How AI assistants choose which businesses to recommend

Ask an AI assistant to recommend a business and it will hand you a short, confident list. It feels like a judgement about quality. It is not. It is the model assembling an answer from a few signals, and once you know what they are, the shortlist stops looking random.

Two sources feed the answer

Every recommendation an assistant gives comes from one of two places, usually both:

1. What it learned in training. The model has read a huge slice of the web. Businesses that were described clearly and often, on credible pages, are the ones it can recall and name.
2. What it reads live. ChatGPT, Gemini and Perplexity can search the web while answering. Perplexity does this on almost every query. So the freshest, best-cited pages at the moment of the question also shape the answer.

This is why two assistants give different names for the same question. One leans more on training, another on live search. It is also why being absent from the live, citable web hurts you even if you have been around for years.

The signals that decide the shortlist

Across both sources, the same handful of signals keep deciding who gets named:

• Clarity. Can the model state, in one line, what you do and where? A business described as "ducted heating installation in the Hills District of Sydney" is easy to recommend. "Comfort solutions for your home" is not.
• Credible mentions. Reviews, local press, reputable directories, customer stories. These are the evidence the model uses to justify naming you. Quality and relevance beat raw quantity.
• Consistency. The same business name, phone and service area everywhere. Conflicting details make a model uncertain, and uncertainty loses you the slot.
• Relevance to the exact question. "Best emergency plumber in Geelong" and "best bathroom renovation plumber in Geelong" can return different names. The model matches the specific service, not just the trade.
• Recency and citability. For the live-search part, having current pages that are easy to cite puts you in the running at the moment it counts.

Notice what is not on the list: how good you actually are at the job, how long you have traded, or how much you would like to be recommended. The model cannot see those directly. It sees how well you are described and referenced.

Why this is good news for smaller businesses

Because the signals are about description and evidence, a sharp small operator can beat a much larger competitor. A solo electrician with a clear site, real reviews and a couple of local mentions can outrank a big firm that hides behind vague marketing copy. The model rewards legibility, and legibility is free.

It also means the businesses AI names are not always the ones you would expect. We see this constantly in the Australian AI Visibility Index: the leaders in a category are often not the biggest brands, but the ones the assistants understand best.

What this means for you

If you want to be recommended, you are really trying to do three things: be easy to describe, be backed by credible evidence, and be consistent everywhere. That is the whole game, and it lines up exactly with why a competitor might be getting picked ahead of you, which we cover in Why ChatGPT recommends your competitor.

The one thing you cannot skip is looking. Because the assistants disagree and pull from different sources, you have to check each of the four rather than guess from one. A free AI check shows you what they currently say about your business and, just as useful, who they name instead.