Someone asks Gemini for the best project management tool for a 12 person agency and gets four names back. If one is yours, that is a lead you never saw arrive. If none is, that is a loss you never saw either. The questions worth answering are how often that happens and on which prompts.
You can track brand mentions in Gemini, but only if you are precise about which Google product you mean. A method built for one produces garbage on another.
Table of Contents
Gemini, AI Mode and AI Overviews are three different products
Most guides skip this, which is why so much Gemini advice is wrong. People say “Gemini” and mean any Google surface with generated text.
The Gemini app is a chat assistant. You open it deliberately and have a conversation. It answers from the model, and sometimes runs a web search first and grounds the answer in what it found. When it does not search, there are no sources to inspect.
Google AI Mode is a conversational version of Search itself, where the generated answer takes follow up questions. AI Overviews is the generated block above the classic blue links, which are still underneath it.
| Surface | Where it lives | How a user gets there | Sources |
|---|---|---|---|
| Gemini app | Its own app and website | Opens an assistant deliberately | Only on grounded answers |
| AI Mode | Inside Google Search | Switches into it from a search | Links inline and in a panel |
| AI Overviews | Top of the results page | Runs an ordinary search | Links above the organic results |
The same prompt sent to all three returns different brands in a different order, and the clicks land in different places in your analytics. This post is about the assistant. For the other two, see our guide on optimizing for Google AI Overviews, which covers the Search side including AI Mode.
What is measurable in Gemini, and what is not
Three things move the answer underneath you, before run to run variation:
- The account. A signed in user with history and personalization can get a different answer from a signed out one on the same prompt, so your customer’s Gemini is not yours.
- Whether it searched. A grounded answer names more brands and shows sources. An ungrounded one reflects what the model absorbed in training, which is slower to change and harder to influence.
- Country and language. Ask in Spanish from Spain and in English from the US and you get different brand sets, not translations.
There is no ranking position in Gemini. The only ordinal signal is where your brand appears relative to the others in the answer, which matters because the first name in a list of recommendations gets read and the fifth often does not.
This is a sampled measurement, closer to polling than to a census. Google publishes no report of how often your brand is named in the assistant, so every number you see, ours included, is an estimate. The direction of a good sample over twelve weeks still holds up, even when a single answer is noise.
Method one: manual checks with proper hygiene
Manual checking is fine for a first look and for reading one answer closely. Most people do it badly: one run, logged into their own account, then quoted in a meeting.
- Use a temporary chat or a signed out session. Your personal history is a thumb on the scale.
- Set the locale you care about. Ask in the language your customers use, from the country they are in. Three markets means three checks.
- Repeat the prompt three to five times in fresh sessions. One answer tells you almost nothing. Five tell you whether you appear roughly always, roughly never, or somewhere in between.
- Record the date, the exact prompt, whether sources appeared, and the brands named in order. A screenshot without a date is not evidence of anything a month later.
- Do not follow up in the same chat before recording the result. Follow ups inherit context and change the next answer.
What it cannot give you is history. You will not rerun forty prompts across five markets by hand every Monday, and if you try you will do it inconsistently. See our guide to tracking brand mentions across AI assistants.
Method two: scheduled tracking
What a tool adds is repetition: the same questions on a schedule, in a clean session, with every answer kept. That buys four things manual checks cannot:
- The same prompt set over time, so a drop is a real drop and not someone rephrasing a prompt.
- Competitors measured on the same answers, counted out of the same response text on the same day.
- Sentiment. Being named as the expensive option and being recommended both count as a mention.
- The sources behind the answer. When Gemini grounds an answer, the pages it used are the lever you can act on.
We compared the tools that do this in a separate post. LLM Pulse tracks Gemini on every plan alongside ChatGPT, Perplexity, Google AI Mode and Google AI Overviews, so you can put one prompt against the assistant and both Search surfaces and see how far apart they land. Prompts run weekly by default, with daily and monthly available.
The caveat applies to every tool including ours: a scheduled run uses its own clean session. That is the right baseline over time, though it is not what your logged in customer with two years of history sees. Read the result as an index, since it counts answers and not people.
Building a prompt set for Gemini specifically
Google’s assistant gets used heavily for comparison and recommendation questions, and for local and shopping intent. Weight the set that way. A keyword list reused as it stands will miss most of it.
- Category: “best CRM for a small sales team”, “tools for freelance invoicing”.
- Comparison: “X vs Y for enterprise”, “compare warehouse management options”.
- Alternatives: “alternatives to X”, “what to use instead of X offline”.
- Best X for Y: the qualifier decides the answer. “Best accounting software for a UK sole trader” names different brands from “best accounting software”.
- Local and shopping: “where to buy X near me”, “running shoes for flat feet under 120 euros”.
- Branded, kept separate: “is X any good”, “X pricing”, “X reviews”.
Read that last bucket on its own. A prompt that names your brand will mention your brand, so mixing branded prompts into your visibility number inflates it and hides the trend that matters.
A weekly prompt produces about 4.33 scheduled answers per model per month, so 50 prompts across five models is over a thousand answers a month. Ten prompts is not enough to see a rate move, and fifteen non branded questions plus five branded ones is a reasonable first set.
What to track
Five numbers cover most of it, and the definitions matter more than the names. Our breakdown of which GEO metrics matter goes deeper on each.
- Mention rate: answers naming your brand divided by total answers, times 100. This is the headline number.
- Position weighted visibility: the same count weighted by where you appear. First brand named counts 100%, second 50%, third 33%, and so on. We call it the AI Visibility Score. It separates “listed last among eight” from “recommended first”.
- Share of voice: your mentions divided by the mentions of every brand in your tracked set, times 100. Tells you whether a flat mention rate means stability or means everyone else is growing.
- Sentiment: the split of positive, neutral and negative framing across answers that mention you.
- Cited sources: the domains grounded answers pull from. The most actionable of the five, and tracking which sources get cited is where the work usually starts.
Where Gemini traffic shows up in analytics
GA4 has an AI Assistant default channel group. Google defines it as “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”, matched when the medium is exactly ai-assistant or when the referrer matches Google’s list of known AI assistant sources (Google’s channel group documentation).
Gemini generally does not append campaign parameters to the links in its answers, unlike ChatGPT, which adds utm_source=chatgpt.com. Attribution therefore depends entirely on the referrer surviving the click, and referrers get stripped often enough that your AI Assistant number is a floor rather than a total.
The same documentation defines the Organic Search channel to include “Google’s AI Overviews and AI Mode”, so clicks from those two land in Organic Search instead, and no single GA4 channel covers everything Google’s AI features send you.
LLM Pulse’s AI Traffic analytics connects to GA4 (and to Plausible, Piano, PostHog and Adobe) and reports the human visits arriving from AI assistants next to the visibility data.
How to improve what Gemini says about you
The levers are mostly off your own site, and they are slow.
Grounded answers pull from pages that already rank and from the sources that dominate your category, which in practice means review sites, comparison pages, forums and trade publications more often than your homepage. Getting named there moves more than another landing page does. Your own content has to be quotable: direct answers near the top, specifics instead of adjectives, tables a model can lift a fact from.
Ungrounded answers are harder, because they change on the model’s schedule, which you do not control. Google publishes an AI features optimization guide, and the honest summary of it is that there is no separate Gemini ranking system to optimize against. Start from your cited sources report: the domains appearing in answers about your category are a target list.
FAQ
Does Google give me a report of my brand mentions in Gemini?
No. Search Console has a Generative AI performance report, but it covers AI Overviews and AI Mode, not the Gemini app, and reports impressions only: “how many times links to your site were shown to a user in a generative AI feature on Google Search”. No clicks, no CTR, no position. Dimensions are pages by canonical URL, countries, dates and devices, and the usual limits apply, including the 1,000 row cap (Google’s documentation). It also only counts links being shown, so it never tells you your brand was named without one.
Do Gemini and AI Overviews give the same answers?
No, and do not assume the brands overlap. They ground differently and answer at different lengths. Run both if both matter to you, rather than treating one as a proxy for the other.
How many runs before the number means something?
For a manual check, three to five runs tell you whether you appear reliably. For a tracked number you want dozens of answers per prompt, which is why weekly scheduling adds up faster than it looks: 4.33 answers per prompt per model per month is a usable series within a quarter.
Does Gemini always show its sources?
No. Sources appear when the answer grounded itself in a web search. An answer produced from the model alone has none, and that is a data point in itself: if your category consistently returns ungrounded answers, fresh content will influence it slowly.
Can I track Gemini in different countries and languages?
Yes, and you should if you sell in more than one market, because brand sets genuinely differ by locale. Treat each country and language pair as its own prompt set: averaging hides the market where you are invisible.
Is there a Gemini equivalent of rank tracking?
Not in the sense of a position from 1 to 10. The closest equivalent is mention rate combined with position weighted visibility, which together answer how often you appear and how prominently. Anyone selling a “Gemini ranking” is putting a familiar label on a different measurement.
