AI visibility is measured by asking the engines the questions your buyers actually ask, recording whether your business gets named and which URLs get cited, and counting the answers with the denominators intact. That is the short version of how AI visibility is measured, and it fits in one sentence because the definition is not the hard part.
The hard part is that almost nobody who sells you an AI visibility score will show you those denominators, and almost nobody reads back the words the engines use. Doing both is what turns a number into a decision. Here is the full method we run for clients, including the part that gets skipped.
Why the percentage you were quoted probably means very little
A single number compresses everything. Twelve questions across five engines produces sixty answers. Turn one answer from a loss to a win and the headline percentage jumps about eight points. Nothing in the business changed. One answer did.
That is not automatically dishonest. It is a ruler, and the person selling you the measurement chose where the marks go. Three ways that goes wrong:
- The question set is picked to flatter. Choose twelve questions the client already wins and the score is excellent. It also tells you nothing about the questions they lose.
- The competitor list is written by the seller. Share of voice is only meaningful against a disclosed field. Against a hand-picked, convenient field it is decoration.
- The question set changes and the old score is still compared to the new one. A rebuilt panel can move a number thirty points in either direction without a single page changing.
So the first question to ask anyone selling AI visibility is not "what is my score". It is "how many questions, which engines, how many answers, and who is on the competitor list". If any of those four cannot be answered in one breath, the number is not a measurement yet. We wrote a separate piece on how to spot a bent ruler: Shape the Work, Never the Ruler.
How is AI visibility measured, then? Four things you can actually count
Strip the composite score away and four countable things remain. All four are checkable by hand, which is the point.
- Mentions, with the denominator attached. Not "you have 30% visibility". Instead: "you were named in 6 of 60 answers". Both numbers, every time. Six of sixty is small. Six of twelve is a different story, and the raw fraction forces that conversation.
- The citation source. Which URL did the engine point at when it named you? This is the most ignored and most useful field in the whole exercise. If the engine cites your city's directory page rather than your own site, your site is not doing the work you think it is.
- Query by query results. Which specific questions you lose, and who wins each one. A category-level average hides the fact that you are strong on "how much does it cost" and invisible on "who is the best", and those two need opposite fixes.
- The rolling rate across cycles. One month is a snapshot. Three consecutive readings is a trend. A business at 30% for three cycles is in better shape than a business that spiked to 45% once when a competitor's site was down for a rebuild.
That is measurable, defensible, and repeatable. It still leaves out the step that produces the actual recommendations, which is the next section.
Read back the words the engines use
Keyword research answers a different question than the one you need. An autocomplete tool tells you what people type. It does not tell you what gets a business cited. Run a keyword tool on a category and you get a list of search phrases. Run it again next month and you get roughly the same list, because the way people type does not change quickly.
Meanwhile the engines publish their preferences in every single answer they give. Whatever words they keep using to describe the businesses they name are, by definition, the words they treat as the canonical language of that category. Nobody has to guess at those. You have to read them.
So we build three lists from the answers we already collected:
- The words in the sentence where the engine names you. If it names you and describes you as "an interactive web app that pairs a live star map with scripture", those are not our marketing words. Those are the engine's words for you, and they are worth more than any phrase a tool suggested.
- The phrases that recur across all the answers. Not once. Repeatedly, across engines and across questions. Those are the category's operating vocabulary.
- The domains being cited. Where the engines actually get their answers tells you which third-party listings carry weight in that vertical, and which assets you are missing.
Autocomplete has exactly one job in this process: confirming that a phrase the engines use is also something a human being types. Presence only. It never establishes importance, because it cannot. Search volume is a decent proxy for interest and a poor proxy for whether an engine will put your name in a sentence.
A worked example, with the uncomfortable part left in
One client is a Christian astronomy resource: an interactive sky map paired with Scripture. Twelve questions, five engines, sixty answers, six of which named the site. Ten percent mention rate, with the denominators visible, which is the only form that number should ever take.
The words were the interesting part. The engines kept describing the category with "interactive star map", "live star map", "biblical astronomy", "creation astronomy", and "bible verses". On the site, two of those show up once each, on the home page. The other three do not appear on the home page, the sky map page, or the constellation pages at all. The site speaks in "Scripture" and "Psalms". The engines speak in "bible verses" and "creation astronomy". Same subject, different words, and the engines are the ones writing the recommendations.
Then the citation sources, and this is where the recommendation changed completely. The domains cited across those sixty answers were apps.apple.com forty-eight times, christianbook.com twenty-one times, amazon.com eight times. Not blogs. Not competitor websites. Retail and app listings.
Which meant the next asset was not another blog post. It was a listing the engines could read, and for that category a book is the cheapest honest way to get one. No percentage would ever have told us that. The percentage said "you are at ten". The words and the citation sources said "you are missing a store page, and the book is already written".
Same method on a service business produces the same kind of clarity. The answers around an electrician keep repeating "panel upgrades", "24-hour emergency", and "standby generator". The answers around a gas contractor keep repeating "licensed and insured" and "tankless water heaters". Write a post that never says those words and you have written a post the engine has no reason to connect to that category.
What to never accept from anyone selling AI visibility
- A score with no denominators. Six of sixty and six of twelve are not the same claim.
- A score that improved after the question set changed. Ask whether the questions moved. If they did, the old number is retired, not beaten.
- Share of voice against an undisclosed competitor list. You cannot audit a field you cannot see.
- A promise to make ChatGPT or Perplexity recommend you. There is no lever marked recommend. The work is shaping the signals those engines read, and the engines still decide. We put a business in position to be the answer. Nobody can sell you the answer itself.
- A report with no record of the actual questions asked. Without those, there is nothing to reproduce and nothing to argue with.
How to measure it yourself in an afternoon
You do not need software to start. You need discipline about the denominators.
- Write down ten to fifteen questions your buyers actually ask. Not your service list. The sentences people use in your inbox and on the phone. "Who installs a tankless water heater in Riverview" is a question. "Water heater services" is a category label.
- Ask all of them on five engines: ChatGPT, Perplexity, Google's AI answers, Claude, and Grok. Use a fresh conversation for each question so you are not coaching the model with context.
- Record three things per answer: were you named, which URL was cited, and which competitors were named. Paste the raw answer into a spreadsheet. Do not summarize it, because the words matter later.
- Count. Mentions over total answers, per engine and overall. Keep the fractions, not just the percentages.
- Read the words. Highlight every phrase the engine used to describe the businesses it named. Whatever repeats is the language of your category. That list, not a keyword tool export, is what your next piece of content should be built on.
Repeat it monthly with the same questions and the same engines. The first reading is a baseline, not a score. The third reading is a trend, and the trend is the only thing worth reporting to anyone.
Frequently asked questions
How is AI visibility measured?
By asking a fixed set of buyer questions across several AI engines and counting, for each answer, whether the business was named and which URL was cited. The measurement is the counts. Any percentage is a compression of those counts and is only trustworthy when the counts are shown alongside it.
Can I measure AI visibility without paying for a tool?
Yes. Ten to fifteen questions, five engines, one spreadsheet, about two hours a month once the question list is set. Paid tools save the labor of running the prompts and storing the answers. They do not improve the quality of your question list, which is where most of the measurement quality actually lives.
Does AI visibility replace SEO?
No. It sits on top of SEO and depends on it. If a crawler cannot read your pages, an engine cannot cite them. What changes is the target: SEO has historically optimized for a ranked list of links, and AI answers optimize for being the specific business named inside a paragraph. The technical foundation is the same. The content shape is not.
How long before AI visibility moves?
Longer than a ranking report and less predictably. Engines refresh what they know at their own pace, so a page that is live and readable today may not influence an answer for weeks. We treat three monthly cycles as the earliest honest read, and we say so up front rather than reporting a monthly swing as progress.
What counts as a good AI visibility score?
There is no universal good number, and anyone quoting one is selling. What matters is the direction across cycles, which specific questions you win, and who wins the ones you lose. A business named in the answers to the three questions that produce most of its revenue is in a better place than a business named broadly and vaguely in twelve.
Which engines should I test?
The ones your buyers use, plus the one you would not have guessed. For most local businesses that means ChatGPT and Google first, Perplexity and Claude next, and Grok because it indexes differently and sometimes surfaces a business the others miss. Five is a reasonable field. Two is a sample so small that one answer decides the story.
If you want the version with the denominators
We run this measurement every month for clients, keep the question set locked so the trend means something, and report the words alongside the counts. If you have been quoted an AI visibility percentage and want to know what is underneath it, that is a twenty-minute conversation.
Book a call at e320.io/book, or email keith@e320.io.
About the Author

Keith Marsh
Founder & Fractional CMO · E320 Consulting
Keith is a Dartmouth-educated marketing executive with 15+ years driving measurable growth for SMBs and established brands. As founder of E320 Consulting, he's delivered 40%+ ROI improvements for clients like Freedom Boat Club Tampa Bay. He specializes in fractional CMO leadership, GoHighLevel CRM strategy, lead generation systems, and converting marketing spend into predictable revenue.
Read Next
SEO Services in Apollo Beach, Florida: A Practical Guide for Local Businesses
Apollo Beach zip 33572 keeps growing, and so does search competition. What local SEO actually covers for Apollo Beach businesses: Google Business Profile, chamber and community links, on-page basics, timelines, and costs.
Who Owns Your Website and Ad Accounts?
Whose name is on the account is who owns it. The three ownership arrangements, the four questions to ask before signing, and what leaving actually costs when the accounts are not yours.
Free · 45 Min · No Contracts
Turn These Insights Into Real Growth
Book a free Discovery Voyage with Keith Marsh — E320 Consulting's founder and fractional CMO. In 45 minutes, we'll audit your marketing, identify your top 3 growth opportunities, and give you a clear roadmap.
Whether you hire us or not, you'll leave with real, actionable insights.
Keith Marsh personally on every call · Ruskin, FL · (656) 565-8375
