If you want to track visibility in Gemini, the first thing to understand is that one answer won't tell you much. This guide walks you through gemini tracking as a repeatable routine: you set up a fixed group of buyer questions, collect several answers for each, and score every answer for mentions, tone, accuracy and sources. By the end you'll have a record you can read by topic, so you can see how brand mentions in Gemini differ by topic, and where it describes your brand well and where it leaves you out or gets something wrong.
You don't need special software to start. A spreadsheet, a few fresh Gemini chats and a short list of approved facts about your company are enough. A tool can take over the collecting later, and we'll point out where that makes sense.
Step 1: Decide what the tracking is for
Before you write a single prompt, pick one brand or product, one audience and one decision the results should help you make. That might be "does Gemini include us when someone asks for a shortlist in our category," or "does Gemini explain our main feature correctly," or "which buyer concerns come with a good or bad description of us." Keep it narrow enough that you can review every answer and act on it.
Write this down as a short tracking brief. It should name the audience, the product, the market or location if that matters, and the decision. You're done with this step when someone else on your team could read the brief and understand what the numbers are meant to support.
The most common mistake here is treating the brand name alone as the whole program. If you only ask Gemini about your own name, you learn how it describes you when people already know you, which is the easy case. Most of the useful information is in the questions where buyers haven't named anyone yet. And if you later want to cover Google's other search surfaces too, the roundup of the best Google AI Overviews tracking tools is a good next read.
Step 2: Build a prompt set from real buyer questions
Now write the questions themselves, in the words a buyer would actually use. Include some that mention your brand and some that don't, and spread them across awareness, consideration and decision-stage intent where it makes sense. A few templates you can fill in:
- "Which [category] options support [use case] for [buyer type]?"
- "How does [Brand] handle [specific capability or constraint]?"
- "Is [Brand] a fit for a team that needs [requirement]?"
- "What should a buyer check when choosing a [category] provider for [use case]?"
Replace the brackets with verified details about your product and your buyers. If you're unsure how to balance branded and unbranded questions, it helps to read about the four types of prompts worth tracking before you finalize the list. A second read on branded, unbranded and competitor prompts covers the same balance from another angle.
Then give every prompt a topic label, such as product fit, integrations, pricing, implementation or one specific use case. These labels are just a working set for your own sheet, so use whatever groups match how your buyers think. The labels are what let you read results by topic in Step 7.
Once tracking starts, keep the wording fixed. That's the part people find hardest, because you'll notice a clumsy prompt on day three and want to fix it. If you change the wording, you can't tell later whether the result moved because Gemini changed or because your test did. If a prompt really is broken, retire it and add a new one as a separate group.
DeepSmith can help at this step. Its Discover Prompts feature suggests candidate prompts from your product, persona and buyer-stage context. Treat what it gives you as a starting list. Read each one, remove the vague or duplicate ones, and keep only the questions your audience could really ask. You're done when the list covers your priority topics and buyer stages, the wording is stable, and every question has a topic label or is left unlabeled on purpose.
Step 3: Fix your Gemini test conditions and write them down
Gemini's answers can shift with the surface you use, your location and your account settings, so decide on one setup and record it. For manual checks on a computer, open Gemini's web app, type the prompt into the text box and submit it. Start a new chat for every separate test instead of continuing an earlier conversation, because a long chat carries its own context into the next answer.
Gemini picks a model automatically by default. If you choose one by hand, write down the label you see on screen and keep using it, since the available choices change over time and you don't want to rely on a list that may be out of date. Location matters too. Gemini may use your general location and, if you allow it, your precise device location. In the web app, the location settings show whether precise location is in use, with a blue indicator when it is and gray when it isn't.
A Temporary Chat is one way to cut down on personalization from your past chats. On a computer, open Gemini, select Temporary chat next to New chat, enter the prompt and submit it. Google's help pages say it needs a signed-in personal Google Account, and it isn't available in Gemini in Chrome or Gemini in Google Messages. Temporary Chats don't give personalized responses and don't show up in your recent chats, but Google keeps them with the account for 72 hours, so don't think of them as never stored. Because some account types can't use the feature, it's fine to document whatever condition you used instead of treating Temporary Chat as the only correct way to test.
Record these for every run: the date, your location or locale, whether the model was automatic or chosen, the account context, and whether it was a regular chat or a Temporary Chat. If a setting isn't available to you, write that down instead of guessing. You're done when every observation can be tied to a date, a prompt and a known set of conditions.
The mistake to watch for is mixing ordinary chats, Temporary Chats, different locations and different model choices in one trend line without noting the change. One more habit is worth building here. Keep confidential company information out of personal Gemini chats, since Google's privacy information says chats can be used for personalization when memory is on.
Step 4: Collect several answers for each prompt
A single Gemini answer is a sample, and the next one may look different. Run each fixed prompt in separate new chats, or use a tracking service that collects each prompt on a schedule and keeps the answers (the roundup of tools to track your brand's visibility in Gemini lists the options). As a practical starting point, collect three independent responses per prompt in each measurement cycle. That number is our suggestion, and no research has established it as a minimum. For prompts that swing a lot or claims that carry real weight, collect more and always say how many you collected.
There is some research behind the idea of repeating. One empirical study of repeated AI-search runs reported source overlap of 34% to 42% between consecutive days, and brand-set overlap of 45% to 59%, with wide variance. Those figures come from that study's own campaigns and conditions, so they aren't a Gemini benchmark and they aren't a forecast for your brand. They're a good reason not to trust a single run. If you want a fuller explanation of why this happens, why AI answers change from run to run covers it.
DeepSmith checks the prompts you track on a schedule you set, and it keeps the full answer history for each one, so you aren't copying and pasting every answer by hand. Gemini tracking is included on the Scale plan and higher. If you need several samples inside a single cycle, confirm the collection settings available to your account, or do separate manual runs and label them as manual.
You're done when each prompt has several separately recorded, valid responses for the period and you know exactly how many. A useful habit is to keep two counts, the runs you planned and the answers you actually captured. A tool error or a missing response says nothing about whether Gemini would have named you, so it should never be counted as an omission.
Step 5: Save the full answer and the sources it shows
For each valid run, keep the whole answer text and any source links that appear with it. In the Gemini Apps interface, a Sources button may show up inline or below the answer, and Gemini can also show related content, which may include public websites, uploaded files or connected Google Workspace documents. Record each link as it's displayed and note whether the interface called it a source or related content. Don't assume that every related link was used to write the answer.
Keep those original links apart from what Double-check response gives you. Double-check uses Google Search to find content that's likely similar to or different from the statements in the answer. On a computer, you select More under the response, then Double-check response, and then select a highlighted statement to see what Search found. According to Google, green means Search found content likely similar to the statement, but the linked page isn't necessarily what Gemini used to produce the answer. Orange means Search found content likely different from the statement, or found nothing relevant. No highlight means there wasn't enough to evaluate, or the statement wasn't meant as a fact. These are checking cues, so don't read them as a final verdict on whether a claim is true.
For each displayed source, log the page or domain, whether it belongs to your brand, a competitor or neither, and whether Gemini marked it as a source or only as related content. If an answer displays no source at all, write "no source displayed." If you couldn't retrieve the answer or its evidence, log that as a collection failure, which is a different thing.
You're done when the answer and its original links are preserved and any Double-check links sit in their own labeled field. The common mistake is counting a Search link from Double-check as if Gemini had cited it. Even when Gemini shows sources, the answer can still be wrong, so keep a healthy distance between "a link was shown" and "the claim is correct." The same distinction shows up in the way AI citations and brand mentions are two separate signals.
Step 6: Score mentions, tone and accuracy as separate fields
Use one row per prompt run, and fill in these fields independently of each other:
| Field | How to code it |
|---|---|
| Valid response | The answer was captured and can be reviewed. If not, log the failure reason. |
| Brand mentioned | Yes if the answer names your brand or a verified alias, otherwise no. Count it once per response, however many times the name appears. |
| Brand sentiment | Positive, neutral or negative, based on what the answer says about you. If you're absent, use "not applicable" instead of neutral. |
| Description accuracy | Check each material claim against approved, current facts. Accurate, partly accurate, inaccurate or not assessable works as a simple rubric. |
| Owned citation displayed | Yes only if a displayed source is a page on your own site. Record the exact page. |
| Other source type | Competitor, third party or other. Keep related content separate from source citations. |
The accuracy rubric is our own suggested audit method, so it's not a Gemini feature, and you can adjust the labels to fit your team. What matters is that tone and accuracy stay apart. A positive answer can still be wrong, and a negative one can be a fair description of a real limit. To keep reviewers consistent, give them a short fact sheet of approved claims and have them write down the exact claim they judged right or wrong.
If you're using software to detect your brand name, keep a list of aliases and spot-check the results. The same study that measured overlap also points out that fixed substring matching can miss abbreviations and paraphrases, and can flag common word fragments by mistake. A person reviewing a sample catches most of that.
DeepSmith tracks sentiment as one of its visibility metrics, alongside mention rate and citation rate. A sentiment label from any tool tells you about tone, and it doesn't prove the statement was accurate, so keep your own accuracy check in place. If you're comparing tools for this part of the job, the guide to the best AI brand sentiment tools can help. You're done when each valid answer has its own mention, sentiment, accuracy and source records, with "not applicable" or "not assessable" used where they honestly apply.
A mistake that's easy to make is treating a citation as proof of a mention, a recommendation or a correct description. They're three different observations, and a brand can be named without its site being cited, just as a site can be cited without the answer recommending the brand.
Step 7: Calculate your rates and read them by topic
With a clean log, you can work out a few simple numbers to track visibility in Gemini over time. Calculate each one at the prompt-run level and always state the denominator.
- Mention rate: valid responses that mention your brand, divided by all valid responses.
- Owned citation rate: valid responses with at least one displayed citation to a page you own, divided by all valid responses.
- Sentiment mix: the count of positive, neutral and negative descriptions among the responses that actually discuss your brand. Show the count next to the share.
- Accuracy rate: claims judged accurate, divided by claims you could review. Say how many were not assessable.
Then group the results by the topic labels from Step 2. For each topic, show the number of valid responses, the mention rate, the owned citation rate, the sentiment counts and any notable accuracy errors. Keep the individual prompt results one click away, so that a topic average never hides a single question with a very different outcome. A missing answer should never be silently counted as "not mentioned" or "not cited." When you compare two periods, use the same prompts, the same conditions and the same denominator logic. If you have no earlier baseline, call the result your starting position instead of implying a change from zero.
This is the step where a platform saves the most effort. DeepSmith's AI Visibility overview reports mention rate, citation rate and share of voice with trends, and the prompt views show per-prompt rates and answer history, including category-level results so you can see which groups of buyer questions get more or fewer mentions and citations. It also classifies cited source URLs as yours, a competitor's or neither. If you'd like to see how a spreadsheet compares with a tool as the prompt list grows, when a manual prompt-tracking sheet stops scaling is a fair look at that tradeoff, and the wider category is covered in our roundup of the best AI visibility tools.
You're done when each rate has a stated numerator, denominator, period and count of valid responses, and each topic result traces back to individual prompts. The mistakes to avoid are comparing a period full of collection failures with a fully observed one, and reading a small one-off sample as a stable ranking.
Step 8: Turn the repeatable gaps into an action
Now look at the brand mentions in Gemini that need attention, which are the rows where Gemini leaves your brand out, mentions you without citing your pages, shows a competitor or third-party source, or says something inaccurate. Pick one documented issue, give it an owner and write down the next content or product-information action. It's better to fix one gap and measure it than to start ten.
DeepSmith can show which of your pages get cited and which prompts bring those citations in, along with the exact competitor pages cited for the prompts you track. That's useful for deciding which page to update or which page to write. It doesn't mean that publishing or editing a page will make Gemini cite it, so keep your expectations modest and your measurement honest. If you're looking for ways to act on a gap once you've found it, the guide to the best tools to optimize your brand for Google Gemini is a reasonable place to look.
After you make a change, run the same questions again under the same conditions and report what you saw, including the sample size and any uncertainty. Keep the original question set stable for the next cycle, and add new buyer questions as a separate, labeled group so the trend stays readable. You're done when each priority gap has an evidence record, an owner or next action and a planned re-measurement.
What to do next
Keep the same sampling and review rhythm, and expand the prompt set only when new buyer questions or real product changes call for it. If you'd rather not collect and score every answer by hand, you can see how Gemini tracking works inside DeepSmith and start a free trial to try it on your own prompts.
Related guides
- How to Check Your ChatGPT Visibility Manually, Step by Step: Check your ChatGPT visibility by hand: write a prompt set, run each in a clean session, log mentions, position and citations in a sheet, and repeat on a.
- How to Choose the Right Solution for ChatGPT Visibility Tracking: Decide between manual checks, a general AI visibility tool and a ChatGPT specialist using criteria such as prompt volume, budget, reporting needs, team time and how much detail.
- What Your CMS Needs to Support for AEO: Evaluate whether your CMS supports AEO by checking structured data and schema control, clean semantic markup, fast updates, redirect and metadata handling, and what to ask vendors before.
- Enterprise AI Visibility Platform Requirements: Security, Governance and Scale: Build a requirements checklist for an AI visibility platform covering SSO, role permissions, audit logs, API access, data residency and scale limits, so you can score vendors on.
- Entry-Level vs Enterprise AI Monitoring: Is the Price Gap Justified?: Weigh what enterprise AI monitoring adds over entry-level plans, such as coverage, integrations, governance and support, and decide whether those extras justify the price gap for your situation.


