Picture your best buyer, mid-decision, typing your category into ChatGPT or Gemini and asking which options they should consider. The model answers with a short list of names. You have no idea whether yours is on it. You cannot see this moment, it leaves no entry in your analytics, and it may be happening hundreds of times a day while you optimize things you can actually measure. That blind spot is the problem, and closing it is the first real step toward improving your presence in AI answers, because you cannot fix what you never see.
Tracking your brand in AI search is different from tracking Google rankings, and pretending otherwise leads to frustration. There is no fixed position to look up, because the answers vary between runs, users, and phrasings. What you can do is measure presence as a rate: how often, across a stable set of the questions your buyers ask, the models name you. Set that measurement up once and run it on a schedule, and the invisible moment becomes a number you can watch move.
You cannot improve what you never measure
Every other channel you invest in has a feedback loop. Search has rankings and clicks, email has opens, ads have impressions. AI search, for most brands, has nothing, which means effort there is flying blind: you publish, you earn coverage, and you have no idea whether any of it changed whether a model names you. That absence of feedback is why AI-search work so often stalls, because without a signal, you cannot tell what is working, and without knowing what is working, you cannot do more of it.
The fix is to build the feedback loop yourself, because no dashboard hands it to you by default. It does not have to be sophisticated. It has to be consistent: the same questions, asked the same way, on the same schedule, with the results recorded so you can compare over time. The moment you have that, AI search stops being a black box and becomes a channel you can manage like any other, with a baseline, a trend, and evidence of whether your work is moving the number. Measurement is not the glamorous part of this, but it is the part that makes everything after it possible.
Share of Model: the metric that matters
The metric to track is what you might call Share of Model: across the set of buyer questions where a model could name you, how often it actually does. If you ask twenty relevant questions and the model names you in six, your Share of Model is roughly thirty percent for that set. It is a rate, not a rank, and that is the point, because a rate captures how AI answers actually behave: probabilistic, varying, but measurable in aggregate. Watch that rate over months and you have a real signal of whether your AI presence is growing.
Share of Model works because it mirrors how buyers actually encounter you through AI. They are not checking a ranking, they are asking questions and seeing which brands come up, so the honest measure of your visibility is the fraction of those moments you are present in. It also gives you something to move: a baseline percentage today, a target percentage later, and a way to tell whether the coverage you earned or the pages you rewrote actually raised it. A single vague sense of doing better in AI becomes a concrete number you can hold your work accountable to.
Build a prompt panel and run it on a schedule
Start by building a prompt panel: a fixed list of the real questions your buyers ask when they are looking for what you offer. Write them the way a human would type them, cover the range from broad category questions to specific comparison and recommendation questions, and keep the list stable so your measurements stay comparable over time. Twenty to forty well-chosen prompts is plenty for most brands. The panel is the instrument, and like any instrument, its value comes from being consistent, so resist the urge to constantly change the questions.
Then run the panel on a schedule across the engines your buyers use, and record what happens. Ask each question, note whether the model named you, and do this the same way each cycle so the comparison holds. Monthly is a sensible cadence for most brands, frequent enough to catch real movement and slow enough to avoid chasing the natural noise between individual runs. The discipline is in the repetition: the same prompts, the same engines, the same schedule, recorded the same way, so that the trend line you build actually means something instead of reflecting random variation.
Write prompts the way buyers actually ask
Your panel is only as good as its prompts, and the most common mistake is writing them like a marketer instead of a buyer. Buyers do not type your brand name and ask how great you are. They ask about their problem, their category, and their options, often without naming any brand at all. So your prompts should mirror that: the unbranded category question, the best-option-for-a-situation question, the comparison question, the recommendation question. Those are the moments where a model either names you unprompted or does not, which is the visibility that actually matters, because it happens before the buyer knows you exist.
Include a spread of intents so the panel captures your presence across the buyer’s path, not just one slice of it. Broad category questions test whether you are in the consideration set at all. Specific, situation-based questions test whether you are the recommended fit for a particular need. Comparison questions test how you stand against named rivals. A panel weighted toward branded questions flatters you, because a model asked about your brand will usually say something about it, while the honest test is whether you come up when the buyer never mentioned you. Write the prompts your buyers would actually type, and your Share of Model measures real visibility instead of a comfortable illusion.
What to record beyond a yes or no
A simple named-or-not tally is a good start, but the richer signal is in the details around each mention. Record how you were described, because the framing the model uses tells you what it believes about you and whether that matches the position you want. Note which competitors appear alongside you, because that maps your real competitive set in the model’s eyes, which is often not the set you assumed. And when the answer came from a live search, note which sources it cited, because those citations show you exactly which pages and outlets are feeding your presence.
These details turn tracking from a scoreboard into a diagnostic. If the model consistently describes you in a way you do not like, that is a nameability problem to fix in your content and coverage. If a competitor keeps appearing where you do not, that is a specific gap to close. If certain sources keep getting cited, those are the pages and publications worth reinforcing. The yes-or-no rate tells you whether you are winning. The surrounding detail tells you why, and why is what you act on. Capturing both is the difference between knowing your score and knowing your next move.
Track across engines, not just one
Your buyers do not all use the same AI tool, so tracking one engine gives you a partial and possibly misleading picture. ChatGPT, Gemini, Claude, and the AI answers inside Google each draw on different training and different retrieval, which means your Share of Model can be strong in one and weak in another for the very same question. A brand that is named consistently by one engine and skipped by another has a specific, addressable gap, but you will never see it if your panel only ever asks one tool. Run the same prompts across the engines your buyers actually use, and record the results separately so you can compare.
The differences you find are diagnostic, not noise. If you show up well where an engine leans on live search but poorly where it answers from training memory, that tells you your retrievable content is strong and your broad, learned presence is weak, which points to different work. If one engine cites specific sources and another does not, the citations from the first show you which pages are carrying you. Treating each engine as its own measurement turns a single vague visibility number into a map of where you are strong, where you are absent, and why. That map is worth far more than a single blended score that hides the gaps inside it.
Watch the trend, not the single run
The biggest mistake in AI-search tracking is overreacting to one run, because individual answers vary and a single absence proves nothing. Ask the same question twice and you may get named once and skipped once, not because anything changed but because that is how these systems behave. If you treat every run as a verdict, you will chase noise, celebrate flukes, and change your strategy based on randomness. The signal lives in the aggregate: your Share of Model across the whole panel, watched over months, not any single answer on any single day.
So build the discipline to read the trend and ignore the jitter. One month’s number is a data point, not a conclusion. Three or four months of the same panel, run the same way, reveal the real direction: whether your presence is climbing, flat, or slipping, and whether the work you did between measurements moved it. This is why consistency of method matters more than frequency of checking. A stable panel run monthly and read as a trend tells you the truth. A panel you obsessively re-run and reinterpret every few days tells you mostly about variance. Patience with the measurement is what turns tracking from anxiety into insight.
Turn tracking into action
Tracking is only worth the effort if it changes what you do, so close the loop deliberately. Each cycle, look at the questions where you were absent and treat them as your targets: write the clean answer-first content that resolves them, earn the outside corroboration that makes you safe to cite, and sharpen the description you want the model to repeat. Then watch the next cycle to see whether your Share of Model on those questions moved. That is the entire management loop, measure, act on the gaps, measure again, and it is exactly the loop AI search has been missing.
Over a few cycles this compounds into something no competitor flying blind can match: a clear picture of where you stand, evidence of what moves the number, and a prioritized list of the questions worth working on next. To track your brand in AI search well, you do not need expensive tooling, you need a stable prompt panel, a consistent schedule, and the discipline to act on what it shows you. Build your panel this week, run it once to set a baseline, and you will know more about your AI presence than nearly everyone you compete with.