and profiles that confirm you, and a brand.txt that spells out the rules.
It feels personal when it happens. A prospect asks an assistant about you and is told you are a software company when you are a consultancy, or that you are based in Basel when you moved to Zug in 2019, or that "Müller & Partner" is "Mueller Partners AG". But there is no malice in it, and usually no bug either. Language models write from the text they retrieved, and they weigh agreement across sources. If the web says three different things about you, you get a blend.
The five usual causes
1. You have had more than one name. A rebrand, a merger, a legal-form change from GmbH to AG. The old name lives on in directories, old press, PDF brochures, and someone else's "our partners" page. To a retrieval system those are two entities with overlapping evidence.
2. Your own pages never say it plainly. A homepage headline that reads "Shaping tomorrow, together" tells a machine nothing. If the one sentence describing what you do, for whom, and where exists only in a PDF or an image, the assistant will write its own.
3. A namesake. A company with a similar name in another country or industry, with more web presence than you. The assistant is not confused about you; it has merged you with them.
4. Inconsistent details across listings. Address, phone, founding year, headcount, the umlaut in the name: when Google Business, LinkedIn, the commercial register and your footer each say something slightly different, the model averages, and the average is wrong.
5. Nothing authoritative confirms you. No Wikidata entry, no register link, no profile on the sites assistants treat as reliable. Your own claims are one source; without corroboration they carry little weight against a wrong but well-linked third-party page.
The fixes, cheapest first
State it on the homepage, in text. One sentence in the first paragraph: who you are, what you do, for whom, where. Not a tagline; the sentence you would want quoted. This is the single highest-value change and it takes ten minutes.
Add Organization structured data. JSON-LD on the homepage with name, legalName, alternateName for the forms you accept, url, address, foundingDate, and sameAs links to every official profile: commercial register, LinkedIn, the industry association, Wikidata if you have one. This is the machine-readable version of your imprint, and it is what disambiguates you from the namesake.
Publish a brand.txt. A small text file at the root of your site listing your official name, the spellings and former names that are not correct, and the rules (capitalisation, the legal suffix, the product versus the company). It is a new convention and not every system reads it, but it is the one place you can state the rules explicitly, and it costs nothing. The free analysis on this site tells you whether yours exists and parses.
Make the listings agree. Go through the ten places your company appears most, Google Business, LinkedIn, the register, Apple Maps, the two or three directories your industry uses, and make name, address and description identical. Boring, decisive, an afternoon.
Get the old name retired. Ask the directories and partners still using it to update. Where you cannot, add a line to your own site: "formerly X", which gives retrieval systems the bridge they are missing.
Earn one authoritative mention. A Wikidata item with your register number, a trade-association member page, a regional business journal profile. The AirOps study of AI citations found third-party pages 6.5 times more likely to be the source of a brand mention than the brand's own site; the same mechanism decides whose version of your name wins.
Then check
Ask the assistants the direct question, "what is [your company]?", in each language you operate in, and read the sources they used. The wrong ones tell you where to go next. Repeat in a month; entity descriptions move slowly, and a correction you made in November shows up around February.
For the broader picture of how assistants build an answer from retrieved pages, see How AI search works; for the trust signals they weigh, E-E-A-T for AI search.
