How AI decides which businesses to recommend
What happens between someone asking ChatGPT for a plumber and a name appearing in the answer, and the three places a local business can actually influence it.
The most useful thing to understand about AI recommendations is that there is no ranking algorithm.
With Google there is a scoring system. It is enormously complicated and nobody outside Google knows it fully, but it exists, it produces an ordered list, and an industry has spent two decades inferring its shape.
An assistant naming three plumbers is doing something different. It is closer to what a fast, slightly credulous researcher would do with ten minutes and a search box.
That difference is good news, because it means the levers are ordinary.
What actually happens
Four steps.
1. Someone asks. “Who’s a good physio in Castle Hill?” Note that this is a question, not a keyword. People type differently to an assistant than they do to Google, more conversationally and with more context.
2. The model looks. For anything local or current, it searches the live web, because its training data is months out of date and it knows that. It also draws on what it already learned during training, which is where long-established businesses have a quiet advantage.
3. It picks sources. This is the step that decides everything. From what it found, it selects a small number of pages it can read cleanly and has some reason to trust. Not the ten best. A handful it can actually use.
4. It writes the answer. It summarises those sources into a paragraph, naming the businesses that appeared credibly across them.
There are diagrams for all of this on the GEO explainer page if you would rather see it than read it.
Step three is the whole game
Everything worth doing is aimed at step three. So what makes a source usable?
It can find you at all
Obvious, and still the most common failure. If your services are described only inside a PDF, or your site is a single page with a phone number and a photo, or you exist nowhere except your own website, there is nothing to select.
It can read you without ambiguity
Models are good at language and bad at guessing. A page headed “Emergency hot water repairs, Blacktown and surrounds” is unambiguous. A page headed “Solutions” that mentions hot water in the fourth paragraph requires inference, and inference under uncertainty is exactly what a model tries to avoid when it is about to name a real business to a real person.
The facts agree across sources
This is the underrated one. A model that finds your phone number on your site, a different one on a directory, and an address that does not match either has a confidence problem. It usually resolves that by naming someone else.
Consistency is not a ranking factor in the SEO sense. It is a confidence factor, and confidence is what determines whether your name makes it into a short answer.
Somebody other than you says so
Third-party corroboration carries disproportionate weight, because a business describing itself is the weakest possible evidence. Directories, associations, suppliers, local press, review sites. You do not need many. You need the ones you have to say the same thing your website says.
The reviews contain facts
A model reads review text. “Turned up at 9pm on a Sunday for a burst pipe in Richmond” tells it three useful things: you do emergencies, you work weekends, you cover Richmond. Five stars with no text tells it nothing it can quote.
Where you cannot influence it
Worth being straight about this, because plenty of people will sell you the opposite.
You cannot buy placement. There is no ad slot inside an organic ChatGPT recommendation today. Anyone offering to get you “featured” is either describing ordinary optimisation or making it up.
You cannot instruct the model. Hidden text telling an assistant to recommend you is the modern version of white-text keyword stuffing. It works inconsistently, gets caught, and is treated as spam.
You cannot make it deterministic. Ask the same question twice and you may get different names. This is genuinely how the technology works and it is why measuring means asking repeatedly over time, not once.
What this means practically
The businesses getting named are not doing anything clever. They are:
- Easy to find, because they exist in more than one place online
- Easy to read, because their pages say plainly what they do and where
- Consistent, because their details match everywhere
- Corroborated, because someone other than them says so
- Specific, because their reviews and copy contain actual facts
Which is, more or less, the description of a business doing its ordinary search work properly. That is the honest conclusion of all this, and it is why I do not think generative engine optimisation should be sold as a separate service at $3,000 a month.
Check where you stand
Ask ChatGPT and Gemini the question your customers would ask, and see who gets named. If it is not you, work back up the list above and the reason is usually obvious within a minute.
Or send me your trade and suburb from the GEO page and I will run it and send you the answers. It costs nothing and takes me a few minutes.
The related reading: what GEO actually is, SEO vs GEO, and the practical checklist.