How Does Google Gemini Decide Which Businesses to Recommend?

A client asks Google Gemini for someone who does what you do. It comes back with three names, a sentence about each, and a note about what to look for. You are not one of the three, and there is no obvious reason why not.

The useful question is not whether Gemini likes your website. It is what Gemini is actually doing between the question and the three names, because that sequence has one step in it that you control and several that you do not.

This post walks through the sequence in plain English, shows what Gemini can and cannot read about a small business, and ends with the one change worth making this month.

How does Gemini decide which businesses to recommend?

It answers from two sources at once. Some of the answer comes from what the model learned when it was trained, which is a fixed snapshot of the web from some time ago. The rest can come from a live look at Google Search, which is current. For questions about real businesses, the live look does most of the work.

Out of what that look returns, Gemini names the businesses it can describe confidently. Confidence here is not a score you can see. In practice it comes down to whether several independent places say the same thing about you, in terms that match the question the buyer asked.

So the honest one-line answer is this: Gemini recommends businesses that are easy to describe. Not the best ones, not the closest ones, and not the ones with the nicest site. The ones whose description holds still across every place it can read.

The learned half and the live half

Google publishes the mechanism for the live half. Its developer documentation for grounding with Google Search describes a tool that lets the model run a search and build its answer from what comes back, with links to the pages it used. The consumer app draws on Search in the same spirit for anything that would otherwise be out of date.

The learned half is the opposite. It is large, general, and old, and it is where a model gets its sense of which brands exist and what words go together. It is also where stale descriptions of your business survive longest, because nothing in it gets corrected when you change your pricing page.

That split explains something that confuses a lot of owners. Gemini can be right about your industry and wrong about you in the same paragraph. The industry part came from the learned half. The part about you came from whatever the live look happened to surface, which may be a directory entry you forgot existed.

The same two-part structure sits behind the other assistants, with different plumbing. We walk through the equivalent sequence in how ChatGPT decides which businesses to recommend, and the source-first version of it in how Perplexity chooses its sources.

AGREEMENT, AND WHAT THE BUYER HEARSILLUSTRATIVE

Named outright

4 of 5 sources agree

For that, people usually go to [name], who works with exactly this.

Named with a hedge

2 of 5 sources agree

There is a [name] in that area, though I would check their current focus.

Not named at all

0 of 5 sources agree

Here are three well-known options, and some things to look for.

Illustrative counts, not measurements. The point is the direction: the more independent places that describe you the same way, the further up this list your buyer's answer lands.

AnswerHalo answerhalo.com

CLICK TO ENLARGE

What Gemini can actually read about you

Published text that Google has indexed. That is the whole list. Your own pages, your listings, the roundups you appear in, the review profiles, the interviews, the association directories. If it has a public address and Google has been to it, it is in scope.

Everything else is not. Your results with clients, the reason people refer you, the standard of your work, and the reputation you have in your own city are all real and none of them are readable. An assistant is not making a judgment about you when it leaves you out. It is reporting an absence.

Google's own guidance for site owners about its AI features in Search makes the same point in the other direction: the material these features draw on is the ordinary indexed web, and the ways to be eligible are the ordinary ones. There is no separate submission process and nothing to buy.

Why it names your competitors instead of you

Usually because they are easier to summarize. A competitor described in four places as an executive coach for first-time founders is a low-risk name to give. A better coach described everywhere as helping people reach their potential is a risk, because the assistant cannot tell whether that matches the question.

The second common reason is that the buyer's question was narrower than your pages are. Someone asking for help with a specific problem in a specific city is asking a question with a small field of plausible answers, and a general description does not enter that field at all.

Neither of these is a ranking problem, which is why the usual response of publishing more content rarely moves it. The fix is not more words. It is the same words in more places, saying one specific thing.

The reading that tells you which of the two is happening to you is a short one. Ask the assistant your buyer's question and look at who it names and what it says about them. Our walkthrough of that is in how to check if AI recommends you.

What to change this month

Write one sentence that says who you help and with what, in the words your buyer uses, and put it in the same form on your homepage, your listings, and every profile you control. Not a variation on each. The same sentence.

Then get one outside page to describe you that way too. A guest piece, an association profile, a partner's page, a genuine roundup. One is enough to move you from a single claim to an agreement, and agreement is what the figure above is about.

Do not try to write for Gemini specifically. Google's help documentation for the Gemini apps describes a general-purpose assistant, not a directory with a submission process, and there is nothing to opt into. What helps in Gemini is the same thing that helps everywhere: being described plainly and consistently.

Then leave it and re-read in a month. Answers vary between runs on their own, so a change you make on Tuesday is not something you can confirm on Wednesday. One reading a month is enough to see a direction, and it is a reading rather than a guess.

If you want the first reading done for you, the free check asks ChatGPT 25 real buyer questions from your niche at $0 and emails a plain-English readout: your score, who got named in your place, and the one fix to start with. It runs as soon as you ask for it, and the report lands within the hour between 7am and 5pm Central, or by 10am the next morning outside those hours.