DeepSmith

Nov 26 · AEO & AI Visibility

19 min read

How to Track Your Brand in Google AI Mode

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
White cover line reading Track Your Brand in AI Mode over a dark charcoal background with a line-drawn chat bubble, connected link nodes and layered answer cards.

If you lead marketing and someone asks whether Google AI Mode ever mentions your brand, you probably don't have a clean answer yet. This guide walks you through Google AI Mode tracking by hand, from writing the questions to logging the answers to comparing results month over month, so you end up with numbers you can defend instead of a few screenshots. It covers only the conversational AI Mode surface. If you're weighing software for the wider Google picture, our roundup of the best Google AI Overviews tracking tools is a good place to compare options, and the steps below work whether or not you ever buy one.

You may see people call this an AI Mode rank tracker, but AI Mode doesn't give your brand a fixed position, so what you measure is whether you're included, which pages are linked, and how often that repeats. The short version to track AI Mode looks like this. Build a fixed list of real buyer questions, run each one as a fresh AI Mode sample under conditions you write down, save the full answer and every cited URL, and then compare brand mentions and owned-page citations across repeated runs and over time. The rest of this guide is how to do each part without fooling yourself.

You'll need a shared spreadsheet, a Google account you can use consistently for testing, and a couple of hours for the first wave. After that, each wave takes less time because the list and the sheet already exist.

1. Define what counts as a result

Before you run a single question, decide what you're going to count. In AI Mode there are two signals that people often blur together. A mention is an answer that names your brand. A citation is a link to one of your pages as a source. Either one can happen without the other, so you want to record them in separate columns from the start.

Here are the definitions to agree on with your team:

  • Valid answer sample: AI Mode gave an answer or a set of web links for the question. An answer with no link to your site is still a valid observation.
  • Brand mention: the answer names your brand, a product name, or a brand variant you've agreed on. Add a note when the reference is ambiguous.
  • Owned-page citation: at least one link in the answer points to a page on your domain.
  • Competitor mention or citation: write down who it was and, for citations, each linked URL.
  • No answer or collection failure: a page that didn't load, an access problem, or a broken run is not the same as an honest answer without a link. Never quietly turn missing data into a "no mention."

Save exact URLs, not just domains. A page-level record tells you which piece of content earned the link, and that's what you'll want later when you decide what to update. If you want more background on why the two signals matter separately, our piece on AI citations versus brand mentions covers it.

You're done with this step when everyone agrees on the brand names, product names, domain variants, competitors, and the separate rules for mentions and citations.

Common mistake: treating an answer that names your brand as proof that Google cited your site. Record both fields on their own, because they will disagree more often than you'd expect.

2. Build your prompt list from real buyer conversations

AI Mode is conversational, so the questions you track should sound like conversation. Start with what buyers actually ask. Good sources include customer interviews you have permission to use, sales calls, support tickets, community threads, and on-site search queries. Keep the wording close to the original where you can, remove anything personal or confidential, and group questions that are really the same question.

It helps to sort the list by buyer intent. Here are templates you can adapt to your category:

  • Awareness: "What should I consider when choosing a [category]?"
  • Consideration: "Which [category] options work for a team that needs [requirement]?"
  • Decision: "How does [brand or product] compare with [alternative] for [use case]?"

These are templates, not real customer wording, so swap the brackets for your own category and the language your buyers use. Don't make a separate prompt out of every small wording change. A new prompt is worth adding when the use case, the audience, the buying constraint, or the competitor changes.

A reasonable starting size is 20 root prompts. That number is a manageable place to begin, not a standard. Keep a stable core group that you never change, because that's what lets you compare waves, and keep a separate group for new questions. For each prompt, log the exact text, the intent label, where the question came from, and the date you added it. If you're used to working from keyword lists, our explanation of why AI prompts differ from SEO keywords shows how the two lists should look different.

You're done when every core prompt reflects a real buyer question, carries an intent label, and has wording you'll keep the same from now on.

Common mistake: shrinking real questions into keyword fragments. A phrase like "crm pricing small team" is not how someone talks to AI Mode, so the answer you get may not match what your buyers see.

DeepSmith can help at this point if you'd rather not start from a blank sheet. Discover Prompts generates a starter set from your product, persona, and buyer-stage context in DeepSmith, and you define the questions you track from there.

3. Set your sampling conditions before you test

Several things can change what AI Mode shows you, so if you want to track AI Mode in a way you can repeat, write them down before you start. At a minimum, log the date and time, the country or market, the device type, the browser or app, the state of your Google account, and whether personalization settings were on. If location permission is part of the test, record that as well. Google says current location can affect Search results, and that precise location may be used if you grant permission.

For a controlled comparison, keep the market, device class, account state, and settings the same in every wave. If what you really want is to see what one audience experiences, use that audience's normal settings and label the group clearly. Don't mix the two kinds of samples in one set of numbers, because then you can't tell whether a change came from your brand or from the settings.

Personalization deserves a closer look. Google describes Search Services History and Personalized Recommendations in Search services as two separate controls. You can reach Personalized Recommendations through your Google Account settings under Data & privacy, and Search Services History can be turned off on its own. Google also says these settings are being updated and rolling out gradually, so the exact path may look different for you. According to Google's documentation on how AI Mode personalizes responses, answers can draw on past searches and interactions when the relevant history and recommendation settings are on. That documentation says the experience is for people 18 or older in the US and in English, so don't assume it describes everyone who uses AI Mode.

If you're unsure why AI Mode needs its own samples, our comparison of AI Overviews and AI Mode for tracking explains the difference. Availability varies by country and language, so check Google's AI Mode help documentation for your market.

You're done when every observation can be grouped by a recorded, consistent cohort, and any change in conditions shows up in your log.

Pro tip: if you care about both a controlled view and a real-user view, run them as two separate tests and report them separately. Don't assume a private window or a fresh tab removes every trace of personalization or location, because Google doesn't say that it does.

4. Run independent repeat samples

One answer from AI Mode tells you very little, because the same question can come back differently from one run to the next. So you want repeats. For a starter protocol, run each of your 20 core prompts three times per wave. That gives you 60 answer observations. Treat those numbers as a practical suggestion for a first process, not a Google standard and not a statistically proven sample size. Three repeats won't give you certainty, but they will show you quickly whether one lucky answer is being mistaken for a pattern. If you want to understand why results move around, our guide to why AI answers change from run to run goes into it.

Use a separate, fresh conversation for each repeat of a root prompt. If you ask the same question again inside the same conversation, the earlier answer becomes part of the context, and you're no longer measuring a fresh sample. Run the same wording under the same conditions each time, and keep every result, including the ones you don't like. It's tempting to keep the answer that supports the story you want to tell, and that will quietly wreck your trend.

If you also want to study follow-up behavior, write a short, fixed sequence of follow-up questions and store each turn as its own observation, linked to the root prompt and the conversation it came from. Google describes follow-ups as a way to go deeper, so a follow-up is really a different prompt with earlier context behind it. Keep follow-ups out of the count for your root prompts.

You're done when the core set has been run repeatedly, and every attempt has a date, an environment, the exact question, and the answer.

Common mistake: re-running a prompt in the same continuing conversation and calling it an independent repeat. That measures a chain of conversation, not a fresh sample.

A word of caution on how you run these. Google's spam policies say that scraping results for rank-checking and other automated access without express permission counts as machine-generated traffic that goes against its policies and Terms of Service. So for a do-it-yourself process, run the questions by hand in the interface rather than building something that queries Google automatically.

For each run, save the full answer as text, not just a yes or no. Keep the sentence or passage that names your brand, every link shown as a source, the final URL after any redirects when you can get it, and any competitor names and links.

A sheet with the columns below works well. It's the same structure whether you keep it in a spreadsheet or a database.

FieldWhat to record
Run IDA unique identifier for this observation
Wave, date, and timeWhen the sample ran, using one stated time zone
Root prompt ID and exact textA stable ID plus the prompt as you entered it
Conversation turnRoot prompt, or a numbered follow-up linked to its root
Market and location contextCountry or market, and any known location condition
Device and access contextDesktop or mobile, browser or app, account state, and settings
Result statusValid answer, valid links-only response, access failure, or other issue
Full answerThe saved text or an archived copy of the response
Brand mentionYes or no, the exact brand or product name, and the excerpt
Owned citationsEach exact URL and its domain
Competitor mentions and citationsThe name and each exact cited URL, if any
NotesAnything unusual, or a reason the sample can't be compared

Google says AI Mode may return web links in cases where it doesn't have high enough confidence in the quality or helpfulness of an AI response. Count that as a valid result type and not as an error. Keep the result status in its own column, separate from "brand mentioned" and "owned citation," since both of those can be no on a perfectly good sample.

You're done when another person on your team can trace any number in your report back to a dated prompt, the saved answer, and the exact URL, or to a documented absence of one.

6. Score mentions and citations separately

Once a wave is finished, work out your numbers from valid answer samples only. Always state the denominator and show the raw counts next to the percentages, so nobody has to guess what "30%" is a percentage of.

  • Mention rate: valid answers that name your brand, divided by valid answers.
  • Citation rate: valid answers with at least one link to your pages, divided by valid answers.
  • Owned URL frequency: the number of valid answers that cite each exact page of yours. A page can be cited many times across runs.
  • Competitor visibility: the same counts for each competitor, using the same definitions and the same denominator.

These are working definitions for this guide, not formulas Google sets. Keep mentions and citations as two numbers instead of blending them into one score. If you also want a share-of-voice style figure, decide up front what the numerator and denominator are, whether you're counting answers or individual brand appearances, and use the same rule in every wave. Our overview of AI visibility metrics and KPIs can help you choose which ones to report.

If someone asks for an AI Mode rank tracker style report, this is the honest version: rates and cited pages, not a single rank. Here's a worked example with invented numbers, only to show the arithmetic. Say 60 valid answers include 18 that name your brand and 9 that cite at least one of your URLs. Your mention rate is 18 out of 60, or 30%, and your citation rate is 9 out of 60, or 15%. In a later wave of 60 valid answers you get 24 mentions and 15 answers with an owned citation, so the rates are 40% and 25%. That's a 10 percentage point change in each. None of these figures are research findings or customer results.

You're done when your report shows the numerator, the denominator, the percentage, and the exact prompt group for every rate.

Common mistake: saying a content change caused a rise in citations because the rise came after you published it. Sampling like this can show that two things moved together over time. By itself it can't show that one caused the other. If you're wondering how much movement to treat as real, our article on day-to-day noise in AI visibility scores is a good companion.

7. Compare waves, prompts, and cited pages

With a few waves in the sheet, compare like with like. It helps to look at three levels, one after the other.

  1. The whole group: mention rate and citation rate, with raw counts and denominators.
  2. Each prompt: which questions reliably produce a mention, an owned citation, a competitor citation, or nothing at all.
  3. Each page: which of your exact URLs show up, which prompts cite them, and whether the page behind a link changed between waves.

Read the answer excerpts before you trust a score. A mention can be qualified, out of date, or buried inside a comparison that doesn't favor you. A citation might point to a page you didn't expect. Write those distinctions down, and flag any classification you aren't sure about so a person can look at it. If competitors matter to your read, you can also line them up using the same sheet, and our guide to benchmarking your AI visibility against competitors shows one way to set that up.

Google Search Console can add some context here. Google's help pages describe a generative AI performance report that includes impressions for AI Mode, with views by page, country, date, and device, and an export of the chart and table data. Dates in that report use Pacific Time. Google said the insights had rolled out worldwide as of August 31, 2026, and also says the report may not show up for a site without enough qualifying impressions. The documentation I reviewed describes impressions, not which brands an answer named or which competitors it cited. So use Search Console as a companion signal for your site's overall visibility, and keep sampling the answers yourself for the prompt-level picture.

You're done when every change you claim can be traced to the same prompt group under the same conditions, and someone has looked at the underlying answers and URLs.

Common mistake: comparing a personalized mobile sample from one wave with a different market or a desktop sample from another, and then blaming the difference on your brand's visibility. Check the conditions column first.

On a larger program, the manual version of this gets heavy. DeepSmith's AEO area tracks mention rate and citation rate separately, shows per-prompt results with the full answer history, and has a Pages view that lists which of your pages AI cites and which prompts drive those citations, along with competitor citations by platform. Google AI Mode is one of the ten engines DeepSmith covers, and it's part of the Enterprise and Custom plans, not the lower tiers, so check which plan fits before you plan around it. If you're choosing software, our roundup of the best AI visibility tools lists the options. If you'd like to see how a sheet compares with a tool, our write-up of an AEO tool versus a DIY prompt-tracking spreadsheet is worth a read.

8. Turn what you find into the next test

When the same gap shows up in several waves, use it to form a hypothesis, not a promise. For example: for your comparison prompts, competitors were named more often and their product pages were cited more often than yours. A sensible next move is to look at whether you have a clear, accurate page that answers that buyer question, change it if you don't, and then repeat the same measurement.

Keep the original prompt group and the same sampling conditions, so your Google AI Mode tracking stays steady and the next wave stays comparable. If you add or reword prompts, put them in the separate group so your baseline doesn't shift without anyone noticing.

You're done when each content action you propose is tied to a specific prompt and a specific answer or page gap, and the follow-up measurement is on the calendar with the same protocol. If you want help turning those gaps into a plan, our guide to turning AI visibility gaps into a content backlog picks up where this step ends.

What to do next

Start small this week. Write down your definitions, pick your 20 core prompts, and run the first wave of three samples each under conditions you've recorded. Then put the next wave on the calendar before you close the sheet, because the value of this comes from the second and third waves, when you can see what moved.

If you'd rather have prompt tracking and answer history in one place, you can start a DeepSmith free trial and see how it handles your prompts. The trial runs for seven days, and engine coverage depends on your plan, with Google AI Mode available on Enterprise and Custom.

Frequently asked questions

How do I track my brand in Google AI Mode?

Build a stable list of real buyer questions, run each one several times in fresh AI Mode conversations under recorded conditions, and save the answer and source URLs for each run. Then calculate mention rate and owned-citation rate separately across the same prompt group over time.

Does Google Search Console show brand mentions in AI Mode answers?

Not as far as the documentation I reviewed goes. The generative AI performance report gives impressions with views by page, country, date, and device, and it isn't described as a prompt-level tracker for brand mentions or citations. Use it alongside direct answer sampling, not instead of it.

How many times should I test each prompt?

Google doesn't prescribe a number, and none of the sources I reviewed offered a validated one. A practical starting point is three independent samples per prompt per wave, reporting the raw counts and denominators. If a decision rests on a small or inconsistent result, run more and say how you did it.

Why can two AI Mode samples look different?

Prompt wording, conversation context, timing, device, market, location, and personalization settings can all change an answer. Google also says AI Mode can break a question into subtopics and search for each at once. Record your conditions and compare like with like, and don't treat one answer as a permanent position.