The mechanism
Why does AI get my company wrong?
An agent asked about your company does not consult a database of companies. It searches, opens a few pages, and writes a conclusion from whatever came back in the seconds it had. When the answer is wrong, the cause is almost never invention. It is substitution: your own account of yourself was unreachable, undated, uncorroborated, or attached to the wrong name, so something else filled the space. There are four mechanisms, they are mechanical rather than editorial, and each one is checkable on your own site this afternoon.
A model with nothing to read does not stop. It answers from the nearest available source, and that source is rarely you.
It cannot read you, nobody else vouches for you, nothing you say carries a date, or your name resolves to another company.
Each one is a fetch, a lookup, or a search anybody can run. None of it requires guessing at what a model was thinking.
It is substitution, not invention
The word people reach for is hallucination, and it sends them looking for the wrong fix. A model asked about a company it cannot find does not return nothing. It returns the most plausible account it can assemble from what it did find, and it states that account with the same confidence it would state a fact. So the useful question is never "why did it make that up". It is "what did it read". That question has a literal answer, because the searches an agent ran and the pages it opened can be recorded, and the answer is usually a page you did not write.
Mechanism one: an agent cannot read your site
This is the one that surprises people, because the site looks perfect in a browser. An answer-time agent fetches the URL and does not run your JavaScript, so a page that assembles itself client-side returns navigation, a headline, and nothing to quote. Across the sites Oomira has probed under the current checks, 14 of 96 returned a shell rather than prose to an anonymous fetch. A second, quieter version: robots.txt written years ago for search crawlers, which names none of the answer-time agents, or names them in a block. 6 of 96 sites served an ordinary browser normally and turned a named agent away. Check it by fetching your own homepage with curl and reading what comes back.
Mechanism two: nobody except you says it
An agent weighing a claim looks for a second voice. When every figure about a company traces back to that company’s own site, a careful model does what a careful buyer does: it hedges. The hedge is the sentence that costs the meeting, and nothing in it has to be false for that to happen. This is the mechanism behind the most frustrating category of wrong answer, where everything stated is accurate and the reader still comes away unconvinced. The fix is not louder copy on your own domain. It is evidence that lives somewhere an agent can read and you do not own.
Mechanism three: nothing you publish carries a date
A claim published once and never stamped reads as permanently current to a generous model and as unverifiable to a careful one, and neither reading is what you wanted. Worse, it means an agent has no way to prefer your newer statement over an older third-party one, because as far as the page is concerned neither has a time. Most companies discover this as an answer describing a product they stopped selling, or a headcount from two funding rounds ago, delivered with total confidence. The claim was true when written. Nothing on the page said when that was.
Mechanism four: your name resolves to somebody else
This is the failure that makes the other three irrelevant, because everything downstream is measured against the wrong company. Ordinary-word company names, names a character or two from a larger brand, and names shared across jurisdictions all resolve badly, and a model reaching for a canonical entry finds the wrong one or none. When it finds none, a scam-check page or a review listing for a similar domain is sitting right there. Structured data does not fix it on its own, because a search-based agent may never parse your markup. Prose that names the near-misses explicitly does, because a model holding a page about somebody else needs a stated negation to discount it.
The half you can check yourself this afternoon
Three of the four mechanisms are mechanical, and mechanical things can be tested without anybody’s opinion. Fetch your own homepage the way an agent does, with curl and no browser, and read what comes back: if the argument for your company is not in that response, an agent never saw it. Open your robots.txt and look for the answer-time agent names rather than the search crawler names, because they are different lists and a file written for search mentions none of the second. Search a public knowledge graph for your company name and see which entity comes back. Each of those is a minute’s work and each one either clears a mechanism or names it.
The half you cannot, and why
The fourth check is what four agents say about you to somebody who has never heard of you, and you cannot run it in your own chat window. Your agent has your history, your earlier questions, and frequently your own website open in the conversation, so it answers you warmly and a stranger gets something else entirely. That gap is the measurement. An Oomira report asks the four cold (blank slate, no memory of you, nothing supplied by us, one live search tool each), keeps every search and every page opened as a receipt, and grades what came back against a record where each fact carries its source and the date it became true. It runs the mechanical checks above as 45 named probes and prints the arithmetic beside the score. Across the 91 companies whose sites have been probed under the current checks, measured 2026-08-16, the median reachability score was 50 out of 100: the median company asking this question is about half-open to the agents answering for it.
Questions this helps answer
- Why does ChatGPT describe my company incorrectly?
- Is AI hallucinating about my business, or reading something wrong?
- Why does an AI answer about us cite a competitor or a directory instead of our site?
- Can I change what AI says about my company?
- How do I check what an agent can actually read on my website?
Net Promoter®, NPS®, and Net Promoter Score® are registered trademarks of Bain & Company, Inc., NICE Systems, Inc., and Fred Reichheld. Agent NPS applies the Net Promoter method to AI agents and is not affiliated with or endorsed by Bain & Company.