Brand Mentions in Gemini: How to Track Visibility, Citations and Recommendations

Brand mentions in Gemini are answers where Google's Gemini names your brand when someone asks a question. A mention is not the same as a citation, which links to your website, or a recommendation, which tells the person to choose you. You track all three by running a fixed set of prompts on a schedule, recording every answer and counting each signal against a stated denominator.

Brand mentions in Gemini: the short answer

1. A mention is a name in the answer text. If Gemini writes your brand name, that is a mention. Nothing more is implied.
2. A citation is a link, and Gemini does not always give one. Google's own help pages say that not every Gemini response includes sources, so a mention with no citation is normal.
3. A recommendation is a separate, rarer event. Being listed among five options is not the same as being told "choose this one".
4. The same prompt can return different answers. Grounding, personalization, location, conversation history and model updates all move the result, so one screenshot proves nothing.
5. You measure patterns, not positions. A prompt set, a schedule and a record per answer turn a moving target into numbers you can compare month to month.

This page covers Gemini specifically. If you want the general concept first, read what AI search visibility is and why it matters. If you want the same method for OpenAI's engine, the sibling page on monitoring brand mentions in ChatGPT covers it. Everything below is specific to Gemini: how it builds answers, what to count, what to control and what to disclose.

What counts as a brand mention in Gemini, and what does not?

Most arguments about Gemini visibility are really arguments about definitions. One person says "Gemini mentioned us", another says "no, it linked to a review site that talked about us", and both are half right. Fix the vocabulary before you count anything. The table below gives the six signals we record for every Gemini answer, with the Prime AI Visibility reporting term in brackets where one exists.

Signal	Definition	How you see it in the Gemini app	Record as
Mention (Brand named)	The answer text names your brand or an approved name variant	The name appears in the written response	Yes or no per answer
Citation (Website cited)	The answer includes at least one link to a website you own	Your domain appears in the Sources panel or as an inline link	Yes or no per answer, plus the URL
Recommendation (Explicit recommendation)	The answer clearly tells the person to choose, try or prefer your brand	Language such as "the best fit here is", "I would go with" or "start with" attached to your name	Yes or no per answer
Co-mention	Your brand is named in the same answer as at least one tracked competitor	Two or more tracked names in one response	Yes or no, plus which competitors
Sentiment or accuracy issue	The answer says something about your brand that is negative, outdated or false	A wrong price, a discontinued feature, a wrong headquarters, a "known for poor support" line	Yes or no, plus the exact sentence
Source reference	A page that talks about your brand is in the sources, but it is not your website	A review site, a directory, a news story or a competitor's comparison page in the Sources panel	The referring URL

Three rules keep the table honest.

First, one answer counts once per signal. An answer that names you four times is one mention. An answer with three links to your site is one citation with three source links. Prime's published measurement definitions work the same way, and it is the reason "Source links" is reported separately from "Website cited".

Second, a citation without a mention still counts as visibility. Gemini sometimes links to a brand's page without writing the brand's name in the text. The person can still click through, so the answer is visible even though Brand named is zero.

Third, a source reference is not a citation of you. If Gemini cites a review site that praises you, the review site earned the citation. You earned a mention at best. This distinction matters when you decide where to look first: a missing citation of your own site sends you to your own pages (are they indexed, do they answer the question plainly), while a source reference tells you which third-party pages appeared in Gemini's sources for your category. Neither is proof of why Gemini chose what it chose. A link is observed evidence, not a window into the model's reasoning.

Fourth, the accuracy column is not optional. Google's own guidance says Gemini responses can be inaccurate and should not be relied on as professional advice. When Gemini is wrong about your price, your features or your locations, the mention is working against you, and no mention count will show that.

Why does the same question give different Gemini answers?

Gemini is built in a way that makes single observations unreliable, and you need to understand the five moving parts before you can control them.

Grounding. Google documents that Gemini can ground a response with Google Search, meaning the model runs one or more searches, reads the results and writes an answer with citations to what it found. In the Gemini API, that comes back as grounding metadata: the search queries it used, the web sources it drew on and which parts of the answer each source supports. In the consumer app you see a Sources button when links are available. Google's help page is blunt about the gap: not all responses include sources, and if there is no Sources button, Gemini provided no links for that response. So whether your brand gets a citation depends partly on whether the model decided to search at all. Indexing is worth checking too. Google Search Central's guidance on AI features says that to appear as a supporting link in AI Overviews or AI Mode a page must be indexed and eligible to show with a snippet. That rule is written for Google Search's own AI features, not for Gemini, and Google does not publish an equivalent rule for grounded Gemini answers. We treat an unindexed page as the first thing to rule out, not as a documented cause.

Personalization. Google states that Gemini Apps can personalize responses using the memory of past chats, content and activity in Google apps the person has connected, and saved instructions about how they want Gemini to respond. These features require a personal Google Account and are not available on work or school accounts. A signed-in marketer who has spent months discussing their own brand with Gemini is the worst possible tester of that brand's visibility.

Geography. Gemini can use device location with permission, and it can use saved Home and Work locations from the Google Account. For a local brand, a prompt tested from the office will not represent a buyer two states away. For a national brand, record the country and language you tested from anyway. Google does not document how account region and language affect a grounded Gemini answer, which is a reason to hold them constant, not a reason to ignore them.

Date and model behavior. Google updates Gemini's models and features over time, and its release notes will not tell you which change, if any, affected a specific answer. Grounded answers also change when the underlying search results change. Treat every answer as a dated observation and never compare answers gathered weeks apart as if the engine had stood still.

Conversation state. A brand missing from the first answer often appears when the person asks a follow-up. That is real buyer behavior, but for measurement you need a fixed starting point. Every tracked prompt goes into a fresh chat with no prior turns.

One more gap sits underneath all of this: the Gemini API and the Gemini consumer app are not the same surface. API-based observations do not carry a person's account history, and they can differ from what a signed-in person sees. Prime AI Visibility records answers through API-based checks and says so on its methodology page. Whichever surface you use, name it in your report. If you are curious how this compares with OpenAI's approach to retrieval, see why ChatGPT and Gemini reach different brand answers.

How to track brand mentions in Gemini with a repeatable prompt set

Here is the method we use. It works with a spreadsheet or with a platform, and the point is that anyone on your team can repeat it next month and get a number that means the same thing.

Step 1: build a prompt set of 20 to 40 questions

Write the questions your buyers actually ask, in their words, not yours. Cover three stages: learning about the category, comparing options, and deciding. Add prompts that name a competitor, prompts that name you, and prompts that describe a situation without naming anyone ("we are a 40-person accounting firm and need a client portal"). If location matters to your business, write location-specific versions. Keep the exact wording in a shared document. Changing "best" to "top" or adding "in 2026" is a new prompt, not the same prompt.

Step 2: fix the controls

Decide once, write it down, and do not vary it without noting the change:

• Account state. Signed out, or a dedicated test account with personalization turned off and no chat history. Never a personal account.
• Location. Location permission denied in the browser, or a stated test location, with the country and language recorded.
• Surface. Gemini web app, Gemini mobile app or the Gemini API with grounding enabled. Pick one and name it. Record the model name if the app shows one.
• Session. A new chat per prompt. No follow-ups in the measured run.
• Runs. Two or three runs per prompt per cycle, because a single run cannot separate a real change from ordinary variation.
• Schedule. Monthly is enough for most brands. Weekly makes sense during a launch or after a major content change.

Step 3: record every answer in full

Save the complete response text, the date and time, the account state, the location setting, the surface and every URL in the Sources panel. Screenshots are useful as evidence but poor as data. You want the text, because the classification in the next step has to be checkable by someone else later.

Step 4: classify each answer against the six signals

Go through the definitions table above for every saved answer. Two people should classify the first cycle independently and compare, because "is this a recommendation or just a mention?" is exactly where teams disagree. Write down the tie-break rules you agree on and version them. When the rules change, the numbers are no longer comparable, and your report must say so.

Step 5: count with stated denominators

Percentages without denominators are how reports on brand mentions in Gemini lose credibility. For each metric, say what the top number is and what the bottom number is. Answers that failed to return or were refused are "missing" and do not appear in either number, but you report how many were missing. If 30 prompts times two runs requested 60 answers and 57 came back usable, the denominator for your rates is 57, and you say so.

This is the discipline the cross-engine guide to tracking AI brand mentions applies to every engine. Gemini simply has more controls to lock down than most, because it sits so close to Google Search and to the person's Google Account.

Which AI visibility metrics matter for Gemini?

Keep the metric set small when you report brand mentions in Gemini. Every metric below comes from the record you built in Steps 3 and 4, and every one has a denominator you can defend.

Metric	Formula	What it tells you	Common misread
Visible in AI answers	Answers that name the brand or cite its website ÷ usable answers	How often a buyer meets you at all	Treating it as a ranking; there is no position in an answer
Brand named rate	Answers naming the brand ÷ usable answers	Whether Gemini knows you belong in the category	Counting a name inside a cited page title as a mention
Website cited rate	Answers with at least one link to your site ÷ usable answers	Whether your own pages appear among the answer's sources	Counting a review site's link as your citation
Source links	Count of individual citation links across all answers, with links to your own site tallied separately	Which pages, yours and other people's, the answers pointed to	Reading five links in one answer as five visible answers
Explicit recommendation rate	Answers that clearly recommend the brand ÷ answers naming the brand	Whether being known turns into being chosen	Calling any positive adjective a recommendation
Co-mention rate	Answers naming the brand and a tracked competitor ÷ answers naming the brand	Who Gemini puts you next to	Assuming co-mention is bad; in comparison prompts it is expected
Accuracy issue rate	Answers with a false or outdated brand statement ÷ answers naming the brand	Whether visibility is helping or hurting	Ignoring it because the mention rate looks good

Two notes on scope. Explicit recommendation, co-mention and accuracy issues use answers naming the brand as their denominator, not all answers, because you cannot be recommended or misdescribed in an answer that never mentions you. And "usable answers" excludes missing answers, which you report as a separate count so the reader knows how complete the cycle was.

You will notice what is not on the list. There is no Gemini rank, because an answer has no rank. There is no traffic estimate, because answer observations do not convert into clicks by any published method. There is no forecast. Prime AI Visibility's methodology page makes the same exclusions for the same reasons, and it adds a 0 to 100 Prime AI Visibility Score as a summary of observed visibility across a check's scope, not as a universal Gemini ranking.

What must be controlled or disclosed in every Gemini report?

A Gemini visibility number without its conditions is a rumor. Put this block at the top of every report, and if any line changes between cycles, flag the comparison as not like for like.

• Prompt set version. The exact wording, the count of prompts and the date the set last changed.
• Personalization state. Signed out, or a named test account with personalization off, memory empty and no connected Google apps.
• Geography. Country, language and whether location permission was granted, and to what location.
• Date window. The dates the answers were collected. A cycle collected over three days is one observation window; note it.
• Surface and model. Gemini web app, mobile app or API with grounding, plus any model name shown. If the surface changed, the series restarts.
• Runs per prompt. How many times each prompt was asked in the cycle.
• Denominators. Requested, returned, usable and missing answer counts.
• Classification rules version. Who classified, and which version of the tie-break rules they used.
• Who is reporting. If a vendor produced the report, name the vendor and any commercial interest. Prime AI Visibility sells the platform described later on this page, so when we publish Gemini numbers we say so.

This is what "comparable" means in practice. On the Prime platform it is formalized as Comparable change: a comparison that uses a sufficiently aligned question set, tools, persona, market and measurement method. Without that alignment, a jump from 30 percent to 45 percent Brand named might be a real gain, a personalization leak, a location change or a Gemini model update, and nobody can tell which.

Worked example: 30 prompts over one month

The numbers below are illustrative. They show the arithmetic, not any client's results.

A regional payroll software company writes 30 buyer prompts and runs each twice in a signed-out Gemini web session with location permission denied, on the first Tuesday of the month. It requests 60 answers. Three return with no usable content, so 57 are usable and 3 are reported as missing.

• Brand named: 21 of 57, or 37 percent.
• Website cited: 9 of 57, or 16 percent. Six of those nine answers also named the brand, so Visible in AI answers is 24 of 57, or 42 percent.
• Source links: 34 individual citation links across the 57 answers. Eleven of them, spread over the 9 cited answers, pointed at the company's own site: eight to the pricing page and three to a comparison article.
• Explicit recommendation: 4 of the 21 answers that named the brand, or 19 percent.
• Co-mention: 17 of 21, or 81 percent. Two competitors appeared in almost every one of those answers.
• Accuracy issue: 3 of 21, or 14 percent. All three repeated a starting price that changed eight months ago.

What the team does with it: the price error is the first fix, because it is hurting 14 percent of the answers that already include them, and the old price still lives on two third-party listings that show up as source references. The pricing page drew most of the owned links. That does not prove Gemini prefers it, but it makes that page the obvious first place to correct the number and to add the plan comparison buyers keep asking about, as a hypothesis the next cycle can test. The low recommendation rate is not an emergency. It is the metric to watch after the fix, on the same prompt set, under the same controls, next month.

Should you track Gemini manually or with a platform?

Manual tracking of brand mentions in Gemini is the right way to start, because it forces you to write the prompt set and the rules. In our experience it breaks down as the prompt set, the runs per prompt and the number of people involved grow, because saving, classifying and the denominator math take longer each cycle and the controls start to slip.

A platform earns its place when it locks the controls for you and keeps the raw answers. If you are evaluating tools, what an AI visibility tool measures explains the category and its limits without vendor language.

Here is our own disclosure. Prime AI Visibility is our product. Verified on the live site on 23 September 2026: Gemini checks are available on the Growth plan at $179 per month, and on the Agency and Enterprise plans, alongside ChatGPT, Perplexity, Claude and a Google AI Overview-style preview using Google-grounded answers. The free Flash plan checks ChatGPT only, and Starter covers ChatGPT, Perplexity and the AI Overview-style preview. One question in Gemini costs two account credits, monthly checks are the default, and weekly or temporary daily schedules are optional. The current plan and credit details may change, so check them before you decide. Prime records answers through API-based checks, which can differ from a signed-in person's consumer-app experience, and it does not report rank, traffic estimates, forecasts or causal attribution. If that workflow fits how you want to work, the walkthrough of how the saved-evidence workflow runs shows each step from question discovery to comparison.

Common mistakes when tracking brand mentions in Gemini

• Testing from your own account. Months of your own Gemini chats about your brand are personalization fuel. Use a clean, signed-out session or a dedicated test account.
• Counting a source reference as a citation. A review site linking to you in the Sources panel earned that citation. Record it as a source reference and go read what the review site says.
• Comparing across surfaces. API answers, web app answers and mobile app answers are three series. Do not splice them.
• Reading one run as the truth. Two runs of the same prompt can differ. Report the rate across runs, and say how many runs there were.
• Rewording prompts between cycles. Every rewrite resets the series. Version the set and keep the old wording live until you have a full cycle on the new one.
• Declaring victory on mentions. Being named in 60 percent of answers while a third of them repeat a wrong price is not a win. Accuracy issues belong on the same page as the mention rate.
• Attributing a change to a page edit. You changed a page and Brand named went up. So did Gemini's model, possibly, and so did the search results feeding it. Report the change and the edit side by side, and let the next two cycles tell you whether they move together.

References

1. Google AI for Developers, Grounding with Google Search (Gemini API documentation, 2026). https://ai.google.dev/gemini-api/docs/google-search
2. Google, View related sources from Gemini Apps (Gemini Apps Help). https://support.google.com/gemini/answer/14143489
3. Google, Get personalization in Gemini Apps (Gemini Apps Help). https://support.google.com/gemini/answer/16598623
4. Google, Find places and get directions in Gemini Apps (Gemini Apps Help). https://support.google.com/gemini/answer/16622866
5. Google Search Central, AI features and your website (2026). https://developers.google.com/search/docs/appearance/ai-features
6. Google, Learn about responses from Gemini Apps (Gemini Apps Help). https://support.google.com/gemini/answer/16279220

Next steps

1. Read the cross-engine guide to tracking AI brand mentions to apply the same record-per-answer discipline to ChatGPT, Perplexity and Claude.
2. Review the reporting vocabulary Prime uses for Brand named, Website cited and Explicit recommendation so your spreadsheet and any platform report count the same things.
3. When you are ready, create a Prime AI Visibility workspace and bring the 20 to 40 prompts you wrote in Step 1.