AI Visibility Tracker: Track Brand Share in AI Search

TL;DR: An AI visibility tracker samples prompts against AI assistants on a schedule and reports whether your brand was named and whether your URL was cited. Those are two different events and most reporting blurs them. Define your metrics precisely, build a prompt set that reflects real demand, record variance alongside every rate, and treat the output as a work queue rather than a scoreboard.

Key Takeaways

• It samples, it does not observe. A tracker runs your prompts and records answers. It is not watching real user conversations, and no tool has access to those.

• Mentions and citations are separate events. Being named in an answer and having your URL cited are different outcomes with different fixes. A blended score cannot tell you which applies.

• Report variance, not just the rate. The same prompt produces different answers across runs. A rate without a sample size and a range is not a measurement.

• Your prompt set is the measurement instrument. Change it and the number moves for reasons unrelated to performance. Version it and freeze it between comparisons.

• You cannot edit training data. Claims about correcting a model's training are wrong. What you influence is what gets retrieved and what third-party sources say.

• The output is a work queue. A dashboard nobody acts on costs more than it saves. The point is identifying which prompts you lose and why.

Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.

What a tracker actually does

An AI visibility tracker maintains a list of prompts, sends them to AI assistants on a schedule, records the answers, and reports whether your brand appeared.

Related reading: AI Search Visibility with Telepathic: 2026 Guide

That is the whole mechanism, and understanding it precisely prevents most of the misinterpretation that follows.

What it is not

The first row matters most. Every figure a tracker produces describes a synthetic sample that you defined, not observed demand. That does not make it useless; it makes it a survey rather than a census, and surveys need sample sizes and error ranges.

What it samples

That last row is worth knowing. Assistants do not always search. Some answers come from the model's own parameters and some from live retrieval, and your content can only influence the second kind.

What good output looks like

If a tool gives you a score without the underlying answers, you cannot diagnose anything from it. The answers are the data; the score is a summary of the data and a poor substitute for it.

Why that matters for buying

When comparing tools, ask to see one prompt you lose, end to end, with the full answer and every citation. That single request separates products that store the evidence from products that store a number.

The three metrics, defined

Three metrics do the work. Each needs a definition precise enough that two people computing it from the same data get the same number, which is a lower bar than it sounds and one most dashboards fail.

1. Mention rate

Definition: the share of runs in which your brand name appears anywhere in the answer text.

Mention rate = runs naming your brand / total runs

What it tells you: whether assistants associate your brand with the topic at all. A low mention rate means you are not in the consideration set.

What it does not tell you: whether your content earned it. Mentions often come from third-party sources rather than from your site.

2. Citation rate

Definition: the share of runs in which a URL on your domain appears in the answer's sources.

Citation rate = runs citing your domain / total runs

What it tells you: whether your pages are being retrieved and used as source material. This is the metric your content work moves most directly.

A necessary caveat: only answers that invoked retrieval can cite anything. If you want a cleaner denominator, compute citation rate over runs that cited any source at all, and say which denominator you used.

3. Recommendation rate

Definition: the share of runs where your brand is presented as a suggested option rather than merely referenced.

Recommendation rate = runs recommending your brand / total runs

What it tells you: the closest proxy to commercial outcome. Being listed as one of three suggestions is different from being named in passing.

The hard part: this requires classifying answers, which means a documented rule for what counts. Write the rule down, apply it consistently, and spot-check the classifier against your own reading.

Why the separation matters

A single blended visibility score collapses four different situations into one number and tells you nothing about which of four fixes applies.

On sentiment

Sentiment classification is useful and frequently oversold. Treat it as a flag prompting you to read the actual sentence, not as a metric to report. Any sentiment score you cannot trace back to retrievable text is decoration.

Metrics that sound rigorous and are not

Several metrics circulate in this category that cannot be computed as described. Knowing which is useful when reading a vendor's feature list.

Share of model voice

Usually defined as your share of brand mentions compared with competitors across tracked prompts. That is computable and reasonable, provided two things are true.

First, it is a share of your prompt set, not of a market. Adding prompts you lose lowers it; adding prompts you win raises it. Second, it requires a fixed competitor list, since a new entrant appearing in answers changes your share without anything about you changing.

Use it, define it as share across your frozen prompt set, and never present it as market share.

Semantic positioning or vector proximity

Some material claims trackers measure how close your brand's vector sits to industry terms, and that services can optimise that position.

No external tool has access to a commercial model's embedding space. You cannot measure your position in it and nobody can tune it directly. What you can do is write text that clearly associates your brand with the topics you want, which influences retrieval through ordinary means.

The underlying advice, which is to write focused content on your actual area rather than diluting across unrelated topics, is sound. The measurement claim attached to it is not.

Training data influence

Claims that a service can correct a model's training or ensure your information appears in training datasets do not describe anything achievable. Training corpora are assembled by model providers, updated on their schedule, and not editable by third parties.

What is achievable is influencing retrieval: what your site says, what third-party sources say, and whether crawlers can reach you.

Agentic readiness scores

An emerging category of metric claiming to measure how autonomous purchasing agents will evaluate your products. Agentic commerce is a real direction and these scores currently measure a vendor's model of a future behaviour rather than that behaviour.

Treat them as speculative. The underlying work, which is clear product data and structured pricing, is worth doing for ordinary reasons.

The test to apply

For any metric, ask: what exactly was counted, out of what denominator, and could I reproduce it from the raw answers?

If the vendor cannot answer that, the metric is a summary of something unstated, and you should not put it in a board deck.

Building your prompt set

Your prompt set is the measurement instrument. Build it badly and every number that follows is measuring the wrong thing, confidently.

Where prompts come from

Not from a keyword tool. Prompts are conversational and longer than keywords, and keyword databases are built on a different behaviour.

The intent split

A representative set spans four intents, because your performance differs sharply across them.

Branded prompts inflate your numbers and belong in the set in proportion, because tracking only branded prompts produces a flattering figure that moves with nothing.

How many

Below about 50 prompts, variance dominates and monthly movements tell you nothing. Above a few hundred, you are usually adding near-duplicates rather than coverage.

Version and freeze it

This is the discipline most teams skip. Give the set a version number and a date. Do not add or remove prompts mid-quarter.

If you change the set, your rate changes for reasons unrelated to performance, and the comparison across that boundary is meaningless. When you must change it, note the version change on the chart and treat it as a new baseline.

The review cycle

Revisit the set quarterly against fresh sales calls and support tickets. Language shifts, new competitors appear, and a set built eighteen months ago is measuring a market that has moved.

Change it deliberately, at a quarter boundary, with the version recorded. That keeps the instrument current without destroying your trend line.

Setting a baseline

A baseline that survives scrutiny needs a sample size, a date, a prompt set version and a variance figure. A single number with none of those is an anecdote with a percentage sign.

Why variance matters here more than elsewhere

The same prompt asked twice produces different answers. Sources differ, phrasing differs, and sometimes the brands named differ entirely. That is normal model behaviour rather than a fault.

So one run per prompt tells you almost nothing. Five runs gives you a usable rate. Fewer than that and you will report a change next month that is sampling noise, and someone will act on it.

What to record

How to report it

Instead of a bare citation rate of 23%, report it with its shape:

Citation rate: 23%
  Prompt set v1.2 (120 prompts), 5 runs each, 600 runs total
  Measured 1 to 7 October 2026
  Engines: ChatGPT, Gemini, Perplexity
  Range across runs: 19% to 27%

That version is defensible in a review, and the range is what stops a move from 23% to 25% being presented as progress.

What counts as a real change

As a working rule, treat a movement smaller than your observed run-to-run range as noise. If your range is 19 to 27, a shift to 25% is not a result.

Look instead at the pattern underneath: which specific prompts changed state, and did the change cluster around pages you worked on? Prompt-level movement is far more informative than the headline rate.

Re-measure on a fixed cadence

Monthly for most teams, with the same set, the same runs and the same engines. Quarterly for the full review including the prompt set itself.

Consistency of method matters more than frequency. A monthly measurement taken the same way every time beats a weekly one taken differently.

Acting on what it tells you

A tracker earns its cost only if its output becomes work. The useful unit is not the dashboard, it is a weekly list of specific prompts you lost and the reason for each.

The weekly triage

Thirty to sixty minutes. Pull the prompts where you were not named or not cited, read the actual answers, and sort each into one of four buckets.

Four buckets, four owners, and each produces a specific task rather than a general ambition.

The one to check first

Not retrieved. If your pages exist and never surface, no amount of content work helps, and the cause is usually a crawler block or client-side rendering. Confirm in server logs by user agent and status code.

That single check has saved more programmes than any content strategy, and it costs an hour.

What to report upward

Senior stakeholders cannot act on a score. They can act on a list of named gaps with owners and a note on what shipped.

The honest attribution position

You will be asked what revenue this produced. The accurate answer is that AI referral attribution is genuinely difficult: many assistants pass no referrer, some answers produce no click at all, and the influence often shows up as better-informed buyers rather than tracked sessions.

Say that plainly, then offer what you can measure: citation rate on commercial prompts, self-reported source at intake, and branded search volume. Overclaiming here is how measurement programmes lose credibility the first time someone checks.

What a tracker cannot do

Being precise about the limits is what makes the rest of the measurement trustworthy.

It cannot see real user conversations

No tool has access to what people actually ask assistants. Every figure describes your synthetic prompt set. Treat the output as a survey of a sample you designed, with all the caveats that implies.

It cannot edit or correct training data

This is the most common overclaim in the category. Training corpora are assembled by model providers on their own schedule. No service updates them, corrects them, or inserts your information into them.

What you can influence is retrieval: what your pages say, whether crawlers can reach them, and what third-party sources assert. When a model's output improves after you fix something, that is almost always retrieval, not retraining.

It cannot fix a hallucination directly

When an assistant states something false about your product, there is no correction mechanism to submit to. What you can do is address the sources: correct outdated pages of your own, update third-party listings, and publish a clear, current statement of the facts.

That works over time through retrieval and third-party corroboration, not immediately and not reliably. Some providers offer feedback mechanisms, and they are not a guaranteed route.

It cannot tell you about traffic or revenue

It reports answer content. Sessions and conversions come from your analytics, and AI referral attribution is incomplete by nature, since many assistants pass no referrer and many answers produce no click.

It cannot distinguish retrieval from parametric knowledge, usually

When an assistant names your brand without citing anything, that may come from training or from an uncited retrieval. Most tools cannot tell you which, which matters because only one is influenceable by your content.

A useful proxy: if the answer cites nothing at all, your content work is unlikely to have driven it.

What this adds up to

A tracker is a good instrument for a narrow question: across prompts I chose, how often do assistants name and cite me, and who do they name instead?

That question is worth answering and the answer drives real work. Everything beyond it, in either direction, is either someone else's tool or nobody's.

How retrieval actually works

Worth understanding accurately, because the mechanism explains why some tactics work and why others cannot.

The sequence

1. Someone asks a question.

2. The assistant decides whether to search. Not every answer triggers retrieval.

3. If it searches, it typically rewrites the question into one or more queries, which is why near-me becomes a city name.

4. It retrieves candidate documents from a search index.

5. It selects passages from those documents rather than whole pages.

6. It composes an answer, often citing some of what it used.

What each step implies

Step five is the one that changes how you write. A page that buries its answer in the sixth paragraph may be retrieved and still contribute nothing quotable, which is why opening passages carry so much weight.

Where embeddings come in, accurately

Retrieval systems commonly represent text as numeric vectors so that semantically similar passages sit near each other, which is how a query about stomach pain can match a page about gastritis without sharing words.

That is real and it is useful to know. Two things follow that are often stated incorrectly.

You cannot measure or tune your position in a commercial model's vector space. No external tool has access to it.

And you do not need to. The practical consequence of semantic matching is simply that clear, focused, on-topic writing matches more queries than keyword-stuffed writing does, and that covering one subject properly beats scattering across unrelated ones.

The practical summary

Four levers. Everything else in this category is either one of these restated or something nobody can actually do.

Off-page authority and where the line is

Third-party corroboration genuinely matters. Assistants cross-reference sources, and a brand described only on its own website has nothing supporting it. That observation is correct and it leads some practitioners somewhere they should not go.

What genuinely works

The common thread is that each produces a real artifact a real person made for a real reason.

Two tactics that cross a line

Manufactured community discussion. You will see advice to ensure your brand is frequently mentioned in natural-seeming conversations on Reddit, X and similar platforms, with a service arranging the mentions.

That is astroturfing. It breaches platform rules, which in Reddit's case means bans for accounts and often domains. It also runs into FTC guidance on endorsements, which requires disclosure of a material connection between an endorser and a brand. Paid or arranged posts presented as organic opinion are exactly what that guidance addresses.

The risk is not only enforcement. Communities detect this reliably and the reputational outcome when they do is worse than the visibility was worth.

Bought links positioned as AI authority signals. Buying high-authority backlinks to signal credibility to generative engines runs into Google's link spam policy, which names buying links that pass ranking credit and link building through automated programs or services. Rebranding the tactic for AI search does not change what it is.

The honest version of the same goal

The goal is reasonable: be present where assistants look. The legitimate route is slower and durable.

If you work with an agency

Ask directly how third-party mentions are obtained. If the answer involves arranged posts, undisclosed paid placement or purchased links, the exposure attaches to your brand and your domain rather than to theirs.

Free versus paid

Manual checking is a legitimate method, not a poor substitute. It is how you should start, and for many teams it is sufficient indefinitely.

The free method

Twenty prompts, five runs each, across three engines, logged in a spreadsheet. That is 300 runs and roughly two hours monthly if you are thorough.

The first row is an advantage worth naming. A person reading the answers notices things a classifier does not: a competitor described more warmly, a feature attributed incorrectly, a framing that disadvantages you.

Also free and worth connecting

Google Search Console's generative AI performance report covers AI Overviews and AI Mode with your own first-party data. That is measurement nobody can sell you a better version of, for your own site.

When to switch to paid

If manual checking has been slipping, that is the signal. A process that only works when someone is diligent is not a process.

What paid should give you

That last row is the one to test rather than assume. Run the spot check before committing, because a tool that undercounts produces a confident, consistent and wrong picture for as long as you keep it.

The cost framing

Entry pricing in this category runs from about $29 to $500 monthly. Compare on cost per tracked prompt at your real volume with the engines you need, since allowances vary far more than headline prices.

Seven mistakes in AI visibility measurement

1. One run per prompt. Answers vary between runs. A single sample is an anecdote, and a month-over-month change built on single samples is noise you will act on.

2. Changing the prompt set mid-quarter. The set is the instrument. Add prompts you win and the rate rises with nothing improving. Version it, freeze it, and note the version on every chart.

3. Blending mentions and citations. Four different situations collapse into one number, and the number cannot tell you which of four fixes applies. Report them separately.

4. Reporting a rate without a range. If your run-to-run range is 19 to 27, a move to 25 is not a result. Report the range alongside the rate and most false progress disappears.

5. Tracking mostly branded prompts. They flatter the number and move with nothing. Keep them to a proportionate share and weight category and comparative prompts, which is where the commercial outcome lives.

6. Treating the dashboard as the deliverable. A tracker earns its cost through the weekly triage: which prompts were lost, why, and who fixes each. A score nobody acts on is a subscription for reassurance.

7. Believing you can edit training data. You cannot. What moves is retrieval, and conflating the two leads teams to buy services promising something nobody can deliver.

The pattern

Five of seven are measurement hygiene rather than strategy. That is characteristic of a young category: the hard part is not deciding what to do, it is producing numbers that survive someone checking them.

The one that matters most

Mistake six. Every other item on this list is a methodology problem that makes your numbers worse. That one makes the whole programme pointless, because a measurement that produces no work has no reason to exist.

If you cannot name three things you changed this quarter because of what the tracker showed, cancel it and go back to manual checking until you can.

FAQs

What is an AI visibility tracker?

A tool that sends a list of prompts to AI assistants on a schedule, records the answers, and reports whether your brand was named and whether your URL was cited. It samples prompts you chose rather than observing real user conversations, which no tool can do.

How is it different from a rank tracker?

A rank tracker reports your position for a keyword on a results page. An AI visibility tracker reports whether a synthesised answer named you and which sources it cited. There are no positions to report, so the output is a rate rather than a rank.

What metrics should I track?

Mention rate, citation rate and recommendation rate, computed separately because they have different causes and different fixes. Report each with a sample size and the observed range. Treat sentiment as a flag that prompts you to read the answer rather than as a headline metric.

How many prompts do I need?

100 to 150 for a single product in one market, split across category, comparative, problem-first and branded intents. Below about 50, variance dominates and monthly movements are meaningless. Freeze the set and version it so comparisons stay valid.

How often should I measure?

Monthly, with the same prompts, the same number of runs and the same engines. Run each prompt at least five times, because answers vary between runs. Review the prompt set itself quarterly against fresh sales and support language.

Can a tracker fix incorrect information about my brand?

Not directly. There is no correction mechanism to submit to, and nothing can edit a model's training data. What you can do is address the sources: update your own outdated pages, correct third-party listings, and publish a clear current statement of the facts.

Can I measure my brand's position in a model's vector space?

No. External tools have no access to a commercial model's embeddings, so nothing can measure or tune that position. The practical substitute is writing clear, focused content on your actual subject, which is what semantic matching rewards anyway.

Do I need to pay for one?

Not below about 50 prompts. Twenty prompts, five runs each, across three engines, logged in a spreadsheet takes around two hours monthly and gives you full answers read by someone who understands the product. Pay when scale or multiple brands make that unsustainable.

Setting up AI visibility measurement

Start manually. Twenty prompts drawn from sales calls and support tickets, five runs each, across three assistants, logged in a spreadsheet with the date and the version number. Two hours, no budget, and you will understand your position better than a dashboard would tell you.

Then define your three metrics precisely, record the range alongside every rate, and turn the output into a weekly list of lost prompts with owners. That list is the point of the exercise.

When to buy nothing yet

• Under 50 prompts. Manual checking is sustainable and gives you full answers read by someone who knows the product.

• You have not verified crawler access. If AI crawlers cannot reach your pages, measuring the consequence is premature. Check server logs by user agent.

• Nobody will run the triage. A tracker whose output nobody acts on is a subscription for reassurance.

• You cannot yet define your metrics. Buy after you know what you are counting, not before.

If the gap is execution rather than measurement

Most teams who measure properly discover the same thing: they know which prompts they lose and cannot produce the pages to win them. That is a production constraint, and Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.

It does not track prompts or store AI answers, so it sits alongside a tracker rather than replacing one. For many teams that pairing costs less than a single enterprise measurement tool.

Summary

An AI visibility tracker samples prompts you chose against AI assistants and reports whether you were named and cited. Define mention rate, citation rate and recommendation rate separately, because they have different causes and different fixes, and report each with a sample size and a range so a move from 23% to 25% is not mistaken for progress. Build your prompt set from sales calls and support tickets rather than a keyword tool, version it, and freeze it between comparisons. Remember that no tool sees real user conversations, measures your position in a model's vector space, or edits training data: what you influence is retrieval. And judge the whole programme by whether it produces a weekly list of specific prompts to fix, because a score nobody acts on is the most expensive thing in this category.