DeepSmith

Nov 26 · AEO & AI Visibility

19 min read

What Is an AI Visibility Tracker and How Does It Work?

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Abstract monochrome illustration of layered answer cards linked by lines to nodes and a bar chart, around the cover line Tracking Your AI Visibility.

An AI visibility tracker is a tool that repeatedly checks AI-generated answers to a defined set of questions and measures whether those answers name a brand, cite its pages, or feature competitors. It turns individual answers into a record of observed visibility over time, not a census of everything people see in AI search.

If you have typed a question into ChatGPT or Perplexity and seen a competitor named where you expected to be, you probably wanted a way to check that more than once. That is what an AI visibility tool does. This guide walks through what it measures, how a pile of answers turns into the numbers on a dashboard, and what those numbers can and cannot tell you.

What does an AI visibility tracker measure?

It measures two main things, and they are different from each other. A mention means the answer names your brand. A citation means the answer presents one of your pages as a source. Neither one requires the other.

That is the idea that makes the rest of the dashboard easier to read, so it is worth slowing down on. People often say "visible in AI" as if it were one thing. In practice an answer can do four different things with your brand. Here they are, using Northstar, a made-up customer support software company, as the example.

What the answer didExampleWhat it tells you
Mentioned and citedNames Northstar and links a Northstar page as a sourceBrand presence and source attribution both happened
Mentioned, not citedNames Northstar but links no Northstar pageThe brand appeared, but no page of yours was credited
Cited, not mentionedLinks a Northstar research page for a fact but never says the nameA page was used as a source without the brand being named
NeitherNames no Northstar and cites no Northstar pageNeither signal showed up in that answer

A few cautions go with this table. Being mentioned is not the same as being recommended, because an answer can name a product while warning that it is a poor fit. Being cited is not an endorsement either, since your page might support a statistic while the answer recommends somebody else.

Not every link in a source panel is a citation in the strict sense. OpenAI describes the Sources view in ChatGPT as containing cited sources and other relevant links, and Ahrefs documentation separates sources cited in an answer from pages that were found but not cited. So each tracker has to decide what counts, and that decision shapes its numbers. When you look at a tool, one of the first things to find is how it defines a citation.

Many trackers add more views, such as which competitors appear, which of your pages get cited, how the answer describes the brand, and how all of it changes over time.

How is this different from rank tracking?

The familiar search question is where a page appears for a query. An AI visibility tracker asks a wider set of questions. Does the answer name the brand? Does it link to a page as a source? Does it describe the brand accurately? Does it recommend a competitor instead?

Some tools report an average position, but that number needs its own definition. It might describe where your brand falls in an answer, or where a source falls in a citation list. Neither should be read as a classic search ranking.

How does AI visibility tracking work?

The short version is a chain with five links. A defined set of questions is asked repeatedly, each answer is recorded, each answer is classified for mentions and citations, the results are added up, and the totals link back to the answers behind them. If you are wondering what is AI visibility tracking in one sentence, it is the ongoing measurement of a brand's observed presence in AI-generated answers.

The questions set the scope

Everything the tracker reports comes through the questions you give it. They might cover customer problems, category exploration, comparisons, or questions about your own company. For Northstar, three questions could look like this:

  • "What software helps a small team manage customer support?"
  • "Which support tools work well for an online retailer?"
  • "What does Northstar support software do?"

These do not measure the same thing. The first two show whether the brand turns up when someone is exploring the category and has not named anyone yet. The third puts the brand name in the question, so it is better for checking how the brand gets described. If branded and unbranded questions are mixed together and not labeled, the headline number gets hard to interpret. A high mention rate on questions that already contain your name is not the same as being found by people who have never heard of you.

The same questions are asked again and again

One observation is a question, a platform, a moment in time, and whatever conditions applied when it ran. Repeating the question gives you more observations. It does not give you more customers.

The methodology page from SeenAndCited gives a handy name for this. It calls one monitored question asked of one engine once a probe, and it points out that five questions across two engines make ten probes before you repeat anything. Repetition matters because the same question can produce different brands and sources from one run to the next. It also leaves you with a history of answers, not one screenshot.

The sources behind this guide support repeated measurement, but they do not point to one ideal schedule, so the right rhythm depends on your situation.

Each answer gets classified

For every answer, the tracker works out whether it names the brand and whether it cites any owned pages. It can also note which competing brands appear and which domains and pages are cited.

The counting rules matter here. A tracker has to settle questions like these:

  • Does a brand named five times in one answer count once or five times?
  • Which names, aliases, and product names count as your brand?
  • Which domains belong to you?
  • Is a given link a cited source or just a relevant link?
  • Does the metric count pages, domains, or answers that contain at least one citation?

The research on repeated measurement also points out a weakness of simple name matching against a fixed list of words. Synonyms, abbreviations, and paraphrases can slip past it. So the way a tool recognizes your brand is part of how much you can trust the result.

Valid answers get added up

Counts can be grouped by question, platform, period, topic, brand, domain, or cited page, depending on the product. A careful tool separates a valid answer with no mention, a valid answer with no owned citation, and an attempted check that failed. SeenAndCited, for example, leaves unavailable observations out of its rates and reports measurement coverage separately. That is one documented approach, not a rule for every tracker.

The totals connect back to answers

A headline percentage is a summary. What makes it usable is being able to open the answer history and the page-level detail behind it. When a number moves, you want to ask which questions produced the mentions, which platform produced them, what the answer actually said, which exact pages were cited, and whether those pages belonged to you, a competitor, or a third party.

An AI visibility tool that shows only a score and hides the answers is asking you to take the number on trust. One that links the score to the answers lets you check it. DeepSmith, for example, lets teams inspect the answers behind mention rate, citation rate, and share of voice, including which pages receive citations.

What do the dashboard numbers mean?

Every percentage is a count divided by something, and the something is called the denominator. Before you read any score, find out what was counted and what it was divided by. If you want a wider view of how the whole category fits together, our AI visibility analytics explainer covers the kinds of metrics and reports you will run into.

To make this concrete, here is a teaching model. It is a simple, response-based way to calculate the two main rates. It is not a formula that every product uses, and the numbers below are invented for the example. They are not DeepSmith results, customer results, or industry benchmarks.

Let N be the valid answers included in the scope you are looking at. Let M be the number of those answers that name the brand at least once. Let C be the number that cite at least one owned page. Then:

  • Mention rate is M divided by N, times 100.
  • Owned-page citation rate is C divided by N, times 100.

In this model, a brand named five times in one answer still counts as one answer with the brand present. Several owned citations in one answer still count as one answer with an owned citation. Ahrefs documents this kind of deduplication for brand mentions, page citations, and domain citations. Other products organize their main rates around measured questions rather than pooled answer runs. SeenAndCited, for instance, defines Citation Rate as the share of measured monitored questions where the site was cited. The point is to avoid swapping one denominator for another without noticing.

A worked example

Say a tracker watches ten questions on two platforms, and each question-platform pair gets five observations in the reporting period. That plans for 10 times 2 times 5, which is 100 observations. Ten of them fail or come back unavailable, so 90 are valid. Measurement coverage is 90 divided by 100, or 90%. In this example, failed observations are left out of the rate denominators and shown separately.

Among the 90 valid answers, the outcomes break down like this:

OutcomeValid answers
Brand mentioned and owned page cited10
Brand mentioned, no owned page cited17
Owned page cited, no brand mention8
Neither55
Total90

The brand appears in 10 plus 17, which is 27 answers, so the mention rate is 27 divided by 90, or 30%. An owned page is cited in 10 plus 8, which is 18 answers, so the owned-page citation rate is 18 divided by 90, or 20%.

A fair way to say that is: across 90 valid answers to the monitored questions during this period, the brand appeared in 30% and an owned page was cited in 20%. An unfair way would be to say that 30% of AI users saw the brand and 20% visited the website. The tracker did not measure either of those things.

You also should not add the two rates together to get a total visibility figure. Ten answers contain both signals, and adding would count them twice.

If you had divided the same 27 mentions by all 100 planned observations, you would get 27% and not 30%. Only the denominator policy changed, not the answers. A valid answer with no mention is a real zero, while a failed check is missing evidence, which is why coverage is worth reporting.

Pages are counted differently again

Suppose the 18 answers with an owned citation include Page A cited in 12 answers and Page B in 9, with three answers citing both. That is 18 answers with an owned citation, 21 page-and-answer pairs, and 2 unique cited pages. Each is a real measurement, and none should be relabeled as another.

Share of voice uses another pool

A simple, mention-based version of share of voice divides your counted mentions by the counted mentions for all included brands. Using the same answers, say the monitored brand has 27, Competitor A has 45, and Competitor B has 36. The pool holds 108, so the brand's share of voice is 27 divided by 108, which is 25%, while its mention rate is still 30%.

Add a third competitor with 27 mentions and the pool grows to 135. Your share drops to 20% even though your own 27 mentions did not change. Share of voice describes your slice of a pool you defined, not market share or revenue.

Labels are not standardized

Two published definitions show why you should read the fine print. Semrush describes share of voice in terms of the percentage of brand mentions relative to competitors. Ahrefs documentation dated September 3, 2026 defines AI Share of Voice as a share of impressions relative to tracked brands, with weighting across platforms, and it bases those impressions on Google search volume estimates tied to prompts. Estimated impressions like that are a modeled figure and not a count of people who actually saw the brand in an answer.

So two dashboards can both say "share of voice" and mean different things.

One more thing to watch is that a move from 20% to 30% is 10 percentage points, or 50% relative growth, so avoid writing "up 10%" when you mean points.

Why can AI visibility results change?

Repeated measurement helps you tell a recurring pattern from a one-off answer. It does not turn an unrepresentative set of questions into a representative one. Both parts of that sentence matter.

Variation shows up at several levels, and it helps to keep them apart:

  1. Run-to-run variation, where the same question gives different answers when you repeat it.
  2. Question variation, where a change in wording, constraints, or intent changes what is being measured.
  3. Platform variation, where different platforms name different brands and cite different sources.
  4. Time variation, where later observations differ as systems, source material, and demand shift.
  5. Context variation, where memory or location changes the result.

OpenAI says that saved memories may influence how search queries are rewritten and that approximate location can affect local results. So when you ask ChatGPT yourself and get an answer that does not match your dashboard, that is not automatic proof that the tracker is wrong. Your own session is simply another observation, with its own context.

There is research behind the repeated-measurement idea. A 2026 arXiv manuscript called Don't Measure Once: Measuring Visibility in AI Search looks at how stable both mentioned brands and cited sources are across repeated runs. One of its datasets has daily results from four AI search engines (ChatGPT, Perplexity, Gemini, and Google AI Mode) across four Swiss-German campaign verticals over a 45 to 46 day window. The practical takeaway is that a single observation is a weak description of ongoing visibility. It is a manuscript, not a confirmed peer-reviewed finding, it covers a specific region, and it does not give a universal number of runs to use.

More runs also do not fix every problem:

  • Repeating only questions that contain your brand name still measures a brand-name question set.
  • More observations will not correct a misidentified brand name or a domain assigned to the wrong owner.
  • A smoother trend line does not prove that a content change caused the improvement.

Keep the frame steady when you compare periods

A change from one period to the next is easiest to trust when the things around it stay the same. That means the same questions, platforms, brand and domain definitions, counting rules, measurement coverage, and reporting window length. If you add a lot of new questions or a new platform, the headline rate can fall without anything having gotten worse. When the set changes a lot, say so and start a new baseline.

What a tracker cannot tell you on its own

Some limits are worth keeping in mind when you present numbers to anyone else.

It is not a census of customer conversations. A tracked sample does not tell you every question customers ask or how many people were exposed to your brand.

It is not traffic or revenue attribution. Microsoft's documentation for Bing Webmaster Tools says plainly that citations do not represent traffic, clicks, or user engagement. A tracker observation is not proof of a visit, a lead, or a sale. If you want to judge business impact, put visibility next to your website analytics and commercial reporting, and do not use it as a replacement for them.

It is not a permanent rank. Your order in one observed answer is not a position you hold across all users and future runs, and Bing also says its AI citation reporting does not measure rankings, authority, or importance.

It is not proof of accuracy or cause. OpenAI warns that search results and citations can be incomplete, outdated, or incorrect, so check that a cited page really supports the claim. If visibility goes up after you publish an article, the timing alone does not show the article did it.

And it is not a universal benchmark. No standard good mention rate, citation rate, or score has been established, and a score of 70 out of 100 does not mean you appear in 70% of all AI answers unless the metric is defined that way.

First-party reports are a different dataset

The platforms publish some reporting too. Bing Webmaster Tools has an AI Performance report covering citation activity across Microsoft Copilot and AI summaries in Bing, with total citations, cited pages, and average cited pages. Google announced Search Generative AI performance reports in Search Console on June 3, 2026, covering AI Overviews and AI Mode, with impressions that count how often links to your site were shown. Neither is a mention rate across prompts you chose, so they answer different questions than a tracker does.

What can marketers do with the findings?

A dashboard is most useful when it points you to the next question. These patterns come up often.

Pattern in the dataA reasonable readingA good next question
Mentions are frequent, owned citations are rareThe brand appears without much credit to its own pagesWhich outside sources appear, and what do they say?
Owned pages are cited, mentions are rareYour content supplies facts without the brand being namedWhat role does the cited page play in the answer?
A few pages produce most citationsSource visibility is concentratedWhich questions keep leading to those pages?
Competitors keep showing up for key buyer questionsThe sample shows a competitive gapWhich brands and exact pages appear, and how consistently?
The headline rate drops after many questions are addedThe scope may have changedWhat happens when you compare the original questions on their own?
Mentions rise but descriptions are wrongPresence improved without the message improvingWhat does the answer actually claim?

These patterns give you hypotheses and priorities. They do not prove why something happened, and they do not guarantee that publishing more content will close a gap.

When you report to leadership, describe what was measured. Something like this works: "For our monitored buyer questions, our brand appeared in 27 of 90 valid answers during the period. Owned pages were cited in 18. Visibility differed by platform, so we are reviewing the answers and cited pages behind the gaps."

An AI visibility tracker is useful when its score leads back to a clear question, an observed answer, and a cited page. The goal is not to chase an unexplained number. It is to understand where your brand appears, what the answer says, and how consistently the pattern shows up. If you want to try this on your own buyer questions, you can start a free DeepSmith trial and look beyond the headline score at the answers and pages behind it.

Frequently asked questions

What is AI visibility tracking?

AI visibility tracking is the ongoing measurement of whether a brand and its pages appear in AI-generated answers to a defined set of questions. It commonly separates brand mentions from source citations and follows how both change over time.

What is the difference between an AI mention and a citation?

A mention names the brand. A citation presents a page as a source. An answer can do both, either one, or neither, and neither signal alone guarantees a recommendation.

Does a visibility score show how many people saw my brand?

Not necessarily. A score from monitored prompts usually summarizes observed answers within a stated sample. It should not be treated as audience reach, clicks, or revenue unless a separate measurement says so.

Why does my own AI search show a different answer?

Answers can vary across runs, platforms, time, and context, and OpenAI documents effects from memory and location. Your own search is another observation, and it does not by itself show that the tracked answer was wrong.