How to Track Your Brand in Gemini

Of the four assistants we monitor, Gemini produces the most confused questions. Not because the product is harder to use, but because the word points at more than one thing, and what people are usually staring at when they say "my brand in Gemini" is not the Gemini assistant at all.

The disambiguation comes first. It is most of the article.

Several things are called Gemini

Gemini, the assistant

The chat product you reach at gemini.google.com and inside Google's mobile app. You ask a question in plain language, it writes back a paragraph or a list, and the exchange sits in a conversation with a history. It is the direct counterpart of ChatGPT, Claude and Perplexity: somewhere you go on purpose.

Gemini, the model family

The language models Google trains and versions. They power the assistant, developers can call them through Google's API, and Google embeds them across its products. When a startup says its feature "runs on Gemini", it means the models, not the chat app. Nothing about your visibility lives at this layer, but it is why the name turns up everywhere.

Gemini-derived models inside Google Search

Here is the actual source of the confusion. The AI-written summary above the links on a Google results page, and the conversational search experience Google has been expanding, are generated by models from the same family. They are not the assistant. They live inside the search product, they are triggered by a query rather than opened on purpose, and they are grounded in Google's index and ranking machinery in a way the assistant is not.

Same brand of engine, different vehicle. And, routinely, different answers to what looks like the same question.

Why the distinction is not pedantry

Imagine a marketing manager who types her category into Google, reads the AI summary at the top, sees a rival named and her own company missing, and concludes she has a Gemini problem. She might. She might also have a search problem that behaves nothing like the assistant. The two surfaces differ in ways that change the outcome.

The input is not the same shape. People type fragments into a search box and full sentences into an assistant. "best crm small business" and "we are a small consultancy in Italy, which CRM would you recommend and why" pull in different material, even from related models.
The grounding is not the same. The search surfaces sit on top of the web index Google already ranks. The assistant leans more on what the model absorbed during training, reaching out to the web when it decides to.
The summary does not always appear. Google does not put an AI answer on every query, and what triggers one keeps shifting. An assistant always answers.
One of them is attached to a page of links. In search, the AI text sits next to the organic results you have spent years working on. In the assistant there is nothing underneath to fall back on. You are in the sentence or you are nowhere.

Presence in one is weak evidence about the other. Treating them as a single measurement is how people end up reporting a number that describes a surface they never checked.

Where PSentry stands, plainly

PSentry measures the Gemini assistant, alongside ChatGPT, Claude and Perplexity. It runs your prompt set against the assistant, in each language and market you sell in, and reports whether you were named, how you were described, which sources were used, and which competitors were recommended in your place.

It does not measure Google's AI Overviews. It does not measure Google's AI Mode. It does not measure Copilot. If a tool tells you it covers "Google AI", ask which surface it queries, because these are separate products and an honest answer is a specific one. Ours is: the assistant, and only the assistant.

Checking by hand, so the result means something

Do this once by hand before you buy anything, including from us.

Open the assistant, not the search box

Go to the Gemini app or site directly. If you typed your query into Google and read what appeared above the links, you were looking at a different product, and whatever you concluded belongs to it.

Ask as a customer, not as the brand

"What is [your company]" teaches you nothing. Hand a model a proper noun and it will find something to say about it, and you walk away reassured for no reason. The questions that matter are the ones where your name does not appear: who supplies this category in your region, what a mid-sized manufacturer should use for this job, what the alternatives are to the best-known player in your space. Then read the answer and see who is in it.

Start from a clean session

The assistant knows things about you: earlier conversations, saved information, your Google account, your interface language, your account's region. All of it can shade the answer. Use a temporary or unsaved chat if the product offers one, or a signed-out session in a browser profile you never use. Otherwise you are measuring what the assistant remembers about you, not what it says about your brand.

Ask more than once

Ask the same question several times, each in a fresh conversation, spread over a few days. You will get different answers. That is the model working as designed, not a glitch. A single response is an anecdote. A pattern that repeats is a measurement.

Change the language

Ask the same question in every language you sell in, phrased the way a native buyer would phrase it, not as a translation of your English prompt. Visibility does not cross borders. A brand the assistant recommends comfortably in English can be absent from the German or Spanish answer, because the German or Spanish web says less about it. For an exporter, that gap is usually the most useful thing on the screen.

Write down what you see

For each run, record four things: whether you were named, whether the description was accurate, who else was named, whether there were links. Being described well with no link, and being linked while someone else gets the recommendation, are different situations with different remedies.

What this test cannot tell you

A procedure handed over without its limits is a sales pitch.

The answers are probabilistic. The same prompt, in the same language, on the same day, can return a different set of recommended vendors. Nothing you do removes this. It is why one check proves nothing, and why serious measurement repeats prompts and aggregates them.

Personalization contaminates the sample. Your account, your history, your settings and your language all feed into what you personally are shown. The answer on your screen is not necessarily the one a stranger in your target market receives, and you cannot fully strip that out by hand.

Conversation memory is the quietest trap. Ask about your category, then ask a follow-up that mentions your brand, and everything after that point is polluted: your name now sits in the context window and the model will gladly reuse it. Every question you intend to measure needs its own conversation, opened from nothing.

The assistant changes under you. Models get updated, sometimes quietly. A result you captured last quarter is a historical record, not a current fact. That is the argument for checking on a rhythm rather than once.

None of this reaches into how the assistant actually decides what to say. There is no submission form, no ranking lever, and no setting that nudges Gemini toward your name, only a reading of what it currently does with your prompts, checked again on a twice-a-month rhythm rather than watched live. What a model says about a brand comes from what the web says about it, and moving that is slower and less glamorous than any dashboard suggests. Promising that the next reading will look better than this one is not something the assistant, or anyone measuring it, gets to offer.

Frequently Asked Questions

If Google's AI summary names my brand, am I also named in the Gemini assistant?

Not reliably. The two draw on related models but different retrieval, different triggers, different question shapes. If both matter to you, check both, and keep the results in separate columns instead of averaging them into one comforting figure.

Will PSentry cover Google's search surfaces later?

We are not going to sell you a roadmap in a blog post. Those surfaces are a genuinely different measurement problem, and quietly folding search results into an "AI visibility" number is a shortcut we do not want to take. Today the product covers four assistants, and it says so.

The assistant linked to my site. Is that a mention?

Not the same event. A citation is a link. A mention is your name inside the sentence the buyer reads. A model can cite your documentation while recommending a competitor, and it can recommend you warmly with no link. Track them separately or you misread both.

Why does the assistant describe my company wrongly?

Usually because the web does, or because what the web says is thin enough that the model fills the gaps by inference: outdated pages, a rebrand the internet has not caught up with, a positioning that exists only in your own marketing copy. The fix sits upstream of the assistant, and it is slow.

How often is it worth re-checking by hand?

Monthly at most, unless something happened: a launch, a rebrand, a wave of coverage. Model behaviour toward a brand shifts slowly, so checking daily mostly measures the model's randomness. Our scans run twice a month, roughly the pace at which the picture changes.