What Google does and does not publish about Gemini's business recommendations

As of 29 September 2026 we are not aware of any Google document that sets out how Gemini decides which companies to name in an answer. Google publishes a great deal about how its models work and how content is used, and none of it is a company recommendation formula. So the honest number on this page is zero: zero official pages that we have been able to find specifying such a formula. Anything a vendor tells you about "Gemini's ranking factors" for business mentions is inference, including ours.

That is not a complaint. It is the starting condition for anybody trying to get named, and it decides what work is worth doing: you measure behaviour, because the rulebook is not published.

The direct answer: no public company recommendation formula found in Google's cited documentation

Three things are true at once and are usually confused with each other.

• Google documents its products. There is real, detailed material on the Gemini apps, on the API, on grounding with search, and on how site owners can control what is crawled and shown.
• Google does not document a business ranking. None of that material says: to be named as a recommended company, satisfy these criteria, weighted in this order.
• The absence is not a secret being kept from you. A generated answer is produced from the model's weights and whatever sources were retrieved for that request. There may be no stable formula to publish, which is a different situation from one that exists and is hidden.

How we searched, so you can judge the zero. We looked through Google's own published product and developer documentation for the Gemini apps and the Gemini API, Google's Search documentation on AI features and on content guidance for site owners, and Google's public policy and help material on how recommendations and business information are handled. We did not find a page stating criteria for naming companies. This is a statement about our reading of the public documentation on the date above. It is not a claim that no such page exists anywhere, and if you find one, the change log at the bottom is where it will be recorded.

What Google does publish

These five categories are worth reading directly, because they are the only authoritative material in the area.

• Product help for the Gemini apps, covering what the app can do, what it may get wrong, and how to give feedback on an answer.
• Developer documentation for the Gemini API, including how grounding with search works when it is switched on, and what the response returns about the sources used.
• Search documentation on AI features, covering eligibility for appearing in AI experiences and the controls a site owner has over snippets and indexing.
• General guidance for site owners on useful content, structured data, and the technical requirements for being crawled and indexed at all.
• Policy and safety material describing categories of answer that are handled with extra care, which is relevant if you sell in health, finance, law or anything involving minors.

Read each one on a date and write that date down next to the quotation. These pages are revised, and a quotation with no date is the thing an assistant discounts fastest.

What the documentation does not establish

Six specific things people believe are documented and are not.

• That there is a ranked list of factors deciding which businesses get named.
• That schema markup causes a mention. Structured data helps a machine read your facts. No published document says it earns a recommendation.
• That reviews on a particular site are weighted. No public weighting exists. Our own capture points the other way: the domains cited most often across the whole capture were the assistants' own documentation and the vendors' own websites, with a single well known review site appearing far down the list. Perplexity, September 2026, on aiknowsus.com.
• That a search ranking transfers to an answer. Related, not the same. Being first on a results page is not documented as producing a mention.
• That every answer is grounded in a live search. It is not safe to assume this on any assistant, and there is a separate page on this site about how to test it.
• That a correction request changes an answer. Feedback routes exist. A documented guarantee that using them changes future answers does not.

A reproducible test of company mentions

Because the formula is not published, the only honest way to know where you stand is to count. The test is deliberately simple.

The method

• Write 30 buyer questions for your category, in buyer language, with your name in none of them. Keep them frozen.
• Start a new chat for every question. No memory, no custom instructions, no follow-ups in the same thread.
• Record the surface and the model name the interface shows, plus the date, time and time zone.
• Save the answer text and every displayed source link as a file per answer.
• Score one thing first: did the answer name any company at all. Then score whether it named yours.
• Repeat weekly for four weeks, and report a count per run rather than an average.

The two figures this produces

The first is answers naming at least one company, over the total answers tested. That tells you whether your category gets company-level answers at all, and it is the figure most businesses have never looked at. Some questions produce advice with no vendor in them, and no amount of work on your site will change those.

The second is answers naming your company, over the same total. That is your visibility, and it only means something next to the first figure.

What we have measured, and on what

We have not run this on Gemini and publish no Gemini figures. On other assistants, with dates: on 18 August 2026, across eighteen blind commercial questions on clawlaw.in, ChatGPT made the company the top source on exactly one of them, and its own comparison page was named in the answer as still being the vendor's own editorial page. In September 2026, across 24 batches and 72 conversations on aiknowsus.com, the phrase recording that we were not cited appears 161 times in Perplexity's own self audits of its answers.

Claims, evidence, and confidence levels

Every claim on this page, with what stands behind it and how sure we are. This section exists because the alternative is a page of assertions that a careful reader has no way to grade.

• "No public Gemini business recommendation formula exists in Google's documentation." Evidence: our own documented reading of the public documentation on 29 September 2026. Confidence: moderate to high on what we read, and we cannot prove a negative about every page Google publishes.
• "Assistants sometimes answer without a live search." Evidence: ChatGPT confirmed it had run no live web search for any of 78 blind questions on 27 July 2026, on clawlaw.in. Confidence: high for that run, and it is one run on one assistant.
• "An assistant's own account of its process can be wrong." Evidence: asked afterwards how many questions it had searched for, Perplexity withdrew its earlier statement and said it could not honestly substantiate the claim, in three separate batches in September 2026 on aiknowsus.com. Confidence: high, and it is one assistant.
• "Official sources outrank vendors on how-to questions." Evidence: on 17 September 2026, on clawlaw.in, official court portals took every position above any commercial product on the questions about how to look a case up, and Claude said plainly that no commercial product should rank above the official portal for a question about using that portal. Confidence: high for that question type in that market.
• "Publishing a specific page can be followed by a citation within two weeks." Evidence: ChatGPT cited a named clawlaw.in URL on 6 August 2026, fourteen days after publication. Confidence: the event is high, the causal link is low, because a single sequence is not a cause.
• "Structured data earns you a Gemini recommendation." Evidence: none that we hold. Confidence: we do not make this claim.

What this page cannot settle

It cannot tell you what Google does internally. It cannot rule out documentation we have not seen, in a language we did not search, or published after the date above. It cannot tell you whether a mention you already have came from the model's memory or from a retrieved page, because that is only visible per answer, at the time, from the interface. And it cannot be turned into a checklist that produces a mention, because the thing that would justify such a checklist is exactly the document we could not find.

Common questions

If nothing is published, is any of this work worth doing?

Yes, because the behaviour is observable even where the rules are not. You can see which pages an assistant reaches for, and you can see what it says about them. On 17 September 2026, on clawlaw.in, two enterprise vendors were ranked above the company on a question about finding every case against a company, specifically because they publish explicit court and tribunal coverage lists, and Claude noted that its ordering reflected price transparency and source authority rather than product quality. That is a usable instruction even though nobody published it as a rule.

Does Google Business Profile control what Gemini says about my company?

It is a source of business information and keeping it correct is basic hygiene. It is not documented as a control over what a generated answer says, and treating it as one leads businesses to fix a profile and leave the rest of their public facts contradicting each other.

Why do vendors publish lists of "Gemini ranking factors" then?

Most are inference from observed answers, some of it careful. The problem is presentation: inference written as documentation. Ask for the prompt set, the dates and the log. If those do not exist, you are reading an opinion, which may still be useful if it is labelled as one.

Should I add schema markup anyway?

Yes, for the plain reason that it makes your facts machine readable and cheap to check, and for the unglamorous reason behind one of our own findings: a pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of the company's own site. Claude, 6 August 2026, on clawlaw.in. Being readable is a lower bar than being recommended, and failing it is common.

Can I pay to be recommended by Gemini?

Not through any mechanism we are aware of, and anybody offering it should be asked to document it. We sell measurement, not placement, and we cannot promise a position in any assistant's answer.

How often should I re-read Google's documentation?

Once a quarter is enough for most businesses, and any time an answer about you changes noticeably. Keep the date you read it next to anything you quote from it.

Primary sources and change log

Go to the source rather than to summaries of it: Google's Gemini app help centre, the Gemini API developer documentation including the grounding and search tool pages, Google Search Central's documentation on AI features in Search and its guidance for site owners, and Google's public policy pages on how its AI products handle sensitive categories. Our own observations come from the clawlaw.in programme from July 2026 and the aiknowsus.com audit of September 2026, both described with dates on this site.

Change log. First published 29 September 2026. The claim of zero official pages is dated on purpose. If Google publishes such a document, or a reader sends us one we missed, the claim will be corrected here with the date of the correction and the original wording kept visible.

Disclosure. Written by AI Knows Us, which sells AI visibility measurement, so we have an interest in you measuring things.

What to do first

Do not go looking for the formula. Write 30 blind questions, run them on Gemini once this week, and count two numbers: how many answers named any company, and how many named yours. Then read Google's own help and developer pages with the dates noted. In an afternoon you will know more about your position than any published factor list could tell you.