The short answer
An assistant names a business when it can retrieve a page that answers the question, trust the source, be confident the business is real and specific, and find the business a good fit for what was actually asked. Fail any one of the four and you are not in the answer.
First, retrieval decides who is even considered
Before anything is judged, a shortlist is assembled. Either the model recalls what it absorbed in training, or it runs a live search and pulls a handful of pages back.
That shortlist is short. A live retrieval might read five to ten sources for a question. Everything after this point happens inside that small set, which means the whole game is being in it. Being crawlable, indexed and topically obvious is what buys the ticket.
Second, source trust weights what it reads
Not all retrieved pages carry the same weight. A model has strong learned priors about which kinds of source tend to be reliable, and it applies them.
A review platform, an established directory, an industry publication or a public register are treated as stronger evidence than a company describing itself. This is the single most under-appreciated point in the whole subject, and it is why outside corroboration outperforms another page of your own marketing.
It is also why consistency matters so much. When outside sources describe your business the same way your own site does, every source reinforces the others. When they disagree, they cancel out.
Third, entity confidence decides whether it will say your name
There is a real difference between a model knowing a page exists and a model being confident enough to put a business name in front of a user who might act on it.
Confidence comes from a business being resolvable. One name, one address, one phone number, one description, marked up in structured data with a stable identifier, matching what public records and outside profiles say. When those all line up, the model is dealing with a fact. When they do not, it is dealing with a guess, and it will usually name a competitor it is more sure about.
This is where a surprising number of otherwise strong businesses lose. Their content is good and their identity is fuzzy.
Fourth, answer fit decides which of the finalists gets named
The last filter is the narrowest. Given several trustworthy, well-identified candidates, the model names the one that best matches what was actually asked.
Specificity wins here. A page about one problem, in one sentence, with a number in it, fits a specific question better than a page covering everything at a general level. If someone asks about a niche situation, the business with a page about exactly that situation gets named, even if a larger competitor covers the subject more broadly.
That is the opening for smaller businesses. You cannot outspend a large competitor, but you can be more precisely right about a narrower question, and precision is what this last filter rewards.
What this means practically
Work the four in order. Retrieval is access and indexing. Trust is corroboration you do not control. Confidence is entity hygiene and schema. Fit is writing genuinely specific pages about genuinely specific questions.
Most businesses skip straight to writing more content, which is the fourth filter, while failing the first. That is why the content does not work.