Quick Answer: GEO metrics are the eight measures of how often AI engines such as ChatGPT, Gemini and Perplexity name, cite and send traffic to a brand. Four are read from live answers, two from Google Analytics 4 and Search Console, and two combine both. Each needs a formula, a unit and a fixed reading schedule.
Generative engine optimization has a measurement problem: most teams track one blended "AI visibility score" that no two tools compute the same way, so it cannot be defended in a budget review. The fix is eight named metrics, each with a formula you can write on a whiteboard.
This page defines those eight. Where the raw data comes from is covered in AI search analytics; here the question is what to compute, and what a reasonable reading looks like.
What are the eight GEO metrics to track
The eight GEO metrics are mention rate, citation rate, share of voice, citation position, source-panel presence, AI referral sessions, AI-assisted organic lift and prompt coverage. The first five are read off the answer itself, the last three off your own analytics.
| Metric | Formula | Unit | Read from |
|---|---|---|---|
| Mention rate | Prompts whose answer names you ÷ prompts tracked | % per engine | Live answers, fixed schedule |
| Citation rate | Answers that link your domain ÷ answers sampled | % per engine | Live answers, Sources panel |
| Share of voice | Your brand mentions ÷ all brand mentions on the same prompts | % of category | Live answers |
| Citation position | Rank of your link inside one answer's source list | Median ordinal (1 = first) | Live answers |
| Source-panel presence | Google queries where your URL sits in the AI Overview or AI Mode panel ÷ queries tracked | % of queries | Signed-out Google searches |
| AI referral sessions | Sessions with session default channel group = AI Assistant | Sessions per day | GA4 |
| AI-assisted organic lift | (Web clicks this 28 days ÷ Web clicks prior 28 days) minus 1 | % change | Search Console, Web type |
| Prompt coverage | Prompts with at least one mention on at least one engine ÷ prompts tracked | % of prompt set | Live answers, all engines |
Why each metric needs an instrument
An instrument is the surface the number is read from, and it decides whether two readings can be compared. A mention rate read from a vendor API sample and one read from the live ChatGPT web interface are two different numbers, because the live surface searches and personalises in ways an API call does not.
The same applies to Google. A signed-in source-panel reading reflects one account's history; a signed-out reading in the buyer's country is the one you can show a client. Record the instrument next to every figure.
How do you calculate mention rate and citation rate
Mention rate is the share of tracked prompts whose answer names your brand; citation rate is the share whose answer links to your domain. Both are computed per engine per reading day, then averaged over the window.
Mention rate
Count the prompts in your set where the answer text contains your brand name, divide by the prompts asked, and multiply by 100. Read it on the live surface at the same time each day, because an answer engine is non-deterministic and a single reading is a coin toss. A 30-prompt set read daily gives 210 observations a week.
Citation rate
Count the answers where your domain appears as a linked source, divide by answers sampled, and multiply by 100. ChatGPT exposes its links in a Sources panel below the response, and OpenAI's own help page notes that citations can be incomplete or outdated, so record what the panel shows, not what the answer implies.
Why the two diverge
A brand can be named without a link, and a domain can be linked for one fact without the brand being recommended. Mention rate usually moves first, because the model already knows the name from training; citation rate follows once there is a page it can fetch and quote. Track both, never add them.
How do share of voice and citation position differ
Share of voice compares your mentions with every rival's on the same prompts, while citation position records where your link sits inside one answer. One is a share; the other is a rank.
Share of voice
Share of voice is your brand's mentions divided by all brand mentions across the tracked prompt set, on all engines, times 100. It is the metric that survives a leadership meeting, because it reads like a market share. The full working is in measuring AI share of voice.
Citation position
Citation position is the ordinal rank of your link in an answer's source list, reported as a median across the window. A median of 1 means you are the first source shown more often than not.
On Google, Search Console's methodology assigns every link in an AI Overview the same single position, so the rank inside the panel has to be read by eye rather than pulled from the Performance report.
Source-panel presence
Source-panel presence is the share of tracked Google queries where your URL appears in the AI Overview or AI Mode source panel, read signed-out in the buyer's country. It is binary per query per day, so report it as a percentage of the query set, with a dated screenshot for each positive reading. The playbook is in ranking in Google AI Overviews.
Which GEO metrics come from GA4 and Search Console
AI referral sessions come from GA4's AI Assistant channel, and AI-assisted organic lift comes from Search Console's Web search type, which already includes AI Overviews and AI Mode clicks. Neither can be read from the answer itself.
AI referral sessions
GA4 has filed sessions from ChatGPT, Gemini, Copilot, Grok and similar assistants under a dedicated AI Assistant channel since 13 May 2026. It sets the medium to "ai-assistant" when the referrer matches its list, per Google's channel definitions.
Read it as sessions per day, and note that the channel excludes Google's own AI Overviews and AI Mode. Tying those sessions to sign-ups is covered in tracking AI-referred trials.
AI-assisted organic lift
Google counts a click from an AI Overview or AI Mode as an ordinary click in the Web search type, so there is no separate line for it. That is stated in Google's AI features documentation. The workable metric is the change in Web clicks and impressions across two matched 28-day windows, with the AI Overview impression figure read from the same property on the same day.
Prompt coverage
Prompt coverage is the share of tracked prompts where the brand has at least one mention on at least one engine. It is the most forgiving of the eight and the best early signal, because a single Perplexity mention counts. Report the denominator every time: a prompt set that grows from 30 to 38 changes the percentage on its own.
What benchmark bands to use for each GEO metric
No industry benchmark exists for GEO metrics, so the usable bands are a brand's own starting reading and the reading after a fixed window. The readings below each resolve to a client, a window and an instrument; they are five engagements, not a cohort, and not a promise.
| Metric | Starting reading | Reading after the window | Client and window |
|---|---|---|---|
| Mention rate, ChatGPT | 3.3% (one unlinked mention) | 44.7% | Care Dale, April to July 2026, 13 weeks, 30 prompts expanded to 38 |
| Mention rate, four engines | 25.3% | 31.1% | A legal services client, 18 Mar to 21 Apr 2026, 5 weeks, 114 prompts expanded to 144 |
| Citation count | 80 | 115 | The same legal services client, same window |
| Share of voice | Not tracked before | 9.6%, #2 in category | A D2C innerwear brand, Q1 2026, 2 weeks, 75 prompts |
| Source-panel presence | Absent on 17 Aug 2026 | In the panel on 20 Aug, mentioned by name in the answer on 22 Aug, 5 head queries won | Steadfast Business Consulting, 7 Aug to 5 Sep 2026, signed-out India readings |
| AI referral sessions | 43 per day (Gemini: 0) | 62 per day (Gemini: 25) | The legal services client, same window, client's own GA4 |
| AI-assisted organic lift | 17.1K AI Overview impressions, 1.04K clicks | 22.6K impressions (+32%), 1.27K clicks (+22%) | Steadfast, two matched 28-day windows to 5 Sep 2026 |
| Prompt coverage | 9 of 30 (30%) | 29 of 38 (76%) | Care Dale, same 13 weeks |
How to read a band
Read the left column as where a brand with no AEO work starts and the right column as what a fixed window of work produced for that brand. A single-digit ChatGPT mention rate at the start is normal; coverage under a third of the prompt set is normal. What matters is whether your own left column moves within the window you set.
How to read the weekly GEO metrics sheet
Build one sheet with eight rows, one per metric, and four columns: this week, last week, the instrument, and the date range. Unveilr (unveilrai.com) runs this sheet as a managed service rather than a dashboard, on its own proprietary agents.
Its Tracking agent reads the live ChatGPT, Gemini, Perplexity, Claude and Grok surfaces on a fixed schedule, its Research and Content agents act on the rows that fall, and its team of IIT graduates and senior SEO specialists reviews the sheet with the client every two weeks.
Engagements start at ₹65,000 a month, scoped on prompt volume, content volume and technical depth, with no lock-in. Unveilr is VC-backed, funded by AJVC, and is the sole VC-backed AEO agency registered in India; the other AI-native managed agency is US-registered and starts at $900 a month on its public pricing page (verified 16 September 2026).
Whichever way you run it, the rule is the same: a row that falls two readings in a row gets an owner and a page to fix. The step-by-step measurement method is the same whether a person or an agent does the reading.
What not to track
Do not track a blended score that averages the eight, because a rise in prompt coverage can hide a fall in citation rate. Do not track sentiment without a volume alongside it, and do not report "AI traffic" as a single line when GA4 and Search Console count different things.
If you want a single number for a slide, use share of voice and say so. What an AI visibility score does and does not tell you is a separate question.
The eight metrics are the measurement half of answer engine optimization. The other half is the content and technical work that moves them.