# How to Measure GEO: The 5 Metrics That Track AI Visibility

> Five metrics for AI visibility: citation frequency, share of voice, platform coverage, sentiment and attribution, plus how to set a baseline.

URL: https://llmagnet.com/blog/how-to-measure-geo  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-06-23  
Updated: 2026-11-10  
Category: Guides

Measure GEO with five numbers: how often you appear, your share against competitors, your coverage across engines, how you are described, and which pages get cited. Run each prompt several times and compare against your own baseline, because AI answers vary from run to run. Analytics and rank trackers show only part of this.

## Contents

1. Why SEO Metrics Only Show Part of It
2. Metric 1: Citation Frequency
3. Metric 2: AI Share of Voice
4. Metric 3: Platform Coverage Score
5. Metric 4: Sentiment and Framing Score
6. Metric 5: Content-to-Citation Attribution
7. Building Your Baseline
8. A Worked Example: One Baseline Round
9. What Changes by Business Type
10. A Measurement Cadence
11. Common Mistakes
12. Start This Week
13. FAQ

## Why SEO Metrics Only Show Part of It

Google Search Console now has [generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports), [available to all sites](https://support.google.com/webmasters/answer/16984139?hl=en) since August 31, 2026. They show impressions in AI Overviews and AI Mode by page, country, device and date. They report impressions only, cover Google only, and don’t show whether your brand was named.

Read those reports with care. The chart counts two results from the same site in one AI feature as one impression, while the table by page counts by page, so the totals can differ. The newest data can be preliminary. The report may also not appear if a site hasn’t had enough impressions in these features, so small sites may have only prompt runs to go on.

Analytics tools see visits that arrive from an AI product. They can’t see the buyer who reads an answer and types your name later, or who never visits.

Be careful with claims about AI traffic value. [Siege Media’s analysis](https://www.siegemedia.com/?p=45564) of 78 sites (agency data, January to May 2026, first-party GA4) found a median AI-to-organic conversion ratio of 1.26x, and 1.05x for B2B SaaS. AI traffic was roughly 0.2% to 3% of sessions on most sites. Measure visibility because it shows how buyers meet your brand in answers, and test the revenue case on your own data.

![Median AI-to-organic conversion ratio by vertical in Siege Media’s analysis: finance 1.67 times, consumer and local services 1.29, travel and real estate 1.27, B2B SaaS 1.05, ecommerce 1.01.](https://llmagnet.com/images/blog/charts/measure-siege-verticals.png)

*Source: Siege Media, first-party GA4 data, January to May 2026. Agency data. 1.0x means the two channels convert equally. Site counts per vertical weren’t published.*

Siege Media also found AI matched or beat traditional conversion on 72% of sites, but out-earned it per session on fewer than half of the sites with reliable revenue tracking. A good conversion ratio and a good revenue story aren’t the same thing, so check both before quoting either to your CFO.

No single source shows the whole picture. Each of these answers a different part of the question.

| Source | What it sees | What it misses |
| --- | --- | --- |
| **Search Console AI reports** | AI Overview and AI Mode impressions by page, country, device and date | Other engines, and whether you were named |
| **Analytics** | Visits and conversions from AI referrers | Readers who never click, or who search your name later |
| **Bot logs** | Which crawlers fetched which URLs | Whether anything was used in an answer |
| **Prompt runs** | Who gets named and cited, by engine | The exact wording your buyers use |

![The five GEO metrics and the question each answers: citation frequency, how often do we appear; AI share of voice, how do we compare; platform coverage, on which engines; sentiment and framing, how are we described; attribution, which pages earn it.](https://llmagnet.com/images/blog/charts/measure-five-metrics.png)

*Metric names and framing are our own.*

## Metric 1: Citation Frequency

In a representative set of prompts, how often does AI name or cite you? Record mentions and linked citations separately, since an answer can name you without linking to you, or link to you without naming you.

**How to calculate it:**

1. Build a set of 50 to 100 prompts: category questions, comparisons, problem-solving and feature questions.
2. Run each one several times on your target engines: at minimum ChatGPT and Perplexity, ideally also Google AI Mode, Claude and Gemini.
3. Count the responses that mention your brand, and those that cite your domain.
4. Citation frequency = (responses citing you ÷ total responses) × 100.

**Benchmarks:** we found no reliable public benchmark for this metric. Compare yourself with your own baseline and your named competitors on the same prompts.

In a [SparkToro study](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/) of 2,961 runs, there was less than a 1 in 100 chance of ChatGPT or Google AI giving the same brand list in any two of 100 runs. The authors recommend tracking how often a brand appears rather than its position, and running each prompt 60 to 100 times where you can. The study wasn’t peer-reviewed, and a tracking vendor took part.

Keep those two counts apart in your reporting. In [AirOps’ vendor data](https://www.airops.com/report/the-2026-state-of-ai-search), only about 28% of answers included brands that were both mentioned and cited, so the numbers often diverge.

### How many runs is enough

A presence rate is a percentage, and every percentage from a sample has a margin of error.

![Margin of error at a 50 percent appearance rate, by number of runs: 10 runs plus or minus 31 points, 30 runs 18 points, 60 runs 13 points, 100 runs 10 points, 400 runs 5 points.](https://llmagnet.com/images/blog/charts/measure-margin-of-error.png)

*Our own calculation: 95% interval, normal approximation. Real uncertainty is larger, because repeat runs aren’t fully independent.*

If your true rate is 50%, ten runs could plausibly show anything from about 19% to 81%. That is why SparkToro’s 60 to 100 runs per prompt is a sensible target, and why a handful of runs can’t support a claim about one prompt.

Few teams can afford that across every prompt. The practical compromise is to report the set as a whole. Fifty prompts on three engines at five runs each gives 750 responses, enough to track the overall rate even though no single prompt is well measured. Don’t draw conclusions from one prompt’s movement unless you’ve run it many times.

Prompt wording adds a second layer of uncertainty. Across 142 prompts that SparkToro’s volunteers wrote for the same need, the average semantic similarity was 0.081. Your list is a sample of how buyers ask, so write several phrasings of your important questions. Variation also depends on category: SparkToro saw far more consistency for SaaS cloud providers than for sci-fi novels.

## Metric 2: AI Share of Voice

Citation frequency shows how often you appear. Share of voice shows how that compares with competitors. On the same prompt set, count mentions of each brand and calculate each brand’s percentage of all mentions.

If 200 responses produce 500 brand mentions and you have 75, your share of voice is 15%.

Fix the competitor list before you start, and run the metric before and after a major content or PR push so you can see any change.

Decide whether the denominator is your named competitors or every brand that appears, and don’t switch later. Counting only named competitors produces a higher share and hides new entrants. Log any unfamiliar brand that shows up repeatedly, and review the list each quarter.

## Metric 3: Platform Coverage Score

Engines often cite different sources. [Profound](https://www.tryprofound.com/blog/citation-overlap-strategy) (vendor data, 100,000 prompts run on both engines, published July 2025) found 11% of cited domains appeared in both ChatGPT and Perplexity. A strong result on one engine tells you little about another.

**How to calculate it:** run the same 30 to 50 prompts across five engines and record citation frequency for each. Your score is the number of engines where you pass a threshold you set, divided by the number tested.

For example, with a 15% threshold, a brand above it on Perplexity only scores 1 out of 5. Choose the threshold from your own baseline and keep it fixed.

The score treats every engine equally. If one engine matters far more to your buyers, report it separately rather than adding weights you’ll struggle to defend.

## Metric 4: Sentiment and Framing Score

Being named isn’t enough if you are named as the second choice or with a warning. In SparkToro’s data, one hospital appeared in 69 of 71 ChatGPT responses but was the top mention in only 25.

![City of Hope in ChatGPT responses in SparkToro’s study: named in 69 of 71 responses, and the top mention in 25 of 71.](https://llmagnet.com/images/blog/charts/measure-named-vs-first.png)

*Source: SparkToro, 2,961 runs from 600 volunteers, November and December 2025. Not peer-reviewed, and a tracking vendor took part.*

For each response that names you, label the mention as one of these: primary recommendation, secondary mention, neutral mention, or negative or qualified mention. Read a sample by hand, because comparative wording is easy to misclassify.

Track primary recommendation rate separately. High frequency with a low primary rate points to a positioning problem.

SparkToro also warns that position within an answer is unreliable, with identical order appearing about once in 1,000 runs. So read primary recommendation as a rate across many responses, and never as where you landed in one.

## Metric 5: Content-to-Citation Attribution

This metric shows which of your pages earn citations. When a response cites your domain, record the exact URL. Then group the URLs by topic and format, and compare them with pages on the same topics that aren’t cited.

Use the links engines actually show. We wouldn’t treat a model’s account of where it got something as evidence, because the explanation is generated too.

For Google, the Pages view in Search Console’s generative AI report adds impressions by URL. Run this audit quarterly, and use it to decide what to update first.

## Building Your Baseline

Without a baseline you can’t separate normal fluctuation from the effect of your work. A workable minimum:

- **50 prompts** covering your category, use cases and competitors
- **3 engines:** ChatGPT, Perplexity and Google AI Mode
- **3 competitors** tracked alongside you
- **Two rounds seven days apart,** with several runs of each prompt per round

Calculate how much your own numbers move between the two rounds, and treat that as your noise level. Don’t borrow someone else’s figure. [AirOps’ vendor data](https://www.airops.com/report/the-2026-state-of-ai-search) found only 30% of brands stay visible from one answer to the next and 20% across five runs, so a small change in a small sample may mean nothing.

## A Worked Example: One Baseline Round

The company and figures below are made up for illustration. Quillon runs 50 prompts on three engines, four times each, for 600 responses, and tracks three competitors.

| Metric | Calculation | Result |
| --- | --- | --- |
| Citation frequency | 84 responses cite Quillon, out of 600 | 14% |
| Mention rate | 132 responses name Quillon, out of 600 | 22% |
| Share of voice | 132 of 680 mentions (rivals have 240, 198 and 110) | 19.4% |
| Platform coverage | Citation frequency: ChatGPT 9%, Perplexity 21%, AI Mode 12%. Threshold 15% | 1 of 3 engines |
| Primary recommendation | 33 of the 132 responses that name Quillon | 25% |
| Attribution | 31 of 84 citations go to one comparison page | 37% |

Three actions follow. ChatGPT is the weakest engine, so Quillon compares the pages it cites there with its own. Being named first only a quarter of the time points to a positioning problem, so Quillon reads those answers to see who is preferred and why. One page carries 37% of citations, so keeping it current comes first.

None of this is meaningful until the second round shows how much these numbers move on their own.

## What Changes by Business Type

| Business | Prompts to favor | What to watch |
| --- | --- | --- |
| **B2B SaaS** | Comparisons, integrations and “best for” questions | Siege’s 1.05x is a weak case for direct conversion. Report visibility, and ask new leads how they found you |
| **Ecommerce** | Product, price and “best under” questions | At 1.01x, revenue per session matters more than the conversion ratio |
| **Finance and other regulated fields** | Eligibility, cost and risk questions | Read framing by hand, since a wrong description costs more than a missing one |
| **Small sites** | Start with 20 prompts on two engines | Search Console may not show AI data yet, so manual runs are the baseline |

These are our recommendations, apart from the Siege Media figures.

## A Measurement Cadence

This schedule is a practical starting point, not a research finding:

- **Weekly:** citation frequency on your 20 to 30 most important prompts
- **Monthly:** the full prompt set across all engines, with share of voice and coverage
- **Quarterly:** attribution, sentiment and framing, and a review of the prompt set

Keep prompts fixed between reviews so the numbers stay comparable. Log any content or PR changes with their dates.

## Common Mistakes

**Editing the prompt set mid-series.** Any new or reworded prompt breaks comparison with earlier rounds. Add prompts only at a quarterly review, and report old and new sets separately.

**Blending engines into one score.** A rise on Perplexity can hide a loss on ChatGPT.

**Mixing collection methods.** SparkToro’s authors call it an open question whether API calls match what real users see. Record whether you used the app or the API, and don’t switch between rounds.

**Copying someone else’s benchmark.** Published rates come from different prompts, categories and engines. A 20% citation rate could be excellent in one category and poor in another.

**Reporting a percentage with no count.** Write “14% (84 of 600)”, so a reader can see how much data sits behind it.

## Start This Week

1. Write your 50 prompts and name three competitors.
2. Open Search Console’s generative AI report and export AI impressions by page.
3. Run your first baseline round and repeat it in seven days.

LLMagnet measures AI visibility across seven engines and has an agent that prepares fixes for approval, so the prompt runs don’t have to stay manual. You can start with the [free AI visibility audit](https://ai-visibility.llmagnet.com).

## FAQ

### Why don’t SEO tools fully measure AI visibility?

Rank trackers and analytics mostly see search positions and visits. Search Console now shows AI Overview and AI Mode impressions by page, but it covers only Google and doesn’t show whether your brand was named in the answer.

### What are the five GEO metrics?

Citation frequency, AI share of voice, platform coverage score, sentiment and framing score, and content-to-citation attribution. Together they show how often you appear, against whom, where, how you are described and which pages earn it.

### Is AI traffic really worth measuring?

It depends on your site. The agency Siege Media’s analysis of 78 sites found a median AI-to-organic conversion ratio of 1.26x, and AI was roughly 0.2% to 3% of sessions on most sites. Test the revenue case on your own data.

### What does a minimum GEO baseline look like?

Use 50 prompts, at least ChatGPT, Perplexity and Google AI Mode, three tracked competitors, and two rounds seven days apart, with several runs of each prompt per round.

### How much variance should I expect in AI visibility?

Measure your own, because it depends on the prompts and engines. AirOps’ vendor data found only 30% of brands stay visible from one answer to the next, so small changes in small samples may be noise.

### How often should I measure GEO?

A practical schedule is weekly on 20 to 30 key prompts, monthly for the full prompt set, and quarterly for attribution and sentiment. Keep prompts fixed between reviews so results stay comparable.

### How many times should I run each prompt?

SparkToro’s authors recommend 60 to 100 runs per prompt where you can. Few teams can do that on every prompt, so run a smaller number on each one and report the whole prompt set, where the total response count is large enough to track a trend.