AI brand monitoring means asking ChatGPT, Gemini, Claude and Perplexity the questions your buyers ask, on a fixed schedule, and recording whether your brand is named and how it is described. An AI brand mention is any time an assistant names your business in an answer, most often without a link. Because none of those mentions reach your analytics, monitoring is the only way to know they happened. Each assistant behaves differently, so the useful way to do it is one method applied to all of them, read assistant by assistant.
A mention in a magazine at least produced a spike you could see. A mention inside a private conversation between one person and an assistant produces nothing unless you go looking, and the person who read it may act a week later through a channel that looks unrelated. A business can be recommended many times over and have no idea.
The reverse is just as true and less comfortable. A business can be described inaccurately, or dismissed in a single clause, over and over, and never learn that either. There is no notification and no review to reply to. Somebody asks which of four options to pick, the answer says one of them is fine but known for slow delivery, and the reader quietly chooses differently.
Use one fixed question set across every assistant. Read whether you are present across all of them together, and read whether a change landed one assistant at a time. Watch ChatGPT because your buyers are there, learn on Perplexity because it shows its sources, check Gemini against your Google rankings, and read Claude for how you are qualified.
What an AI brand mention is, and why it leaves no trail
Almost every business judges this by referral traffic, sees little, and concludes nothing is happening. The reasoning is sound and the conclusion is usually wrong.
When an assistant names three brands, the reader rarely clicks. They are mid-conversation. They take the name and act later, through a new tab and a branded search, a direct visit or a marketplace. Every one of those arrives in analytics with no connection to the answer that caused it.
Some assistants pass a referrer when a citation is clicked, so a trickle of identifiable traffic does exist. Treat it as a floor, never as a measure of the whole, because the ratio of readers to clickers is unknown and probably differs by category.
One indirect signal is worth watching over months, with care. If AI recommendations are working, more people should search your brand name directly, since that is the path most of them take. Branded search volume is visible in Search Console and it does move for this reason. It also moves for half a dozen other reasons, so it corroborates a trend you established by asking questions. It is not the measurement itself, and it is far too slow to steer by week to week. Any tool claiming to attribute revenue to AI mentions is modeling, not measuring.
Not every mention is good news
A tally treats every appearance as equal, and they are not.
Being named dismissively counts as present. An answer that says most people choose A or B, though C is cheaper if you do not mind the wait, has mentioned C. A tracker records a mention. A buyer reads a reason not to bother.
Position matters. Named first in a list of three is a different outcome from named fifth in a list of six, and in a short answer attention drops off fast.
Being named in the wrong category is worse than absent. If an assistant consistently describes you as something you are not, every mention sends the wrong buyers. That is a content problem with a clear fix, and you only find it by reading the answers.
Watch the qualifier. Assistants very rarely say a brand is bad. They attach a condition: good if you are on a budget, fine for beginners, worth considering if you can wait. Each reads as neutral in a count and each quietly steers a particular kind of buyer away. Tracking which qualifiers keep appearing tells you what the internet currently believes about your business.
In long answers, record what the mention was doing, not just that it happened. A name can appear in a genuine recommendation, in an aside about the wider market, or in a comparison in somebody else's favor. All three are mentions and only one is good news. Three labels are enough, and the call takes seconds while reading. Skip it and you get a particular kind of bad month: presence looks healthy while your mentions have quietly shifted from recommendation to passing reference.
The habit that catches all of this is simple and rarely done. Once a month, read the full text of five answers rather than only the tally. It takes twenty minutes and it is the difference between knowing you were mentioned and knowing what was said.
Where mentions come from
An assistant is not remembering your brand from an advert. It reads pages and repeats what it finds, so mentions come from wherever your name already appears in a source it trusts. Our research on how LLMs choose their sources covers the mechanics in depth. Three sources carry most of it.
Community threads. A discussion where several people compare options and say what worked is exactly the evidence a retrieval system wants, and it often sits on a domain with very high authority. For many categories this is the largest single source of mentions.
Published comparisons and roundups. Anything shaped like best X for Y gets read constantly, because it matches the question being asked almost word for word. Being in somebody else's list can be worth more than having your own. A single well-placed listing gets read every time somebody asks the question it matches, for as long as it keeps ranking. Most businesses never ask to be included in a list they belong in. The writer would usually rather the page were accurate, so the request is worth making.
Your own pages, but later. They serve as corroboration once the assistant has met you elsewhere, and they carry the specific claims worth quoting. They are rarely the first place you get found. That does not make your site pointless. It makes it the second half of the job, and it explains why sites with good content and no outside presence stay invisible for a long time.
One practice, two levels of reading
Some questions only make sense across all assistants at once, and some only make sense one at a time. Mixing them is why so much monitoring output feels informative and changes nothing.
The aggregate question is whether you are present in your category at all. Being named in one assistant and absent from three is a different position from being named in all four. Over several months, the share of askings that name you across everything is the one number that says whether the program is working, and the only one worth putting in front of anybody senior. Present it with its denominator: named in eleven of forty-five askings can be questioned and trusted. Twenty-four percent is the same fact with the evidence removed.
The per-assistant question is whether a specific change landed. Assistants differ in how quickly they pick things up, so a rewrite can be visible in one within weeks and invisible in another for a quarter. Averaging hides the only feedback you were going to get. Expect the fast movers to show a change first, and give any change at least two readings in the slow ones before deciding it did not work there.
What is meaningless at either level is one blended score across all assistants and all questions. It moves for reasons nobody can name and invites explanations that are invented.
Why assistants disagree about the same business
Two assistants asked the same question can give genuinely different answers. They draw on different material, so a source that carries you in one may be absent in another. They refresh at different rates. Some show sources and some do not, which changes what you can observe rather than what is true. And they compose differently: one may name six options where another names one, so your chance of appearing differs before anything about you is considered.
That last point distorts naive comparisons badly. If one assistant habitually names five businesses and another names one, your presence rate will be higher in the first for reasons that have nothing to do with your standing. What is comparable is each assistant's number against itself over time, not the numbers against each other. That is why one column belongs in every row for every assistant: how many names the answer contained.
Assistant by assistant
The method is the same everywhere. What each assistant is good for is not, and knowing the difference tells you where to spend attention.
| Assistant | Its job in your monitoring | The extra thing to record |
|---|---|---|
| ChatGPT | The scoreboard: where most buyers are | How many names each answer contained, and the exact wording about you |
| Perplexity | The laboratory: it shows which page and passage it used | The full list of cited sources, and the quoted passage |
| Gemini | The test of whether your Google rankings carry over | Your ranking position for the same query, in the same row |
| Claude | The reading on how you are qualified, not just whether you appear | The condition attached to your name, and what the mention was doing |
ChatGPT: the scoreboard
This is the assistant most of your buyers are using, which makes it the one worth watching and a poor instrument for learning what works. If your customers ask about your category anywhere, they ask here, and being absent is a commercial fact regardless of what any other assistant says. But it does not always show where an answer came from, so you can see that you were not named without seeing who was used instead. That is a score without a diagnosis. Watch presence here, and run your experiments somewhere that shows its working.
Because the diagnostic layer is thin, the descriptive layer has to work harder. Ask each question at least five times, each in a fresh conversation, because earlier questions in a thread shape later answers: if question three mentioned your company, question seven may name you for reasons unrelated to your standing. Use a clean account kept only for monitoring, because an account that has researched your own company is primed and produces flattering readings that will not reproduce anywhere else. Both errors make you look better than you are, which is why nobody questions them. Region and language matter too: if you run the checks from one place and your buyers are somewhere else, the readings describe a market you do not sell into.
Phrasing changes the answer more than people expect. Asking for the best providers of a category, asking who to use for a specific problem, and asking how to solve that problem without naming a category will often produce three different sets of names. The third is the phrasing real buyers use most and businesses test least. Spend an hour early on testing three or four phrasings of the same need, pick the set from what you learn, then freeze it. Our walkthrough on how to track brand mentions in ChatGPT covers the hands-on routine.
When a reading shows you absent, diagnose before you write, in this order, because the cheap explanations are the likely ones. First, can anything reach your pages? Answer crawlers are separately named and often blocked by somebody solving an unrelated problem, so a site can be open to search and closed here. Second, are your questions in your vocabulary rather than a buyer's? Third, do the pages you want quoted answer their own question in the first two sentences? Only then consider depth, the expensive answer every supplier is equipped to sell. For the full fix list, see our guide to getting your business on ChatGPT.
Perplexity: the laboratory
Perplexity numbers its sources on every answer, which makes it the one place in this work where a specific edit produces readable feedback. Fewer of your buyers are here than in ChatGPT, so a strong position is worth less commercially. That is why it should be your laboratory rather than your scoreboard. Teams that get value from monitoring tend to land on the same arrangement: watch presence where the buyers are, learn here.
Four things are worth reading in every source list. Which of your pages was used: your homepage usually means nothing deeper was found suitable, a specific article means that page carries the subject, and an old forgotten page means something in it answers well and your newer page is losing to it. Which passage was quoted: the single most actionable fact in this work, because it tells you which of your sentences is doing the job. What was cited alongside you: being cited next to established publications is a different position from being cited next to three competitors. What was cited instead of you: when you are absent, the sources used are a direct answer to what beat you, and you can go and read them.
The experiment is simple and ruined by one common mistake, which is not recording the before. Pick five pages that matter commercially. For each, ask the questions that page should answer and record which source is cited and which passage is quoted, even when nothing of yours appears. Rewrite the opening of each page so the answer to its title question arrives in the first two sentences and stands alone when lifted out. Change nothing else. Wait two to three weeks, read again, and confirm with a second reading a few weeks later.
Watching source lists makes some patterns visible that are guesswork elsewhere. Specific, self-contained statements with a real constraint or trade-off do well. Pages that answer their own title early do well, and the cited passage is very often the first substantive paragraph rather than the best one. Depth shows up as a pattern: a site with one page on a topic gets cited occasionally, while a site with a connected set gets cited across many phrasings. Pages padded to a length or built as long lists of one-line answers do not show up more often. Our guide to Perplexity SEO goes further on what it weighs.
There is a blind spot. Because the feedback is here, you will be tempted to tune for this one assistant. That is fine while the changes are general, and being reachable and quotable mostly are. The test is whether a change would also make the page better for a person. If it only makes sense as a response to observed system behavior, it will age badly. Check presence in the assistants that hide their sources from time to time, and if your position there is not improving while it improves here, you have been tuning to the instrument. Know when to stop, too: once your important pages answer their question up front and their claims stand alone, further wordsmithing buys little, and the remaining gains are in depth and in being referenced elsewhere.
Gemini: does your Google ranking carry over?
Sharing a parent company with a search engine does not mean sharing an index, a selection process or a set of results. Businesses that rank strongly routinely find they are absent from generated answers about their own category, and they find out late because nobody checked. Ranking sorts pages for a person to choose from. Generating an answer assembles a statement from material, and the qualities that put a page at the top of a list are not the ones that make a paragraph worth lifting.
So take the first reading here to test that assumption. Ask your five strongest ranking queries the way a person would phrase them, and record whether you are named. That half hour is the highest-value reading available to a business with existing search strength. Ranking strength does help: being widely referenced and having depth on a subject support both channels. What it does not buy is passage-level suitability, whether your page states its conclusion early and says something specific enough to quote. Our comparison of AEO and SEO works through where the two overlap and where they split.
The column that makes this assistant worth watching is your ranking position for the same query, recorded in the same row on the same day. The interesting cases are queries where you rank well and are never named. They isolate the writing problem from access and depth, which makes them the cleanest test bed on your site. When the numbers move, the pattern is unusually readable.
- Presence rises, ranking flat: your pages became better material. Whatever you changed can be repeated across the site.
- Both rise: something underneath both, usually depth or genuine references from elsewhere.
- Ranking flat, presence falls: check access first. A rule that blocks an answer crawler will not touch your ranking at all.
- Both fall: probably the wider results rather than you. Check whether competitors show the same pattern.
One practical point: also record what appears alongside the generated answer. Being absent from the text while present just below it is a weaker position than being named and a much stronger one than being nowhere. And decide how you will present a finding of absence before you take the reading. A team with years of ranking success will hear it as an attack on that record unless it is framed as a gap in a new channel, one their ranking strength makes easier to close.
Claude: how you are qualified
An assistant that works through a question before answering tends to name fewer things and attach more conditions to each. That changes what presence means. When an answer names six options, being one of them clears a fairly low bar. When it names two and explains why, being one of them is a much stronger signal, and being absent is a much weaker one, because the answer may simply have chosen to be narrow.
Three adjustments make readings here usable. Record how many names each answer contained, or your presence rate cannot be compared with anywhere else. Ask each question more times than elsewhere, because fewer names per answer means each reading samples a smaller window. And keep a short verbatim note of the condition attached to your name rather than a category. Named as the right choice for a specific situation, named as one of several with no reason given, and named as suitable but with a caveat about scale or price are three different commercial positions that all count as one mention.
A rule that saves anxious meetings: only treat absence as evidence when the answer named at least three things and at least one was a business like yours. If the answer named publications or community discussion rather than businesses, it was never comparing suppliers. If the question asked how to solve a problem and the answer gave an approach with no names, that is not a loss, though it tells you the realistic goal in your category is being the source of the approach.
The most actionable output here is a caveat that keeps recurring. If the same reservation about scale, price, breadth or maturity attaches to your name month after month, it is being drawn from real material, and finding the source usually has one of three answers. Sometimes it is a hedge on your own site, written years ago for good reasons and now lifted as a limitation, and the fix is an edit. Sometimes it is third-party discussion, which is a reputational fact you can address with something specific or by changing the thing being complained about. And sometimes it is simply true, in which case the useful response is knowing it is being said in front of your buyers.
If you can only watch two assistants, Claude is probably not one of them. As a third, it is strong, because it answers a different question from the other two rather than confirming them. The exception is a business selling something complicated, where the buying decision involves weighing conditions. There, being the option that survives the reasoning is worth more than being one name among six.
Setting up your monitoring
How many assistants. Two is the practical minimum. With one, you cannot tell a general change from something particular to that assistant, and every conclusion will be about the assistant rather than your business. Three or four covers the realistic range without a reading becoming a morning's work. Include at least one that shows its sources and one that does not. Cap it deliberately, because the arithmetic compounds: ten questions asked five times across six assistants is three hundred askings a month, which stops being an hour and becomes a job. If your buyers overwhelmingly use one assistant, watch that one closely and treat the others as a sanity check.
The question set. Six to ten questions in buyer language is enough to start, drawn from support emails, sales calls and reviews rather than your own vocabulary. Use the same set in every assistant, asked the same number of times, or you lose the aggregate view. Freeze the wording and copy it from one stored list every time, since paraphrasing between readings quietly breaks comparability. Write down why each question was chosen: a question whose purpose has been forgotten is the first one somebody edits when it produces an uncomfortable answer. Over the long run, run two sets: the original, frozen however dated it looks, and a rolling set rebuilt each year from the language customers used that year. The gap between them shows how fast your category's vocabulary moves.
What to record, every row. The assistant, the date, the question verbatim, the share of askings that named you, how many names the answer contained, every other name, the cited sources where shown, the wording about you and what your mention was doing. Then the per-assistant extras from the table above. Keep all assistants in one table so the aggregate view still works, and keep the full answer text for at least the first few months, because you will later want to reread it with a better idea of what matters.
How often. Monthly. Daily and weekly readings are dominated by run-to-run variation and train a team to react to noise. Quarterly readings are too far from any change to attribute it. Read more often only for a few weeks around a discrete fix such as an access rule. When a number moves, ask four things: did the question set change, is the movement spread across several questions, is there a plausible cause in the right window, and does it persist into the next reading. Our guide to AI visibility tools covers when it is worth paying for software to run this instead of a spreadsheet.
Write down the method. One page beside the data: the questions, the number of askings, the assistants in scope, the day of the month, and who runs it. The person who set it up will not be the person reading it in a year, and an undocumented method quietly becomes a different method.
What to do with what you find
Three findings call for three different responses.
Named nowhere. Start with crawler access, since it is quick and binary. Then check whether you appear in ordinary search results for the questions being asked, because if you do not, no assistant that searches the web is reading you either.
Named occasionally, described vaguely. Your pages are probably being read but contain nothing specific enough to repeat. The fix is on the page and fast: lead with the claim, make it checkable, and answer the question the page is named for in its first paragraph.
Named consistently, described wrongly. The assistant has formed its picture of your category from sources that do not describe you accurately. The instinct is to write a page correcting the record, stating that you are not the budget option. That rarely works, because a page arguing against a description still contains the description, and it competes with a dozen sources that say otherwise. What works better is producing specific, checkable claims that make the accurate description easy to assemble, and getting them into places that already get read. Argue less, show more evidence. It is slower and it is the version that holds.
Monitoring also has limits worth stating, so nobody overclaims. It cannot tell you why you were or were not named: the reasoning is not visible, and any confident account of it is reconstruction. It cannot tell you what a mention was worth. And it cannot tell you your question set is wrong, which is why the rolling second set exists. Writing down what you expect before each reading helps too. Most teams are wrong in the same direction, assuming they are better known than the readings show.
This practice will look out of proportion for the first two quarters, because you are building a measuring habit for a channel that may not move for months. The businesses that can later show whether their work paid off are the ones that started collecting before they had anything to show.
How RunOctopus handles this
Monitoring tells you what assistants say about you. Changing what they say takes pages worth quoting, published where they can be found, and that is the work RunOctopus does.
RunOctopus is a done-for-you service. Its AI, Ollie, researches your business, then writes and publishes pages directly on your own website. A team plans, writes, reviews and publishes alongside Ollie. It works with Shopify, WordPress, WooCommerce and Wix.
There is nothing to install, nothing to approve, nothing to upload and nothing to schedule. Pages go live on their own. In the first week you get 50+ pieces, including long-form articles, pages and a custom interactive tool built for your niche. After that it is 175+ pieces of content a month, written, published and linked. You also get reports on where ChatGPT, Perplexity, Claude and Gemini name your business.
There is a 90-day promise: more visibility on Google and in AI answers within 90 days, measured against the 90 days before you joined, or every dollar back. Paste your address into the free site read and Ollie reads the site and shows the plan. No account, no call.