How to Track Your Brand's AI Visibility

Every week a business owner tells me some version of the same thing. They know their customers have started asking ChatGPT for recommendations. They suspect their business is not the one getting named. What they cannot tell me is whether that suspicion is correct, because nothing in their reporting has an answer for it. Analytics does not have a line for it. The rank tracker does not have a column for it. So the question sits there, unmeasured, while the answers keep getting generated without them.

That gap is the actual problem, and it is worth naming plainly. Classic search handed us a scoreboard within a few years of becoming important. AI search became important faster than the measurement caught up, so most businesses are either optimizing for it blind or ignoring it and hoping. Neither is a plan. What follows is how we measure AI visibility for the businesses we run this work for: what the term means once you get specific, what is worth tracking, how to build a question set that reflects real buying behavior, and how to keep the whole thing honest when the answers shift week to week.

What does AI visibility actually mean?

Most people collapse all three into one vague idea of “showing up,” and that is why their tracking never tells them anything actionable.

Mention is your business name appearing in the generated answer. This is the one that wins the customer, because a person asking an assistant for a recommendation is reading names, not evaluating links.

Citation is your site listed as a source underneath or beside that answer. Citations matter for a different reason. They are the model telling you which of your pages it found credible enough to lean on, which makes them the clearest feedback signal you have about what is working. You can be mentioned without being cited, when the model knows you from elsewhere on the web, and you can be cited without being mentioned, when your page taught the answer but the recommendation went to someone else. That second case is the most frustrating one to discover and the most useful one to fix.

Accuracy is the part nearly everybody forgets to track, and the part that quietly does damage. An assistant that confidently recommends you while describing a service you dropped, a region you do not cover, or an outdated way of working is not a win. It is a bad first impression delivered at scale, and you will never hear about it from the customer who read it and moved on.

Why doesn’t Search Console show your AI visibility?

This surprises people, so it is worth being precise. Google has not given us an AI Overview filter, a separate impression type, or a citation report. What lands in Search Console is search performance in aggregate, which means the most important shift in how people find businesses is happening invisibly, inside a number you already look at every month.

One pattern in that data is worth watching, because it is often the first hint. When a page holds its position while its clicks decline over a sustained period, something is answering the query before the person reaches you. That is not proof and it can have other causes, so treat it as a reason to go look rather than a conclusion. We flag exactly that pattern on client accounts during our content sweeps, and it has pointed at a real problem more often than not.

Even a perfect Google-side report would still cover only Google. The assistants people ask for recommendations are separate products, none of which publish a dashboard for the businesses they name. Nobody is going to hand you this number. You have to go collect it.

What should you actually measure?

Presence rate is a simple fraction: the number of prompts where you were named, divided by the total prompts run. Its value comes entirely from repetition, because the absolute number on any single month means very little and the direction across six months means everything.

Citation rate is tracked the same way, but tells you something different. Presence tells you the model knows who you are. Citation tells you which specific pages earned its trust. When one post gets cited repeatedly, that is the format working, and it should shape what you publish next. This is the same feedback loop behind getting cited by ChatGPT and Perplexity, and it is the closest thing to a rank tracker this channel currently has.

Share of voice requires a fixed competitor list, decided before you start and left alone. Write down the four or five businesses you actually lose deals to, then count how often each is named across the same prompt set. This is usually the number that gets an owner’s attention, because “we appeared in six of thirty answers” is abstract while “they appeared in nineteen and we appeared in six” is not.

Accuracy is a log, not a percentage. Every time an assistant says something about your business that is wrong, incomplete, or years out of date, record the exact claim and the exact prompt that produced it. That log becomes a punch list, and it is frequently the highest-value output of the whole exercise.

How do you build a prompt set worth tracking?

Build it in three tiers, because customers arrive at different stages and the assistants answer them differently.

Discovery prompts describe a need with no vendor in mind. “Who produces radio commercials for national franchise brands.” “What kind of company builds a custom jingle.” These reveal whether you exist in the model’s picture of your category at all.

Comparison prompts ask for a shortlist. “Best agencies for AI search optimization.” “Top companies for white label video production.” These are where share of voice lives, and where being absent hurts most, because the person reading is close to a decision.

Problem-first prompts start from the symptom instead of the solution, which is how a large share of real searches begin. “Our website ranks fine but nobody calls, what is wrong.” “How do I get my business mentioned when someone asks an assistant for a recommendation.” These often surface competitors you were not tracking.

Two rules keep the set useful. Keep most prompts unbranded, because a prompt containing your name will nearly always return a flattering answer and teaches you nothing about acquisition. And include geography only if you genuinely sell that way. We work with businesses nationwide, so a prompt set built entirely around one city would measure the wrong market and quietly reward the wrong content.

How often should you check, and who should do it?

The session rules are not fussiness. Assistants personalize from your account history and memory, so a check run inside the account you use all day is measuring your own footprint rather than what a stranger would see. Log out. Open a clean window. Ask the question cold.

Variance is the other reason for discipline. The same prompt can return different names and different sources across runs, because these systems are probabilistic by design. That variance is exactly why one check proves nothing, and why saving verbatim responses matters. A month later you will want to compare wording, not memory.

Whoever runs it should follow a written script, in the literal sense: same prompts, same order, same conditions, results pasted into the same sheet. It is dull work and it is the reason the numbers end up meaning something. Most of the AI visibility “data” I get shown by prospects is a screenshot of one good answer somebody found once, which is a nice moment and not a measurement.

Can you see AI assistants in your analytics?

Set up the GA4 channel group anyway, because that traffic arrives with unusual intent. Somebody clicking through from a generated answer has already been told you are a credible option, which puts them further along than most first-time visitors.

The log side answers a different question: can these systems read you at all. Look for the AI crawlers in your access logs or CDN dashboard and confirm they are being served rather than blocked. This is a precondition, not a metric. A site that quietly blocks AI crawlers cannot be cited no matter how good the writing is, and I have found that exact misconfiguration on sites whose owners were paying for content every month. That is why AI optimization starts with the plumbing.

What do you do when an assistant gets your business wrong?

There is no support line for a wrong answer, and that turns out to be fine, because almost every wrong answer traces back to something you control. A service area that reads differently on your site than on your profiles. A service you retired that still has a live page. A business name written three ways across the web. Faced with conflicting inputs, the model does what a cautious person would do and either hedges or picks the version that appears most often, which may well be the outdated one.

So the repair is unglamorous and effective. Make the fact consistent everywhere, put it in plain sentences on a page you own, and give the systems time to re-read you. This is entity work, the same foundation described in AI search optimization, and it is the single most common gap we find when we audit a new account.

Absence works the same way, with one addition. If your competitors are named for a question and you are not, the gap is usually that they have published a clear, specific answer to that exact question and you have not. That is a content problem with a content fix, and it is the most direct path from a tracking spreadsheet to an actual result.

How do you turn tracking into a result?

The order matters more than the effort. Businesses that skip measurement publish constantly and cannot tell you which of it worked. Businesses that measure first find that a handful of specific questions are doing the damage, and that fixing those is a smaller job than the one they imagined.

Set the expectation honestly at the start, too. Structural repairs can surface within weeks. Becoming the business an assistant reaches for by default across a whole category is a quarterly project, because it depends on the model trusting you, and that trust is assembled from your site, your profiles, your reviews, and mentions on properties you do not own. Anyone promising faster than that is describing something else.

We built our own audit for exactly this reason. The KillerSEOx audit measures how visible and citable a site is to AI systems and names the specific gaps, which is what turns the monthly check into a work list. The program behind it is LLM visibility work, and it sits alongside the classic SEO and AIO foundation rather than replacing it. If the vocabulary is new, answer engine optimization covers the terms and optimizing for Google AI Overviews covers the Google side.

Start this week

Pick fifteen questions your customers actually ask. Log out, run them cold, and paste the raw answers into a sheet with a date on it. Count how many name you, count how many name each competitor, and write down every claim about your business that is wrong.

That is your baseline, and it will take an afternoon. The number will probably be worse than you hoped, which is the useful part, because from that point forward you are working from evidence instead of a suspicion. Do it again in thirty days. The businesses winning this channel are not the ones with a secret. They are the ones who started keeping score while everyone else was still guessing.