How to track your brand in Google AI Mode

At I/O 2026 Google rebuilt the search bar itself around Gemini 3.5 Flash and called it the biggest change to the Search box in more than 25 years. The new bar expands to fit full conversational questions, accepts images, files, videos and open Chrome tabs as input, and hands the heavy queries to AI Mode, the conversational search experience Google has been scaling since 2025. The user numbers explain the urgency: AI Mode passed one billion monthly users by I/O 2026, roughly ten times its audience from six months earlier, and Google reports AI Mode queries more than doubling every quarter.

For a brand, this shifts where the buying decision happens. A classic results page distributes attention across ten links. An AI Mode answer composes a verdict, names a handful of brands, cites a handful of sources, and moves on. If your category's buying questions now run through that interface and your brand does not appear in the composed answer, you are invisible for that conversation regardless of where you rank underneath it. The background on what the feature is and how it behaves lives in our Google AI Mode explainer; this article is about measuring your place in it.

One honest calibration before the method: Google has said on the record in 2026 that AI Mode is not the default way it serves results. Classic results and AI Overviews still carry most query volume. But the adjacent numbers are hard to ignore. Ahrefs measured AI Overviews on 48 percent of queries by March 2026, up from 34.5 percent in December 2025, and Pew Research found users click a traditional result on only 8 percent of searches when an AI summary is present, versus 15 percent without one. The answer layer is absorbing attention faster than most dashboards are absorbing the answer layer.

Rank trackers were built for a deterministic surface: query in, ordered list out, positions stable enough to check once a day. AI Mode breaks every one of those assumptions. When a user asks it a question, the system runs a query fan-out, decomposing the question into parallel sub-queries, often eight to sixteen of them, retrieving results for each, and synthesizing one answer from the combined pool. The answer that comes back is a paragraph with citations, and the brands it names may come from sub-queries no rank tracker ever monitored.

The citation data shows how far the answer has drifted from the ranking. Analyses of Google's AI citations found that the share coming from top-10 organic results fell from 76 percent in July 2025 to 38 percent by March 2026. Pages ranked seventh get cited while pages ranked first get skipped, because the citation went to whichever passage best answered a sub-query. Your position for the visible keyword has become an input among many rather than the score.

Then there is volatility. Generative answers are probabilistic: the same question, asked twice, can produce different brand lists. SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brands, and Gemini-family answers behave the same way. AI Mode adds personalization from search history and location on top. A tracking method built for this surface needs repeated sampling and stored answers, the same discipline GEO measurement demands everywhere else.

In June 2026 Google shipped the first official window into this surface: Search Generative AI performance reports in Search Console, one for Search and one for Discover. The Search report counts impressions your pages earn inside generative AI features, meaning AI Overviews and AI Mode together, and lets you slice those impressions by page, country and device. For the first time you can watch your AI-surface footprint trend over time with Google's own numbers, which is genuinely useful for spotting whether the answer layer is growing or shrinking for your site.

The gaps matter as much as the data. The report shows no clicks, no click-through rate and no query strings, so you cannot see which questions produced the impressions or what the user did next. It does not separate AI Mode from AI Overviews, so you cannot tell a one-shot summary from a multi-turn conversation. And most importantly, an impression only means a link to your page appeared somewhere in an AI response. It says nothing about whether the answer mentioned your brand, recommended it, or named three competitors instead. We break the whole report down in Google Search Console's new AI performance reports, explained.

There is also an access caveat: at launch the report rolled out to a subset of site owners, starting in the UK, with wider availability promised but not dated. If you do not see it in your Search Console yet, that is expected. Either way, treat it as the impression layer of a two-layer measurement stack. The answer layer, what the AI actually said, still has to come from somewhere else.

Tracking starts with the questions, because in AI Mode the question is the unit of competition. Write down 20 to 30 questions a real buyer in your category would type into a conversational search bar: best-tool questions, comparison questions, problem questions, alternative questions. Phrase them the way people talk to an assistant, full sentences with context, because that is what the expanded search bar was rebuilt to invite. Your existing keyword list is raw material here, but it needs translating from keyword grammar into question grammar.

Two rules keep the list honest. First, freeze it. A question list that changes every week produces a trend line that means nothing; you want the same questions scored the same way month after month. Second, separate branded questions from unbranded ones. If the question contains your brand name, the answer will mention you by construction, and mixing those runs into your headline score inflates it. Reachroller excludes branded questions from its headline visibility number for exactly this reason, a choice documented on our methodology page.

Size the list to your capacity to act. Twenty-five questions tracked well, with a content fix shipped for the worst one each week, beats two hundred questions tracked passively. The point of the list is a work queue, and a work queue you never work through is a dashboard.

For each question, the check itself has to be hygienic. Use a logged-out session or a clean profile so personalization does not contaminate the sample, hold location constant, and record everything: the full answer text, every brand named, every source cited, and the date. A screenshot folder works at small scale. A spreadsheet with one row per run works a little longer. The discipline that does not scale manually is repetition, and repetition is the part that makes the numbers mean anything.

Because answers vary run to run, a single check per question is a coin flip with a report attached. You need multiple runs per question per period, and then a rate: out of N runs this week, your brand appeared in K answers, so your mention rate for that question is K over N. Do that across the whole list and you get a portfolio-level visibility score you can trend. The week-over-week movement of that score, per question and overall, is the entire game. What the score means and how to build it from scratch is covered in what is AI visibility.

Store the raw answers, always. Six weeks from now, when a question flips from naming you to skipping you, the stored answers are the audit trail that tells you what changed: a new competitor entered the citations, a source you relied on dropped out, or the engine started favoring a different page type. A score without the answers behind it is a number you have to take on faith.

A mention rate tells you where you stand. Citations tell you what to do about it. Every AI Mode answer cites sources, and those sources are the pages that made the engine say what it said. When you lose a question, the cited pages are the reason: a listicle that ranks you sixth, a review thread that never mentions you, a comparison page a competitor published and you did not. Logging citations per question turns a visibility problem into a to-do list.

The citation patterns are studiable. Cross-engine analyses of hundreds of millions of AI citations found roughly 63 percent point to listicle-style pages, and pages that rank for the fan-out sub-queries are dramatically more likely to be cited than pages that only rank for the visible query. So when your citation log shows the same third-party listicle deciding four of your questions, you know precisely which away game to play. The full playbook for converting citation intelligence into presence is in how to get your brand into Google AI Mode answers.

Everything above can be done by hand, and for a first audit it should be: an afternoon of manual AI Mode checks against your question list will teach you more about your category than a quarter of rank reports. What does not survive contact with a real workload is the repetition. Multiple runs per question, per engine, per week, with stored answers and citation logs, is exactly the kind of loop software should run while you do the part that needs judgment: deciding which lost question to fix next.

Reachroller is that loop as a product. You give it your domain and it builds the buying-question list, runs the checks on schedule, scores mention rates with branded questions excluded, and stores every raw answer and citation so each score has a receipt. When a question is lost, it drafts the fix: a publish-ready, answer-shaped page with slug, title tag, meta description and schema markup, then rechecks whether the answer flipped. Starter is $29 per month for 400 credits and 25 tracked questions, where one credit is one AI answer and a generated fix costs ten.

The honest caveats: Reachroller is a young product, ChatGPT tracking is live today, and Gemini, Claude, Perplexity and Grok tracking are built and rolling out. It does not scrape Google's AI Mode interface directly; Gemini tracking is the closest answer-level signal for how Google's models treat your brand, and Search Console's new report supplies the Google-side impression layer to pair with it. Between the two you get what neither gives alone: impressions from Google, answers and citations from the engines, and a trend line that tells you whether the fixes are working. How that compares to other tools in the space is covered in the best AI visibility tools.