AI Visibility: What It Means, How It Is Measured, and the Five Numbers That Matter

AI visibility is how often, and how favourably, your business shows up when someone asks an AI assistant a question you would want to be the answer to. It is the number behind every "answer engine insights" dashboard, and it is the number most businesses have never measured, because until recently there was nothing to measure it with. Now there is a tracker for every budget, each with its own score, and the scores disagree.

This post is about what the score is actually made of, why two tools give different numbers for the same brand, and the five numbers you can track yourself in a spreadsheet, which are the ones that drive decisions anyway.

What the dashboards measure

Profound, Peec, Semrush's AI toolkit, Ahrefs' Brand Radar and the rest all work the same way underneath. They take a list of prompts, run each one through the assistants' APIs on a schedule, parse the answers, and count. What they count varies by product, but the metrics come from a short list:

• Mentions, or visibility rate: the share of answers that name your brand at all.
• Citations: the share of answers that link to one of your URLs as a source. A brand can be mentioned without being cited (the assistant knows you from training data) and cited without being mentioned (your comparison page was a source for a competitor's recommendation).
• Share of voice: your mentions as a fraction of all brand mentions in the answer set, against a competitor list you choose.
• Position: whether you are named first, third or last. Being first in an answer is roughly what being first in a ranking used to be.
• Sentiment: whether the answer describes you positively, neutrally or with a caveat, scored by another language model.
• Prompt volume: a modelled estimate of how often each prompt is asked, covered in its own post.
• Referral traffic: visits that arrive from the assistants, which comes from your analytics rather than from the assistants.

None of these is wrong. The trouble starts when a dashboard rolls them into one score, because the score hides what moved.

Why the numbers wobble

Ask ChatGPT the same question twice and you will get two different answers. Ask it from London and from Austin, with web search on and off, from a fresh account and from one with memory, and you can get five. The assistants are not deterministic, their web search returns different pages on different days, and every vendor samples the answers from a different location on a different schedule with a different prompt phrasing. Two tools tracking the same brand on the same prompts will report different visibility rates, and both are correct about the sample they took.

That is why the trackers run prompts repeatedly and report rolling averages, and why a week-to-week change of a few points means nothing. The signal is in the trend over months and in the big, stable facts: you are never cited for this cluster of questions; this competitor is named in every answer; the assistant keeps getting your price wrong. Those do not wobble.

The five numbers that matter

If you track nothing else, track these, and track them against a fixed list of questions so the numbers are comparable over time.

1. Mention rate on your twenty questions

Write down twenty questions a real buyer in your category asks: the "best X for Y" ones, the "X vs Y" ones, the "how much does X cost" ones, and a few that mention you by name. Ask each one of ChatGPT, Gemini, Perplexity and Claude with web search on. Count the answers that name you. That fraction, out of eighty, is your visibility rate, and it is the same thing the dashboards show, sampled once instead of weekly.

2. Citation rate

Of those eighty answers, count the ones that link to a page on your site. This is the number that content work moves. Mentions can come from training data you cannot influence this quarter; citations come from pages the assistant found and chose, which you can.

3. Who is named instead of you

For every answer where you are absent, write down who is there. After twenty questions you will have a competitor set of five to ten names, and it will not match the one in your marketing plan. The brands that keep appearing have sources the assistants trust, and looking at what those sources are (review sites, comparison pages, trade articles, forum threads) tells you where you need to be.

4. What the assistants get wrong

Read the answers that do mention you and note every factual error: an old price, a product you discontinued, a location you do not serve, a competitor's feature attributed to you. This is the metric the dashboards score as "sentiment" and it is more useful as a list than as a number, because each item is a fix: a plain facts page on your site, a corrected listing on the source the assistant quoted. For a PR team this list is the whole job; see AI search for PR and brand teams.

5. Referral sessions and what they do

In your analytics, add the assistant referrers (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai) as a channel and watch sessions and conversions from it. The absolute numbers will be small compared with Google, which is normal; what matters is that they convert at a different rate, usually higher, because the visitor arrived with a recommendation rather than a search. Google AI Overview clicks are the gap here: Search Console counts them as ordinary Google clicks, so you cannot separate them.

How to run it yourself

Two hours, once a quarter. A spreadsheet with twenty rows for the questions and four columns for the assistants; in each cell, whether you were mentioned, whether you were cited, who else was named, and anything wrong. Use a browser with no history or memory, turn web search on, and run it from the country your buyers are in. Keep the questions fixed between runs. Three quarters in, you have a trend that is as reliable as any tracker's, because the big facts do not wobble.

This is the baseline SEO Stuff runs for every Done-For-You order: twenty questions, four assistants, every answer recorded on your dashboard before the work starts, and the same twenty run again after delivery. The strategy guide is built from numbers 3 and 4, the competitor set and the error list, because those say what to write and where to be placed.

When a tracker earns its subscription

The spreadsheet stops being enough when you have more questions than you can run by hand, several markets or languages, or a brand that assistants discuss often enough that a weekly check catches problems a quarterly one would miss. That is the case for the trackers, and for Rightcited, which the SEO Stuff team runs: fifty buyer questions a week across ChatGPT, Claude, Gemini and Perplexity, with the fixes done rather than charted, at $499 a brand a month. The comparison with Profound covers where a dashboard fits and where it does not.

Before any of that, find out where you stand. The free SEO + AI search audit includes what the assistants currently say about your category and who they name instead of you, which is numbers 1 and 3 above, done for you, with a report in 48 hours.