Why AI gets company names wrong (and how to measure a name's AI visibility)

Ask an AI assistant "what is Lumio?" and it will tell you about a projector, a reading app or interactive-whiteboard software, depending on the day. Ask about an Italian startup named after a common English word and you get the dictionary definition. Ask about a brand spelled with a "y" instead of an "i" and the model silently rewrites it with the "i", citing a company that is not yours. These are not random glitches. They are predictable consequences of how language models work. And since a growing share of people now discover products and companies by asking an AI rather than a search engine, a name's ability to be recognised correctly, what we call AI visibility, has become a naming criterion in its own right. This guide explains why AI gets names wrong and how to measure the risk before you choose.

How a language model "knows" a name

An LLM has no company register inside it. It has learned, from vast amounts of text, which words tend to appear near which others. When you ask about a name, it does not look up a record: it reconstructs the most probable answer from the associations seen in training, possibly supplemented by a live web search if the product offers one.

Three direct implications follow:

1. Frequency wins. If a name appears a thousand times referring to one entity and ten times referring to yours, the model will talk about the other one.
2. Form matters, not just meaning. The model operates on tokens, fragments of text, and a name written in an unusual way may be split into pieces it has never seen together.
3. Memory is static. A brand born after the training cut-off simply does not exist for the model, unless a web search fills the gap; and then what counts is the search results, and therefore whoever dominates them.

The five typical causes of error

1. Training-data collision

This is the most common cause. Your name, or a near-identical string, is already associated with something else: a company in another country, a discontinued product, a fictional character, a technical term. The model does not know you are the one meant. The more thoroughly the other entity is documented online, the heavier the collision and the harder it will be to shift the association.

The most insidious collisions are the ones trademark searches do not surface: a name that is free in the EUIPO register may be the title of a 2009 video game with thousands of forum pages, and for the model that is the meaning.

2. Tokenization of invented words

Models segment text into tokens based on statistical frequency. Common words are a single token; rare or coined words are broken into fragments. "Zalando" is by now cited often enough to be stable; a new name like "Qvexia" may become "Q", "vex", "ia", and each fragment drags in its own associations ("vex" suggests annoyance, "ia" a Latin ending or an acronym). The result is an approximate reconstruction: the model may "correct" the name into a known word, add or drop letters, or fail to recognise it at all when it appears in a question.

A practical corollary: creative spellings (swapping "i" for "y", dropping vowels, doubling consonants) are the enemy of AI visibility. The model tends to normalise toward the most frequent form, which is the one you did not register.

3. Homophones and voice transcription

More and more interactions with AI happen by voice. A spoken name is first transcribed into text by a speech-recognition system, then interpreted by the model. If the name sounds like a common word or another brand, the transcription will pick the more probable form, and you vanish. "Kwik" becomes "quick", "Fyre" becomes "fire", and a name with sounds that sit ambiguously between "b" and "v" or "c" and "k" is spelled differently on every attempt.

4. Dominance of an existing entity

Even without an exact collision, a name that shares a root with a very famous brand is pulled into its orbit. Call yourself "Teslar" or "Amazonia Tech" and AI answers will tend to mention the giants, present you as their product or partner, or quietly correct the question. This dominance is also why search ownerability is part of the measurement: if the first page of Google for your name belongs to others, any AI with web access will repeat that picture.

5. Descriptive names that dissolve into their category

A name that describes the product ("Fast Deliveries", "Smart Energy Solutions") is, to the model, indistinguishable from the generic description of that kind of service. Ask "what is Smart Energy Solutions?" and you will get an explanation of what smart energy solutions are, not a profile of your company. The name does not carry enough entropy to be an entity.

How NameSapiens measures AI visibility

The "AI visibility" dimension of the Name Readiness Score is neither an opinion nor a number produced by a single prompt. It is the outcome of a battery of repeated probes, run across several models and in several languages where relevant, aggregated into a score. Here is the method, described so you can reproduce it by hand if you wish.

Collision test: "What is X?"

The model is asked the simplest possible question with no context: "What is [name]?", "Who is [name]?", "Tell me about [name]". Answers are recorded and classified: the model responds with a specific entity other than yours (strong collision), responds vaguely or admits it does not know (open ground), or gives a dictionary definition (name too descriptive or a common word). Repetition across models separates an occasional slip from an entrenched association.

Spelling-from-description test

The name is described to the model without being written, as a person who heard it spoken would do: "How do you spell the name of that company pronounced [phonetic rendering] that does [sector]?". The measure is how often the model produces the correct spelling. A name spelled right nearly every time has a stable form; one that gets normalised toward another word has a tokenization or homophone problem that will resurface in every voice interaction.

Disambiguation test

The model is given the name plus minimal context ("[name], Italian logistics software company") and checked on whether it can separate the brand from the entities it collides with, or keeps blending them. This estimates how recoverable a collision is: if one word of context is enough to correct course, the name is defensible with good communication; if the model insists on the other entity, the dominance is too strong.

Search ownerability

Web search results for the exact name and its phonetic variants are analysed: how many top positions are held by established entities, how many domains contain the name, how plausible it is that a new brand could take over that page within a reasonable time. Because many AI assistants integrate web search, this dimension predicts what the model will say tomorrow, not just today.

The four tests yield distinct signals that are weighted and combined. The score does not claim to predict the behaviour of every future model; it indicates how steep the hill is. A name with a strong collision, unstable spelling and search results owned by others has low AI visibility regardless of how good it sounds.

What to do with the result

If you are still choosing, AI visibility is one criterion to use alongside the other six in the Name Readiness Score: a name that is perfect for AI but unregistrable as a trademark is useless, and vice versa. The free Name Check lets you compare several candidates on the same basis.

If the name is already chosen, you can work at the margins:

• publish one unambiguous, consistent brand description, repeated verbatim across your site, social profiles, Wikipedia if you qualify, company registers and authoritative directories;
• use structured data (schema.org/Organization) on your site, so that systems integrating search find a well-defined entity;
• systematically pair the name with the sector in your copy ("[name], logistics software") to supply the context models currently lack;
• stop feeding ambiguity with alternative spellings in your own materials.

If, however, the collision is with an entity far better documented than yours, honesty is in order: you will not win it with content alone. The question then becomes whether to live with the confusion, differentiate the name with a stable added element, or change it while changing is still cheap.

A rule of thumb

Before you fall for a name, do three things that take ten minutes: ask two different AI assistants "what is [name]?", dictate it over the phone to a colleague and see what they write down, and search the exact name in quotes. If all three come back clean, the name has a good chance of being yours in the machines' answers too. If even one fails, you have just saved yourself a rebrand.