How AI decides who to recommend
There is no ranking algorithm to reverse-engineer here, and no dashboard to log into. But the pipeline between a question and a name in the answer is not a mystery either. Four things happen, and you can affect three of them.
Step one
The question gets rewritten before anyone searches
A person types something loose: “who can fix a walk-in freezer in melbourne, urgent”. The assistant does not search for that string. It turns the request into one or several cleaner queries, often adding terms the user never wrote, then runs those.
This matters more than it sounds. You are not competing for the phrase the customer typed. You are competing for the phrases a model thinks that phrase means, which tend to be more formal and more categorical than real speech. “Emergency commercial refrigeration repair Melbourne” is what actually gets searched.
Step two
A shortlist of pages comes back, and it is short
The rewritten queries hit a search index. What comes back is not ten results to browse, it is a handful of documents that will be pasted into the model’s context as raw material. Everything past that cut-off is invisible for this answer, no matter how good it is.
Two practical consequences. First, ordinary search visibility still matters, because that index is usually a search index. Second, being on page two is functionally the same as not existing, which was never quite true before.
Worth knowing: some assistants retrieve live at question time, others lean on what was in training, and several do both and blend the results. That is why the same question can name different businesses in ChatGPT, Perplexity and Google, and why you should test all of them rather than assuming one stands for the rest.
Step three
The model writes an answer it can defend
Now the model has a pile of text and a question. It has to produce a short answer, and increasingly it has to attach citations to the claims in that answer. That constraint shapes everything about what gets used.
It reaches for sentences that are specific, self-contained and attributable. A line stating your service radius, your response time or your licence number can be lifted and pointed at. A paragraph of positioning language cannot, because there is nothing in it to cite.
This is the single biggest lever most businesses have, and it costs nothing but a willingness to be pinned down in writing.
Step four
Agreement across sources breaks the tie
When several candidates could fill the same slot, corroboration decides it. A claim that appears only on your own website is one source saying one thing. The same claim on an industry association register, a council supplier list, a trade directory and a case study on a client’s site is four independent confirmations.
None of those need to be links in the SEO sense. What is being weighed is whether the world agrees you are what you say you are.
What you cannot control, and should stop worrying about
- Model training cut-offs. If a model last trained before you existed, no amount of work reaches into it. Live retrieval is your route in, not training data.
- Which assistant a customer uses. You can only make yourself retrievable and quotable in general and let that play out across all of them.
- Answer phrasing. It is generated fresh each time and it varies between two runs of the identical prompt. Judge patterns across many prompts, never a single reply.
Anybody offering to submit you to ChatGPT, guarantee a citation or place you in an AI Overview is describing something that does not exist.
The short version
Be reachable by the crawlers. Be in the search index the assistant queries. Put specific, checkable claims in plain sentences. Get those same facts confirmed somewhere other than your own website. That is the whole game, and it is unglamorous on purpose.
See what an audit checks