AI search visibility metrics and KPIs that hold up

AI search visibility metrics show whether assistants name your product, use your pages, and draw from the sources around your category. This matters this year because you need a stable way to judge visibility while answers, engines, and cited sources vary.

The useful question is which part of the answer you can measure and improve. A single mention can look reassuring while your own pages remain absent from the evidence. Here are six metrics for an AI search visibility dashboard, with the job each one does and the point at which it turns into reporting theatre.

We do not have an external vendor quote in the sources for this post. The only public figure we can point to is our own: CiteGraph's research page says we have recorded 45,309 citations across 7,260 AI answers. That figure shows the scale of our sample. It does not show that any particular company is visible, that a cited page drove a sale, or that the next answer will behave the same way. Treat your dashboard in the same spirit. It exposes the path to an answer instead of producing a score to display.

Mention rate

Mention rate is the share of sampled answers that name your product or company. It is the obvious first KPI to check because buyers recognise it immediately: the assistant either included you or it did not. In our engine sample, Perplexity named the scanned product in 8.2% of answers, Gemini in 6.8%, Claude in 6%, and the ChatGPT family in 4.4%.

Of the six here, mention rate is the one most often reported as if it were the whole picture, and the one least likely to change even when your actual position does. It is useful for checking whether recognition exists across a fixed question set, but it gives a weak diagnosis. A product can be named without a page from its site being cited. An assistant can also cite a useful page about the category without naming the product in the answer.

That makes mention rate a signal worth watching, not a verdict on your visibility. Use it to spot broad movement by engine and question type. Use citation data to decide which page to rewrite, which comparison to publish, or whether a visibility programme worked.

The engine differences are large enough to keep separate rather than blend into one average. A blended rate can hide the fact that one assistant names you while another does not. Keep the question set and scan method stable, then read the change by engine.

Citation rate

Citation rate measures how often an answer cites your domain or one of your pages. It answers a different question from mention rate: did the assistant use your material as evidence? This is often the better metric when your work is content, technical access, or page selection rather than brand recognition.

The sample gives a useful warning about the gap. The scanned product's own site contributes 1.5% of recorded citations in a category where roundups and competitor pages dominate the evidence, as the breakdown below shows. A company can be present in the answer ecosystem while its own site contributes little of the supporting material.

Citation rate does not promise traffic. It tells you that a page was selected in the sampled answer, not why the selection happened or whether the reader followed the link. Still, it gives you a page-level work queue. Inspect the questions where your domain appears, then compare the cited page with the question the assistant was trying to answer.

If you need the underlying method, the measurement method guide sets out how to record answers, sources, engines, and questions without treating a single result as a trend. How to measure AI visibility walks through one run, from question set to margin. That discipline matters more than a polished score.

Citation share

Citation share measures your portion of the cited sources in a defined category, question set, or competitor group. Mention rate asks whether you entered the answer. Citation share asks how much of the evidence set you occupy once you are there. Share of voice in AI answers, and how to compute it covers the brand-mention version, with citation share as a second measure.

This metric is useful when several products appear in the same answer. A small mention rate can coexist with a strong citation share on the questions where you do appear. The reverse can also happen: your name appears often, but most cited evidence comes from other domains. That distinction helps you separate brand recall from source ownership.

Do not report citation share without naming the denominator. Say whether the share covers all recorded citations, a category, a fixed group of buyer questions, or a selected set of competitors. Our wider sample includes 7,624 distinct domains and 16,506 distinct pages, drawn from 442 distinct buyer questions. Those are different ways to cut the evidence, so they will produce different shares.

Citation share loses its value when the comparison set is chosen after the result. Keep the category and question set fixed before you inspect the outcome. Store the cited URLs, rather than just the percentage, so someone can check whether the share came from useful buyer pages or a narrow cluster of easy questions.

Engine coverage

Engine coverage is the number and spread of assistants in which your product or domain appears. It tells you whether your visibility survives a change in the system answering the question. That matters because the engines in this sample cite different numbers of sources per answer: Perplexity averages 5.3, Claude 4.8, Gemini 4, and the ChatGPT family 3.4.

Coverage should be read alongside mention rate and citation rate. Being named on one engine is a signal. Being cited across several engines is stronger evidence that the underlying source is available to more than one retrieval system. It still does not prove universal visibility, because the sample covers 60 scans of 41 sites in 44 categories, rather than every market or every question.

The practical KPI is coverage by engine and question intent, rather than a single count of engines. Separate buyer questions about alternatives, pricing, use cases, and implementation if those are part of your tracking plan. A brand that appears across engines for one easy query has less useful coverage than a brand appearing across several kinds of buyer question.

For a quick external check, use the free AI visibility checker, then compare what it shows with your own saved answer set. A checker is a starting point. Your recurring sample is the record you can use to judge a change.

Engines that cite fewer sources per answer leave less room for any one page, which is worth remembering when you judge why a page dropped out.

Citation depth

Citation depth measures how much attention a cited page receives in the answer set. At its simplest, record whether a page was cited once, then how often it recurs across the questions and engines you track. You can also record the number of sources in each answer, because a page cited in a short source set has a different position from one buried among many references.

This metric is for prioritisation. A page that recurs across relevant questions is a better candidate for maintenance than a page that appeared once by chance. A page that is cited but never supports the part of the answer a buyer cares about needs a different review from a page that supplies the product fact the assistant repeats.

Citation depth does not measure answer rank. Source order is not a reliable measure of commercial value, and recurrence is not proof that a page is correct. Use the metric to find pages worth reading, then inspect the wording, claims, and question match yourself.

Record the source count alongside the cited page. That gives you context for each citation and helps you compare pages across engines with different answer structures.

Source mix

Source mix is the share of citations coming from each page type. It tells you where assistants are finding the evidence that shapes answers, which is more actionable than a general visibility score. In the recorded sample, listicles and roundups make up 67.5% of citations, competitor product pages 24.4%, community threads 3.4%, the scanned product's own site 1.5%, documentation 1.1%, social posts 0.8%, review platforms 0.8%, YouTube 0.4%, news 0.1%, and Wikipedia 0%.

The pattern gives you a publishing and outreach question. If your category is represented mainly by roundups, inspect the pages that already do the comparison work. If competitor pages carry a large share, make sure the facts buyers need are stated plainly on pages that can be compared. Source mix shows where the answer's evidence currently comes from. It does not prescribe a page type for every category.

A percentage by itself proves nothing. Without the cited URLs and the question set behind it, source mix is just a chart. Keep the page type labels, the domains, and the exact questions together.

One objection is fair: six metrics may be too many for a small team. A founder can waste a week building a dashboard when a handful of saved answer checks would reveal the same problem. That is true when the metrics become a reporting exercise. It is less true when each one answers a different decision: mentions show recognition and citations show whether that recognition rests on your own material. Citation share and engine coverage show how far that evidence spreads, while depth and source mix show where to spend the next hour of work. Keep the metric that answers the decision in front of you.

A short check you can run today

You can make a useful first pass in under an hour:

1. Write down the buyer questions you already use in sales calls or content planning.
2. Run the same questions through the engines you care about and save the full answers with their cited pages.
3. Mark whether your product was named, whether your domain was cited, and which page appeared.
4. Label each cited page by type, such as roundup, competitor page, community thread, or own site.
5. Compare the result by engine and question intent, then choose one page or source gap to investigate.

Keep the first sample as your baseline. Do not change the questions after seeing the result, or the comparison will become a small experiment in self-comfort. The how CiteGraph measures page sets out the measurement boundaries and the fields worth retaining.

We measured one sample, rather than the whole market, and we do not know what a wider scan would show. The evidence here comes from 60 scans of 41 sites in 44 categories and 442 buyer questions. It can show patterns in that sample, but it cannot forecast every answer, category, engine update, or commercial outcome.

The next question to ask is: which buyer question should your page answer better, and which source currently supplies that answer? The answer lives in your own saved citations. See the weekly research on who owns AI answers for how we keep ours.