A one-time check tells you what an AI assistant said once. Tracking tells you what it keeps saying, and where that story is drifting.
Daniel Arons · Jun 2026 · 7 min read
You typed your brand into ChatGPT, read the answer, and felt either relieved or alarmed. That is a useful first look. It is also one of the least reliable signals you can collect, because the answer you saw was a single draw from a system that changes its mind.
Tracking your brand's visibility in AI search is a different exercise. It is not a snapshot you take when you are nervous. It is a measurement habit you run on a schedule, so you can see how AI assistants describe and recommend you, how that shifts week to week, and whether your work is moving the needle. This piece is about that habit, and how to build it without fooling yourself.
Why a single check is misleading
AI assistants do not return the same answer every time. Ask the same question twice and the wording, the brands named, and the order they appear in can all change. Sampling is built into how these models generate text, so two sessions an hour apart are two different data points, not one truth.
The ground underneath the models also moves. Assistants pull from the live web, from their training, and increasingly from retrieval at the moment you ask. A new review, a competitor's fresh page, or a quiet model update can change what gets surfaced without any warning to you.
So when you check once and see your name, you have learned that you can appear, not that you reliably do. And when you check once and miss, you have not learned that you are invisible. You have learned that this particular roll of the dice left you out.
The only way through this noise is repetition. You run the same questions many times, across days, and you look at how often you show up rather than whether you showed up on the one afternoon you happened to look. If you want the conceptual grounding first, our guide to Contextual AI Presence Mapping© explains why presence is a distribution, not a yes or no.
What to actually track over time
Visibility is not one number. If you reduce it to a single score you will miss the parts that matter to revenue. Track a small set of distinct signals, each answering a different question about how AI talks about you.
A fixed, representative prompt set
Everything starts here. Pick a set of questions real buyers bring to AI, the practical ones tied to actual decisions and life moments, not just 'what is [your brand]'. Include category questions ('best tool for X'), comparison questions, and problem-first questions where your name is one of several plausible answers.
Then freeze the set. The value of tracking comes from asking the same things over and over, so a change in the answer reflects a change in the world rather than a change in your wording. Add new prompts deliberately and label them, but do not quietly edit the existing ones, or you lose your ability to compare across time.
Your share of recommendations
For category and comparison prompts, the question is not just 'were you named' but 'how often were you named relative to the competitors named alongside you'. Track who shows up in your space, how frequently each appears, and where you sit in that pack. This is where you see a rival climbing or a new entrant the model has started to trust.
Accuracy and sentiment of the description
Being named is not the same as being described well. Track what the assistant actually says about you. Is the description accurate, or is it repeating a stale claim, the wrong category, a feature you dropped, or a price that is no longer real?
Sentiment matters too. A grudging mention buried in caveats does different work than a confident recommendation. Watch the framing, because a model that names you while hedging is a problem you can fix before it costs you a deal.
The sources driving your mentions
When an assistant cites or leans on specific pages, note them. Over time you will see which sources the models trust when they talk about your category: review sites, your own pages, documentation, forums, press. Those sources are the levers. If you know a particular page is feeding your mentions, you know where to invest, and you can watch what happens when that source changes.
“Being named is the floor. Being described accurately, recommended confidently, and backed by sources you can influence is the ceiling you are actually tracking toward.”
Track per model and per buyer segment
Do not blend everything into one feed. ChatGPT, Claude, Google's Gemini and AI Overviews, and Perplexity draw on different data and behave differently, so they will not tell the same story about you. A page that lifts you in one can be invisible in another.
Run your prompt set against each assistant separately and keep the results separate. You may find you are strong in one and absent in another, which is a far more actionable finding than a single averaged number that hides both.
Segment by buyer too. The questions a first-time shopper asks are not the questions a technical evaluator or a procurement lead asks, and the same model can recommend you to one and skip you for another. Splitting your prompt set by who is asking shows you exactly which audience the AI is failing to connect you with. For the mechanics of this kind of mapping, see how Aethon works.
Set a baseline and watch for drift
Your first few full runs are not a verdict. They are your baseline, the reference point everything later gets measured against. Capture it carefully, because without it you cannot tell whether a later result is good, bad, or just normal variation.
Once you have a baseline, you are watching for drift. Drift is the slow, easy-to-miss change: a description that gets a little less accurate each month, a competitor that creeps up your share, a source that quietly stops feeding your mentions. None of these announce themselves in a single check. They only show up against a stable reference.
Decide in advance what counts as a meaningful move versus noise. Because answers vary run to run, a single dip is not a trend. A sustained shift across repeated runs, on the same frozen prompts, is the signal worth acting on.
“Without a baseline you are not tracking visibility, you are just reacting to whatever the model happened to say the last time you looked.”
Choosing a sensible cadence
Cadence is a balance. Check too rarely and you discover a drop weeks after it cost you. Check obsessively and you drown in the run-to-run noise we already discussed, mistaking normal variation for movement.
For most brands a regular weekly or biweekly run of the full prompt set is enough to catch real change while staying above the noise. Tie extra runs to events you control: a major content push, a website overhaul, a competitor's launch, or a known model update. Those are the moments when you genuinely want a fresh reading.
Whatever you choose, keep it consistent. The discipline matters more than the frequency, because comparable runs at a steady interval are what let you see a trend at all. If you are still scoping the work, our walkthrough on how to audit AI visibility is a good companion to this ongoing approach.
Spreadsheet or tooling
You can start by hand, and there is real value in doing so. A spreadsheet with your frozen prompts, the assistant used, the date, whether you were named, who else was named, and a note on the description teaches you what good tracking even looks like. For a short list of prompts and one or two models, it is a fine place to begin.
The manual approach breaks down on the thing that makes tracking work: repetition at scale. To smooth out sampling noise you need many runs of each prompt, across several assistants, on a steady schedule, with the answers parsed for mentions, share, sentiment, and sources. Done by hand that is hours of copy-paste every cycle, and the moment you get busy, the cadence slips and the baseline goes stale.
This is the honest tradeoff. Manual tracking is cheap to start and expensive to sustain. Tooling carries the repetition, the parsing, and the trend tracking for you, so the habit survives a busy week. If you are weighing options, our rundown of the best AI visibility tool lays out what to look for so the discipline does not depend on your spare time.
Tracking your brand in AI search is a discipline, not a one-off. The brands that get this right are not the ones who checked once and panicked. They are the ones who froze a representative prompt set, set a baseline, ran it on a steady cadence across every major assistant, and acted on real drift instead of single readings. If you would rather see that running on your brand than build it from scratch, you can book a demo and watch the whole picture come together.
Frequently asked questions
How is tracking different from just checking my brand in ChatGPT once?
A single check is one random draw from a system that varies its answers and pulls from a web that keeps changing. Tracking means running the same questions repeatedly over time, so you measure how often you appear rather than whether you appeared on one occasion. That repetition is what separates a real trend from noise.
How often should I track my AI visibility?
For most brands a weekly or biweekly run of your full prompt set catches real change while staying above the run-to-run variation. Add extra runs around events you control, like a content push, a site overhaul, or a competitor launch. The key is keeping the interval consistent so your results stay comparable.
What should I actually measure, beyond whether I was named?
Track several distinct signals: how often you appear, your share of recommendations against named competitors, the accuracy and sentiment of how you are described, and which sources are driving your mentions. Each answers a different question about how AI talks about you. Reducing it to one score hides the parts that affect real buying decisions.
Why do I need to track each AI assistant separately?
ChatGPT, Claude, Gemini, AI Overviews, and Perplexity draw on different data and behave differently, so they will not describe you the same way. You can be strong in one and absent in another. Keeping their results separate gives you findings you can act on instead of an average that hides both extremes.
Can I do this with a spreadsheet instead of a tool?
Yes, and it is a good way to learn what to look for. A spreadsheet with frozen prompts, the model used, the date, and the result works well for a short list and one or two assistants. It breaks down when you need many repeated runs across several models on a steady schedule, which is the point at which tooling keeps the habit alive.
Written by
Daniel Arons
Co-founder & CEO, Aethon AI
Daniel co-founded Aethon AI in November 2025 to close the gap between how marketers measure AI visibility and what AI is actually doing with their brands. Before Aethon, he spent eight years building digital marketing programs in New York across SaaS, financial services, and consumer brands. He holds an MPA from Baruch College and a BA in Public Relations from SUNY Oswego.