← Guides

How AI assistants decide which businesses to name

Published 3 Nov 2026 · 8 min read

AI assistants decide which businesses to name by interpreting your question, usually issuing a search, retrieving a small set of pages, and composing an answer from what those pages say. The names in the answer are, in most cases, the names in the retrieved sources. Getting named therefore depends less on persuading a model and more on being present, clearly described and corroborated in the pages it pulls.

What happens between the question and the answer

When someone asks an assistant who the best joiner in Mansfield is, the reply does not come from a list of approved businesses held somewhere inside the product. It is written on the spot, out of text.

The sequence runs roughly like this. The question is interpreted and rewritten into one or more search queries. Those queries go to a web index. A set of results comes back, and a small number of them are fetched. The text of those pages is placed alongside the original question, and the model writes an answer using it.

That pattern has a name: retrieval-augmented generation, usually shortened to RAG. The model supplies the language and the judgement. The retrieved documents supply the substance, including every business name that ends up in the answer. OpenAI, Google, Microsoft, Anthropic and Perplexity have all publicly described their assistants as searching the web and citing what they find, so the general shape of this is a matter of record even where the internals are not.

The practical consequence is blunt. The question "how do I get named" largely resolves into a different question: what do the retrieved pages say, and is your business in them.

Two modes: recall and retrieval

An assistant can answer in two quite different ways, and the difference changes what you should do about it.

From training data. The model produces a name because that name appeared often enough, in enough contexts, in the text it was trained on. No search happens. This mode is fast, works without the live web, and is frequently out of date — it can name businesses that have closed, repeat an old address, or describe a service a company no longer offers. Training data is fixed at the point the model was built, so nothing you publish this month can affect it until some future model is trained.

From live retrieval. The assistant searches, reads and cites. This mode reflects the web as it is now, which means changes to your own pages, and to the pages that describe you, can show up within days rather than years. It is also the mode in which most buying questions get answered, because assistants tend to search when a question is local, commercial or time-sensitive.

Nearly all the work worth doing targets the second mode. You cannot edit a model's memory. You can change what a crawler finds this week, which is why what an assistant can actually read matters more than almost anything else in this field.

The retrieved pages largely decide the names

Picture the model's position at the moment of writing. It has your question and perhaps a handful of documents. Those documents mention a few businesses between them. The answer has room for two or three names.

A business that appears in none of those documents is not in contention, regardless of how good it is, how long it has traded, or where it sits in a search ranking. This is the most common single cause of the mismatch where a business ranks well and is still never named.

It also reframes the useful question. Not "how do I improve my score", but "which pages get retrieved for the questions my customers ask, and am I on them". That second question you can answer by hand in an afternoon.

Why corroboration across independent sources matters

A model writing a recommendation is doing something close to summarising a small evidence file. One document saying a dental practice in Leeds offers same-day emergency appointments is a claim. Four documents from four different publishers saying it, in similar terms, is a pattern.

Your own website is the weakest item in that file — not because it is untrustworthy, but because it is expected. Every business describes itself favourably on its own homepage. A review platform, a trade body listing, a local news piece and a supplier's stockist page saying the same thing are independent of each other, and independence is what makes a description usable.

Consistency earns its keep here too. If your business name, address and service description differ across the places you appear, you have supplied contradictory evidence rather than corroboration. A model resolving a contradiction may simply name someone else. Third-party signals are, in practice, the lever most small businesses have left untouched.

Why directories and roundups punch above their weight

There is a structural reason these pages dominate citations, and it has less to do with their authority than you might assume.

A page titled "best joiners in Nottinghamshire" is already the shape of the answer. It contains eight named businesses, each with a sentence saying what they do, where they work and what they are known for. A retrieval pipeline that lands on it gets a complete shortlist from one document. Your homepage, however well built, names a single business and is usually written to convert a visitor rather than to summarise a market.

Directories and review platforms have the same advantage at larger scale: consistent fields, plain text, one business per record, and a category label that tends to match how people phrase questions. None of that is glamorous. All of it is easy to extract.

The implication is not that you should buy listings indiscriminately. It is that being absent from the two or three roundups and directories that actually get cited in your category is a larger problem than almost anything on your own site.

Where the public record ends

The broad pipeline above is described publicly by the companies that build these products. The details are not. How many results are fetched, how they are re-ranked before the model sees them, how heavily a source's reputation is weighted, whether a business mentioned once is treated differently from one mentioned twice — none of that is published for any of the major closed models, and anyone stating it as fact is inferring. What can be observed from outside is behaviour: ask the questions, record the names, repeat. Treat claims about internals, including ours, as inference unless they arrive with a citation.

Ask twice, get different names

Two people asking a word-for-word identical question can get different businesses named. So can the same person, minutes apart. Three separate causes stack up: the models sample rather than look up, the search results underneath them change, and consumer apps add memory and personalisation that the APIs behind them do not have.

Location handling adds another layer. An unqualified town name is genuinely ambiguous — Mansfield exists in England, Texas and Ohio, and Boston, Birmingham, Perth, Richmond and Hamilton all exist in several countries. Depending on what the assistant infers about where the asker is, the same words can return answers about the wrong continent.

Treating one conversation as a measurement is therefore a mistake in either direction. The variance is real and cannot be engineered away, so the only honest approach is repeated sampling across a fixed set of questions, read as a trend rather than a verdict.

What this means for what you should do

Work backwards from the pipeline rather than forwards from your website.

  • Find out what gets retrieved. Ask an assistant the ten buying questions your customers actually ask, and note which pages it cites. That list is your real competitive set.
  • Check whether you are on those pages. Not whether you could be, or whether you deserve to be. Whether you are, today.
  • Make your own pages extractable. One plain sentence stating what you do, for whom, and in which towns, written in text rather than baked into a hero image.
  • Make sure you can be read at all. A page behind a cookie wall, or assembled entirely in the browser by JavaScript, may be invisible to a fetcher even though it looks fine to you.
  • Measure repeatedly, never once. Non-determinism means a single absence proves little and a single mention proves less.

None of this is a trick, and none of it guarantees an outcome — the selection happens inside systems nobody outside the labs can inspect. What it does is move you from being absent from the evidence to being present in it, which is the part of the process actually within your control. If you want a starting point rather than a project, a baseline check will at least tell you whether you are in the conversation at all.

Common questions

Does the model have a list of businesses it recommends from?

No. There is no registry and no placement to buy. When an assistant browses, it names businesses it found in the pages it retrieved for that question. When it answers without browsing, it names businesses described often enough in its training data for it to produce the name with confidence.

Why does a directory page get cited more often than my own website?

A roundup or directory page is already shaped like the answer: several named businesses, each with a line describing what they do and where. One such page gives a retrieval pipeline everything it needs. Your own site, however good, describes one business and is usually written to persuade rather than to summarise.

If I ask the same question twice, why do the names change?

These models sample from a distribution rather than looking up a fixed answer, and the search results underneath can change between runs. Two people asking identically worded questions minutes apart can get different businesses named. A single check is close to meaningless; a pattern across repeated checks is not.

Can I tell whether an assistant searched or answered from memory?

Often, yes. Most assistants show citations or a search indicator when they have browsed. An answer with no sources and slightly dated details is more likely to have come from training data. The distinction matters, because the two modes respond to your work on very different timescales.

See whether the assistants name you

Three real customer questions for your trade and town, no signup, about a minute.

Run my free check

Read next