AI Search Visibility Metrics: The KPIs a Report Should Show, and How They Are Measured

What AI search visibility metrics are, and why one score is never enough

AI search visibility metrics show how often, how prominently, and how accurately your business appears in AI answers. They cover ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity, and Copilot, using questions buyers actually ask in your category. AI Search Engine Optimization records them as the baseline of every AI visibility audit, and reports them the way it reports rankings: with the questions listed, the platforms named, the dates fixed, and the competitors counted on the same questions. A single visibility score cannot carry that. It hides which platform named you, which question you were absent from, and whether the number came from one run or twenty.

Owners have already learned to distrust the single number. A forum thread this year asked, “Our AI visibility reports feel made up.” A marketing writer described a similar moment when a 37 percent visibility figure collapsed after three prompts were rerun by hand and produced different answers. A score you cannot rerun is a screenshot. This article is about the metrics that survive being rerun.

In the seven days to September 16, 2026, we asked 26 questions on six AI platforms. Our business was named in two answers and cited as a source in twelve, across three pages. The three most-named competing agencies appeared in 53, 42, and 35 answers. Those are the kind of numbers this page is about: counted, dated, and shown next to the competitors.

The metrics an AI visibility report should show, with their denominators

An AI visibility report should show eight metrics, and every one of them needs a denominator before it means anything. AI Search Engine Optimization reports each as a count over a stated set of questions and platforms in a stated week, never as a lone percentage. Mention rate shows whether the AI names your business. Citation rate shows whether it used your pages. Share of answers compares both with competitors on the same questions. Position and accuracy describe how you were presented, while the platform split and source mix show where you were absent and what the AI read instead. Attributed inquiries show whether any of that visibility reached the phone. The eight are listed below with what each cannot tell you, because a report that omits that column is selling the metric rather than explaining it.

A mention and a citation are different events and a report must count them separately. In the September measurement this site was named in two answers but cited in twelve, which means the AI platforms read its pages to build answers far more often than they named the firm in them. A business in the reverse position, named often and never cited, is living on third-party sources it does not control. The difference between AI mentions and citations decides what the next piece of work is.

For a local business the questions are the service-area ones a buyer asks: who does this work near a named town, who handles this problem in this county, which company to call for this in this neighborhood. The metrics are the same; only the question set changes, and a report that tracks national questions for a business that serves one county is measuring the wrong market.

How AI visibility is actually measured: a report you can read

AI visibility is measured by asking a fixed set of questions of each AI platform, on a fixed schedule, and counting which businesses and which pages the answers name and cite. We publish our own measurement so prospects can inspect the method before buying an audit. The set contains 26 buyer and research questions, asked in the United States on ChatGPT, ChatGPT with web search, Google AI Overviews, Google AI Mode, Gemini, and Bing Copilot, over the seven days to September 16, 2026. Every firm named in any answer was counted, 247 brand mentions in all, and every source the answers cited was recorded. The table below is that week’s result for this firm and its field, and the full record is on the AI answer benchmark page, including the 26 questions in full and a plain statement of what the numbers do not claim.

How Rank Intelligence Engine runs the measurement

Rank Intelligence Engine is the proprietary system AI Search Engine Optimization uses for this work, and its visibility measurement runs in five steps, each with a reason.

1. Write the question set from the buyer’s side. The questions are the ones a prospect asks before hiring, split between buyer questions where being named is the result and research questions where being cited is the result, because the two are won by different pages.
2. Ask every platform, every week, the same way. The wording, the location and the platform list do not change between runs, because a number is only comparable to a number produced the same way.
3. Count every name and every source, not just yours. The competitors and the cited pages are recorded from the same answers, because a good score is relative and the source mix is the map of what to build next.
4. Keep the answers. Every answer is stored with its date and platform, because the count is only evidence if the answer behind it can be shown.
5. Report ranges and trends, per platform. The week’s counts are reported with the previous weeks beside them and the platform split visible, because a single week on a single platform is a sample, not a standing.

Phil Belleville, who has worked in SEO since 1995, reads the result and decides the priority. The system supplies the counts; the decision about which absence costs the business the most is a person’s.

What a good AI visibility score is

A good AI visibility score is one that beats the named competitors on the questions that produce inquiries, measured the same way in the same week. We do not use the 0-to-100 bands published by tracking tools. A score of 30 can be strong in a thin market and invisible in a crowded one. The band also hides which question and platform produced the points. In the September measurement, being named in two answers against a leader’s 53 is a weak standing in absolute terms, and it is also the exact distance the work has to close, question by question. A useful score shows how far you are from the firm the AI currently prefers on the questions that matter. That gives next month’s work a specific target instead of a total to admire.

• A score without the competitors on the same questions is a mood, not a benchmark.
• A score without the platform split hides the platform where the business is absent.
• A score without the question list cannot be aimed at anything.

Why the numbers change between runs, and how to sample so they still mean something

AI visibility numbers change between runs because AI answers are generated fresh each time, from live retrieval that itself changes, on models that are updated without notice. Two identical prompts an hour apart can name different firms; a slight rewording can swap the whole list. AI answers vary, so we measure around that variation with four rules. Use the same questions, worded the same way, every time. Run every platform in the set each week. Keep a fixed seven-day window, and report trends across weeks as a range rather than presenting one swing as a standing. A single run of a single prompt is a screenshot. A fixed set, rerun on a schedule, with the previous runs beside it, is a measurement.

What that looks like in practice

The September measurement is one seven-day window. The week that follows is measured identically, and the report shows both. If the two-answer count becomes five, the report says five against the previous two, on which platforms, on which questions, and whether the competitors moved too. If it becomes zero on one platform, the report says which platform and keeps the other five in view. The number that matters is never the one from the latest run; it is the direction across the runs, read one platform at a time.

What makes an AI visibility report evidence instead of a screenshot

An AI visibility report is evidence when a reader could rerun it. AI Search Engine Optimization applies eight checks before a report leaves the building, and an owner can apply the same eight to any report they have been handed. The questions are listed in full. The platforms are named. The window is dated. Every metric carries its denominator. The competitors are named and counted on the same questions. The answers are kept and can be shown. The platform split is visible. And the report states what it does not claim. A report that fails three of those is a sales document; the free scores that arrive by email usually fail six. The forum poster was not wrong to distrust the reports in front of him. Single-run numbers with no questions, dates, or competitors cannot be checked.

• Listed questions and named platforms, or the number has no subject.
• A dated window and a stated denominator, or the number has no scale.
• Named competitors and kept answers, or the number has no comparison and no proof.
• A platform split and a statement of what is not claimed, or the number is hiding something.

What Search Console and your analytics can and cannot show

Google’s own tools now show two official pieces of AI visibility, and a report should use both while stating their limits. Search Console’s generative AI performance report, announced on June 3, 2026, shows how often a site’s pages appeared inside AI Overviews and AI Mode, by page, country, device and date. It carries no query data and no click data, so it can say a page was shown inside an AI answer and cannot say for which question or whether anyone clicked. Analytics can isolate sessions arriving from AI assistants, although the number is usually small for an owner-sized business. This website recorded one ChatGPT session in the thirty days to September 18, 2026, while its pages were cited in twelve AI answers. That gap is the point. AI visibility mostly does not arrive as a click, so a report that leads with AI referral traffic is measuring the smallest part of the effect.

Google’s guide to AI features in Search, updated July 10, 2026, explains that gap: an AI answer issues several related queries and assembles the result from the pages it retrieves. Your page can shape the answer without earning a visit. Search Console shows the appearance; the prompt-level measurement above shows the answer; only the inquiry shows whether it mattered.

The metric an owner feels: inquiries, attributed

The metric a business owner feels is the inquiry, and AI Search Engine Optimization treats every visibility number as a leading indicator of it. Two signals make the connection honest. The first is self-reported attribution: a “How did you hear about us?” question on the contact and booking forms, with ChatGPT, Perplexity, Gemini, Claude and Copilot as options beside search and referral. It catches the buyer who researched in an AI assistant and then typed the business name into a browser, the journey no analytics tool can see. The second is the count of inquiries and their quality in the same weeks the visibility counts were taken, so the two can be read together. Citations rising while inquiries stay flat points to a page that earns the citation but not the call. That is a problem worth fixing. Citations rising on their own, with no business outcome attached, tell you much less. The reviews owners leave about the alternative describe the difference exactly: rankings moving while they are still waiting for the leads to come in.

Expect a decision at the end of the report

An AI visibility report should end in a decision, and AI Search Engine Optimization writes the decision into the last page of every one. The September measurement produced a clear example. Google’s documentation was cited in 21 of the 26 questions. The three leading agencies were named in 53, 42, and 35 answers, while our firm was cited on three pages and named in two. The decision that follows is not “raise the score.” It is which questions to win first, which page has to become the one the AI reads for each of them, and which platform to watch for the change. A useful review separates what changed, what the evidence shows, and what should happen next. Some changes can be verified immediately. Their effect on citations and inquiries takes weeks of consistent measurement. What it costs to run this as an audit is set out on the pricing and proposals page, and how long before the counts move is on the progress timeline.