GEO metrics for AI visibility

AI visibility metrics that matter

The four main AI visibility metrics pair mention rate with share of voice, and citation rate with prompt coverage. Two supporting measures, sentiment and AI referral traffic, tell you what those four numbers are worth in practice. The table below sets them side by side, with the formula for each and the trap to watch for.

Each metric belongs to a different person on your team. Mention rate and share of voice are brand numbers. They show whether the AI treats you as a real answer in your category. Citation rate is the content team's number, because citation tracking shows which of your pages the engine trusts as a source. Prompt coverage guides planning by identifying the questions where you need new or better material.

How much weight each one gets depends on your market. In a category with only a handful of credible suppliers, the AI keeps returning the same small group of brands, and share of voice becomes the number to watch. In a crowded consumer category, getting mentioned at all is the first hurdle. Before you compare any two of these numbers, make sure they come from the same prompt set and the same date range. If they don't, the comparison means nothing.

What GEO metrics are

GEO metrics are measurements of how present a brand or a web page is inside answers written by large language models. They count appearances and links across a sample of prompts. They also track descriptions. Classic search metrics count positions on a results page, which GEO metrics don't do.

The term comes from academic research. In "GEO: Generative Engine Optimization," presented at the KDD 2024 conference, Pranjal Aggarwal and co-authors scored a source's visibility by how much of an AI answer drew on it and where in the answer it appeared. The authors' affiliations spanned Princeton and Georgia Tech, along with the Allen Institute for AI and IIT Delhi. Their tests showed that content changes raise visibility by 40% in generative engine responses. They also found that what worked differed from one subject domain to another.

Commercial tracking turned that research idea into brand-level counts. Marketers wanted to know whether ChatGPT names their company. That's why most GEO reporting today is built on mentions and citations. You'll see the same measures sold as AEO metrics (answer engine optimization) or LLM visibility metrics. AI search metrics is another label for them. The labels differ but the arithmetic is the same.

Check any "visibility score" an LLM visibility tool gives you before you rely on it. A single blended number can mix mentions and citations with a sentiment measure, in proportions the vendor doesn't publish. Ask for the parts. If you can't rebuild the score from the raw answers, you can't explain a drop in it to anyone.

GEO vs SEO measurement

SEO measures a stable position on a ranked list and the clicks that position earns. GEO measures the probability that a brand appears in a written answer, even when that answer produces no click at all. Those two differences change almost every number you report.

While rankings in Google move slowly and look the same for most users at a given moment, AI answers are generated fresh for each prompt. They also draw on sources that don't line up with the search results. An Ahrefs study of 15,000 prompts examined ChatGPT and Gemini alongside Copilot and found that only 12% of cited links appeared in Google's top 10 for the same prompt. So a page-one ranking doesn't promise a place in the AI answer.

A Pew Research Center study of real browsing behavior found that users clicked a traditional result in 8% of visits when an AI summary appeared, compared with 15% when none did. Links inside the summary got clicked in 1% of visits. If you judge AI visibility by traffic alone, you miss most of what's happening.

Google's AI Overviews lean on ranked pages, as Ahrefs found that 76% of AI Overview citations came from top-10 pages. Your SEO data stays useful for that one surface. For standalone assistants, it isn't a stand-in.

How to measure GEO

To measure GEO, run a fixed set of prompts several times through the AI engines your buyers use. Log every answer, then calculate percentages over a rolling window. It's a sampling exercise, close to running a survey. A consistent sampling method keeps the numbers reliable despite the limits of AI mention tracking.

1. Write the prompt set. Start from the questions customers actually ask your sales and support teams. Group the prompts by topic and by buying stage, from "what is" questions to "best supplier for" questions.
2. Pick the engines. Choose the assistants your audience uses. At DevDay in October 2025, OpenAI CEO Sam Altman reported that ChatGPT alone had 800 million weekly users. Track every engine separately.
3. Repeat each prompt. One run is noise. ThriveStack's summary of University of St. Gallen research puts the minimum at seven runs per prompt per engine per day for a usable detection rate.
4. Log every answer. Record which brands appear and which URLs get cited. Also record how your brand is described. Keep the raw text so you can audit it later.
5. Calculate over a window. Report percentages over a rolling two- to four-week period, so one odd day doesn't swing the trend.
6. Match against analytics. Compare visibility changes with sessions from AI referrers and the conversions those sessions produce.

Rand Fishkin, co-founder of SparkToro, started out believing AI tracking was pointless. After running the numbers, he wrote that he now believes "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric." His study used 60 to 100 runs per prompt per engine for broad consumer categories. Narrow B2B categories settle with fewer runs, because the AI considers fewer brands.

You can collect this data in three ways. They differ in effort far more than in method.

Manual checks are useful for building intuition, but they can't produce reliable percentages, because no one types the same prompt 70 times a week. Whatever method you pick, ask how it handles repeats and whether it queries the consumer apps or the APIs. SparkToro's authors left open whether API calls show the same variation as manual prompts.

Mention rate

Mention rate is the number of AI answers that name your brand, divided by the total number of answers collected, multiplied by 100. It's the base metric for GEO.

For example, say you track 40 prompts in ChatGPT and run each one 7 times in a week. That gives you 280 answers. If your brand appears in 84 of them, your mention rate is 84 ÷ 280 = 30%. Calculate the same figure for each engine on its own and for each prompt group, because a 30% overall rate can hide 60% on "what is" questions and 5% on "best supplier" questions.

Benchmarks depend on the category. In SparkToro's headphone test, the four leading brands appeared in 55% to 77% of 994 answers to 142 different prompts. Across all the categories tested, the three most-mentioned brands averaged 64% in ChatGPT and 73% in Claude. Those are category leaders. A challenger brand at 20% in a crowded category is doing well.

Decide what counts as a mention before you start, and write it down. Does a product name count, or only the company name? What about a misspelling? Does a brand in a list of 12 count the same as a brand recommended alone? Keep every answer in the denominator, even when the AI refuses to name any brand. If you drop those, your rate looks better than it is.

Share of voice in AI

Share of voice in AI is your brand's mentions divided by the total mentions of every brand in your competitor set, across the same answers. While mention rate measures your presence, share of voice measures your portion of the conversation.

In the earlier example, your brand appears 84 times across 280 answers. Your four tracked competitors appear 70, 110, 56 and 30 times. Total mentions come to 350, so your share of voice is 84 ÷ 350 = 24%. One answer can name all five brands, which is why share of voice and mention rate rarely match.

The competitor list decides the result. Leave out a rival the AI names and your share goes up on paper. Keep an "other brands" bucket for names you didn't expect, and check it every month. When a newcomer keeps showing up there, add it to the set and restate the earlier periods, so the trend stays honest.

Share of voice is more closely associated with brand reputation. An Ahrefs study of 75,000 brands found that branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks. Ahrefs itself cautions against inferring causation from correlation, since strong brands earn both. Still, if your share of voice is low, you should look at how often third-party sites discuss you before you look at your own pages.

Citation rate vs position

Citation rate measures the share of answers that link to your domain as a source and is stable enough to track. Brand order changes so much between runs that an average position inside an AI answer tells you almost nothing.

The AI tools SparkToro tested returned the same list of brands in fewer than 1 in 100 runs. They returned the same list in the same order in fewer than 1 in 1,000 runs. Fishkin concluded that "any tool that gives a 'ranking position in AI' is full of baloney." If a dashboard shows you at "position 2.3" in ChatGPT, it's averaging noise.

Citation rate works differently from mention rate. An engine can name your brand without linking to you when its knowledge comes from training data. It can also cite your guide while recommending a competitor. Track both, and keep a short list of the URLs that get cited most. Those pages are the ones the engine trusts, and they tell your content team which formats to repeat.

Engines differ in how they cite. Ahrefs found that 28.6% of Perplexity citations pointed to pages ranking in Google's top 10, well above the other assistants. So a citation rate that's strong in Perplexity and weak in ChatGPT is normal. As the earlier Pew finding showed, a citation is a sign of trust but rarely brings a click.

Prompt coverage gaps

To measure prompt coverage gaps, list the prompts where your brand's appearance rate falls below a meaningful share of runs, then sort them by the brands that appear. Coverage is the share of tracked prompts where you clear a threshold you set, for example a mention in at least 1 run in 5. The gaps are everything below that line.

1. Map every tracked prompt to a topic and a buying stage.
2. Set the coverage threshold and mark each prompt as covered or not.
3. Divide covered prompts by total prompts to get your coverage rate.
4. For each uncovered prompt, record which competitors the AI named and which URLs it cited.
5. Sort the gaps by buying stage, so the prompts closest to a purchase sit at the top.

Of the three kinds of gap this process reveals, a content gap occurs when no brand appears because the AI has no good source to draw on; you can supply one. A prompt where competitors appear and you don't is a competitive gap. Your rivals are present on the third-party pages the engine reads, and you aren't. The third kind is a prompt where your page gets cited but your brand never gets named. That means the page explains the topic well but never ties your company to it.

Your prompt set limits all of this. When SparkToro asked 142 people to write their own prompts on the same headphone question, the prompts had a semantic similarity score of just 0.081. Real buyers phrase the same need in different ways. So write several versions of each core question. Every quarter, add new wording from sales calls and retire prompts nobody asks anymore. A coverage rate of 90% on a stale prompt set is a comfortable number, but it isn't a true one.

Reporting metrics to leadership

Leadership needs a small set of AI search metrics tied to competitors and revenue, shown as trends with honest ranges. The raw tracking data stays with the team. Executives already feel the shift toward AI answers described in Gartner's 2024 forecast, where Alan Antin, VP Analyst at Gartner, said that "GenAI solutions are becoming substitute answer engines," as they replace queries that once went to search engines. Executives need a way to see where their company stands in that shift.

Which AI search metrics to report to your CMO

Start your four-number AI visibility report to the chief marketing officer (CMO) with share of voice against named competitors, because executives most want to know whether the company is winning. Then add the mention rate trend for each engine and prompt coverage on buying-stage prompts. Complete the report with conversions from AI referral traffic. That last one links visibility to money. Semrush found that the average AI search visitor is 4.4 times as valuable as a traditional organic visitor when measured by conversion rate. Small visit numbers can still carry real revenue.

How to present the numbers

Show each metric as a line over rolling four-week windows, with one line per engine. Give a range or a margin next to each figure, and explain in one sentence why AI numbers move more than search rankings. When a number changes, point to the cause you can see in the logged answers, such as a new competitor or a page that started getting cited. A trend backed by an explanation gives you the context to answer questions in the boardroom.

What to leave out of the report

Exclude average position in AI answers because it averages noise. Single screenshots of an answer, good or bad, are also unsuitable because answers change between runs. Use vendor scores only when you can break them down into their components. If an executive asks what drove a change and you can't answer from the components, the metric is costing you credibility.

What to do next

Build a baseline before you change anything. Pick 30 to 50 prompts from real customer questions and run them across two engines for four weeks. Record mention rate and share of voice, along with citation rate and coverage. If you don't have a baseline, you can't show whether new content or PR work moved your GEO metrics, and leadership will treat any later gain as luck. Freeze that prompt set for the first quarter so your comparisons stay clean. Snoika audits websites and tracks brand visibility in Google and AI assistants, with a human review layer on top of its automated workflows. To see where your brand stands now, get a free AI visibility audit.