Ask a marketing team how visible their brand is in AI search, and you'll usually get a single number. Sometimes it's called a visibility score. Sometimes it's share of voice. Either way, it's meant to capture how often ChatGPT, Gemini, or Perplexity brings up the brand.

Single numbers are appealing. They fit on a slide, they're easy to track, and they give everyone something to rally around.

They also hide almost everything that matters.

An answer engine can name your brand without linking to you. It can cite your website once in each of ten answers, or five times in a single answer and never again. It can describe you favorably or unfavorably. These are different behaviors, and they call for different responses. A single score blends them together, so when it moves, you have no way of knowing which behavior changed, or what to do about it.

A better approach is to track a small set of metrics, each of which answers one specific question, and to know which question you're asking before you look at the numbers.

Links and names

Generative search engines surface brands in two ways.

A citation is a link. The engine attributes a claim in its answer to a URL, and that URL belongs to a domain. If the engine cites a page on your site, your domain earned a citation.

A mention is a name. The engine refers to your brand, product, or company in the text of its answer, whether or not it links to anything you own.

The two often diverge. An engine might recommend your product by name while citing a review site as its source. That's still visibility, but it's a different kind, and (as we'll see) the gap between the two is one of the most useful signals you can measure.

Five metrics, five questions

Citation Share answers the question, Of all the sources the engine cited, what portion were mine? Every URL an engine cites adds one citation to the pool for that run, and Citation Share is the fraction of the pool that belongs to your domain. It measures density: how much of the engine's evidence base you occupy.

Citation Coverage answers, In how many answers did my domain appear at all? It's the fraction of prompts for which your domain was cited at least once. It measures breadth.

Mention Count answers, How many times was my brand named? It's a raw volume figure across every response in the run.

Mention Coverage answers, In how many answers was my brand named at least once? It's the mention counterpart to Citation Coverage.

Sentiment answers, When my brand is mentioned, how is it framed? Each mention is classified as positive, negative, or neutral based on the text around it, and each category is reported as a percentage of classified mentions. Net Sentiment (positive minus negative) is the headline figure.

A worked example

Suppose we run 120 prompts about a product category on a single platform. That platform cites about 20 sources per answer, so the run produces 2,400 citations in total.

Our domain is cited 60 times, so its Citation Share is 60 ÷ 2,400, or 2.5%. Those 60 citations are spread across 30 different answers, so its Citation Coverage is 30 ÷ 120, or 25%. And our brand is named in 54 answers, so its Mention Coverage is 54 ÷ 120, or 45%.

Each number tells us something different. A 2.5% share sounds small, but in a category where the engine draws on hundreds of domains, it might place us among the leaders. A 25% coverage rate says we show up in one answer out of four.

The most actionable insight comes from putting two of the numbers side by side. Our brand is named in 45% of answers but cited in only 25%. In 20 percentage points' worth of answers, the engine is talking about us and sending the reader somewhere else for the details. That gap is a content opportunity: The engine already associates us with the topic. It's simply finding better evidence elsewhere.

A single blended score would have hidden all of that.

Density and breadth

Citation Share and Citation Coverage are related, but they can tell very different stories.

A domain can earn high share with low coverage by being cited heavily in a narrow set of answers. (Think of a buying guide that an engine leans on for five citations whenever someone asks about one particular feature.) A domain can also earn high coverage with low share by appearing once in many answers without ever dominating any of them.

These call for different strategies. Low coverage suggests the engine doesn't consider you relevant to much of the topic, so the fix is breadth. Low share with decent coverage suggests the engine knows you're relevant but prefers other sources for most of its claims, so the fix is depth and authority on the questions where you already appear.

The trouble with denominators

Two definitional choices matter more than they might seem.

First, coverage metrics divide by the number of prompts sent, not by the number of answers that happened to contain citations or mentions. Some platforms return answers with no citations at all. (On SearchGPT, that rate varies by topic, and in our data it has exceeded 15%.) Drop those answers from the denominator, and the same brand looks more visible on platforms that cite less often. Keeping the denominator fixed at prompts sent measures every platform against the same base, and it lets us compare how the same prompt performs across platforms.

Second, cross-platform figures should be an average of per-platform values, not a pooled count. Platforms differ enormously in citation volume. In our research, Gemini returned roughly 40 citations per answer, Perplexity about 20, and SearchGPT about 6.

Consider what pooling does. We run 100 prompts on Gemini and 100 on SearchGPT. Gemini produces 4,000 citations, and SearchGPT produces 600. Our domain earns 200 citations on Gemini (5% share) and 90 on SearchGPT (15% share). Pool the counts, and our cross-platform share is 290 ÷ 4,600, or 6.3%, which is nearly identical to the Gemini figure. SearchGPT barely registers. Average the two shares instead, and we get 10%, with each platform counting equally.

The pooled number is really a Gemini number wearing a cross-platform label.

Why we don't report share of voice

Share of voice is a familiar marketing metric, so its absence deserves an explanation. It divides your mentions by the mentions of your brand plus a set of competitors, which means the result depends on which competitors are in that set. Add a rival, and your share drops, even though nothing about the AI's behavior has changed. Because the competitor set is a choice rather than a fixed quantity, the number can't be compared across time periods, analyses, or companies. Mention Count and Mention Coverage avoid the problem because they don't depend on who else you decide to track.

Start with the question

The practical rule is to start with the question and then pick the metric.

If you want to know whether AI engines are using your content as evidence, look at Citation Share and Citation Coverage. If you want to know whether they're recommending or discussing your brand, look at Mention Coverage and Mention Count. If you suspect they're talking about you without sourcing you, compare Mention Coverage to Citation Coverage. If you care how you're being characterized, look at Sentiment. And if you want to know where you're strong or weak, look at any of these filtered by platform, prompt type, or buyer intent.

One caution applies to all of them: Every one of these metrics is an estimate. Answer engines are stochastic, so the same prompt submitted twice can cite different sources and frame a brand differently. A Citation Share of 2.5% is the value observed in one sample of answers, and a second run would produce a somewhat different figure. Each metric therefore needs a confidence interval telling us how far it could plausibly move. That's the subject of the next post in this series, AI Visibility Isn't a Number. It's a Range.

That's why IQRush reports these five metrics separately instead of rolling them into a single score, computes every percentage per platform before averaging, and leaves share of voice out entirely. Each number answers its own question, so when one moves, you know which behavior changed.

The metrics tell us what to measure. The intervals tell us how much to believe what we've measured.

References

In this series

• AI Visibility Isn't a Number. It's a Range.

Research papers

Sielinski, R. (2026). Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement. arXiv:2603.08924.

Sielinski, R. (2026). From Stochastic to Stable: Rank Stability and Structural Sufficiency in AI Visibility Measurement. arXiv:2607.10341.