AI Search Visibility Metrics & KPIs: How to Measure GEO, AEO

Five measures cover AI search visibility, each taken from a fixed prompt list: mention rate (how often an answer names you), citation rate (how often it links your pages), share of answer (your mentions against competitors'), characterization (what the answer says about you), and referral sessions from AI assistants. There is no rank, because answers are generated, not listed.

This guide is about marketing measurement: what to count, how to collect it with a spreadsheet instead of paid software, and what each number cannot tell you. If your concern is what AI says about your reputation, including errors and alerts, our guide to monitoring your brand in AI search covers that side. If you would rather have an analyst run the panel, AI reputation monitoring is a scoped service, and no one can promise that a given engine will name or cite you.

If you want a baseline before deciding anything, a free AI visibility audit, delivered by an analyst, runs a prompt panel like the one below and reports what it finds. Everything in this guide you can also do yourself.

The five metrics

Four of the five come from a prompt panel, a fixed list of questions you run on a schedule and record by hand or with a tool. The fifth comes from your analytics. None of them replaces the others, and each has a blind spot, which is why the table ends with what the metric cannot tell you.

• Mention rate

  Definition

  Share of panel runs in which the answer names your brand

  How to collect

  Panel: mark yes or no for each run, then divide runs naming you by total runs

  What moves it

  Being present in the sources the engine reads, a consistent brand name and descriptor, third-party coverage

  What it cannot tell you

  Whether the mention was accurate or favorable, or whether anyone clicked

• Citation rate

  Definition

  Share of panel runs in which the answer links or cites a page you own

  How to collect

  Panel: record every source URL shown, then count your domain

  What moves it

  Pages that can be crawled and that answer the question directly in their opening passage

  What it cannot tell you

  Whether the answer text is right; not every assistant shows sources on every answer

• Share of answer

  Definition

  Your mentions as a fraction of all brand mentions (yours plus competitors) across the panel

  How to collect

  Panel: list every brand named in each run, then divide

  What moves it

  Competitor activity as much as your own; which prompts you chose

  What it cannot tell you

  Absolute demand; your share can rise because a competitor fell

• Characterization

  Definition

  What the answer says about you, coded with a fixed scheme

  How to collect

  Panel: read each answer and code it (see below)

  What moves it

  The underlying sources: profiles, reviews, press, your own pages

  What it cannot tell you

  Why the engine said it; coding involves judgment, so use two reviewers

• AI referral sessions

  Definition

  Visits in your analytics whose source is an AI assistant

  How to collect

  Analytics: traffic acquisition report by session source, plus a custom channel group

  What moves it

  Whether answers link to you and whether readers click

  What it cannot tell you

  Mentions that produce no click, or later branded searches the answer influenced; the count is a floor

A few notes on coding. For characterization, use a short fixed scheme so two people reading the same answer land in the same bucket: accurate and favorable, accurate and neutral, accurate and unfavorable, inaccurate, outdated, and not applicable. Write the scheme down before the first run and do not change it mid-year, or your trend line breaks.

Mention rate and citation rate are different. An answer can name you with no link, or cite your page while naming a competitor. Record them separately. Our guide to how to get cited by AI covers what tends to move citation rate, and the guide to answer engine optimization covers the page-level side.

Why there is no AI rank

Searches like how to track ChatGPT rankings over time assume there is a list and you sit at a position in it. There is not. A classic results page is a ranked list of links, and a position means something because the list is the product. An AI answer is generated text. It may name three brands or none, it may name different brands on the next run, and the order inside a sentence is a writing choice, not a ranking.

Perplexity's own help center describes the product this way: it searches the internet in real time and each answer includes numbered citations that link to the original sources. It also says it remembers the context of earlier queries in a conversation, which means a follow-up can change which sources appear. Google's documentation on AI features says AI Overviews and AI Mode may use different models, so the links they show can differ. Neither is a ranking system you can hold a position in.

What you can track instead is frequency and character: how often you are named, how often you are cited, what is said, and how that changes month over month. Any vendor that shows you a single AI rank for a prompt is simplifying something that varies from run to run. Ask how many runs sit behind the number.

Is there an AI visibility score?

Some tools publish a single AI visibility score. Treat it as a vendor construct. There is no standard definition or shared formula, so scores from two tools cannot be compared, and a movement does not tell you which underlying metric moved. We do not use one here. Report the five measures separately, and if you use a vendor score as a dashboard headline, ask for the formula and the run counts behind it.

How to build a prompt panel

A prompt panel is the method behind the first four metrics. This is a protocol you can copy. It is deliberately small enough to run by hand.

1. Pick 20 to 50 prompts, in four classes. Use the words a real buyer, journalist or candidate would use, not your marketing copy.

• Identity

  What it tests

  Whether the engine knows who you are and gets the basics right

  Example pattern

  Who is [your company] and what does it do?

• Legitimacy

  What it tests

  Whether the engine describes you as trustworthy, and what it cites for that

  Example pattern

  Is [your company] legit? What do people say about [your company]?

• Comparison

  What it tests

  Whether you appear when a buyer weighs options

  Example pattern

  [Your company] vs [competitor]; alternatives to [competitor]

• Recommendation

  What it tests

  Whether you are named when the buyer has not named anyone

  Example pattern

  Which [category] providers would you recommend for [situation]?

Recommendation and comparison prompts are where mention rate and share of answer are earned. Identity and legitimacy prompts are where characterization matters most. A panel made only of branded prompts will flatter you, because the engine was handed your name. Split the list across all four classes and keep at least a quarter of it unbranded.

2. Choose prompts from real questions. Good sources are your sales and support inboxes, the questions prospects ask on calls, the queries in Search Console, and the follow-up questions the engines suggest themselves. Freeze the list. Add new prompts in a separate batch and mark them, so month-to-month comparison stays clean.

3. Run logged out, in a clean session. Where the engine allows it, sign out. Where it does not, use a fresh browser profile with history and memory turned off. Your own account history can bias the answer toward you, which makes your numbers look better than a stranger's experience.

4. Run with search on and off. Where an assistant lets you switch web search on or off, run each prompt both ways. With search on, the answer is drawn from sources the engine retrieves at that moment, and you get cited URLs. With search off, you see what the model says without looking anything up. They answer different questions: the first tells you which pages and sources matter now, and the second tells you what the model has absorbed.

5. Run each prompt three times. Answers vary between runs, so a single run is an anecdote. Three runs per prompt is a practical minimum for a hand-run panel, not a statistical guarantee. Report mention rate as a share of runs, and treat a small movement on a small panel as noise. Twenty prompts, three runs and two search settings is 120 answers per engine per month. Fifty prompts is 300. Size the panel to what you will actually finish.

6. Keep a fixed monthly cadence. Run on the same day each month, such as the first business day, and record the date. Run it more often only around events you want to measure, such as a content launch or a press cycle, and label those runs separately. Fix the engine list too, for example ChatGPT, Gemini, Perplexity, Microsoft Copilot, Claude and Google AI Overviews or AI Mode, and note the date any engine is added or dropped.

7. Record the same fields every time. For each run: date, engine, search setting, prompt, run number, mentioned (yes or no), cited (yes or no), every source URL shown, the full answer text, and a characterization code. Save the full text. When an answer is wrong, you will need the exact wording and the sources it leaned on to fix the cause.

A blank reporting template

Use one row per prompt per engine per search setting, with the three runs summarized. Copy this into a spreadsheet.

Roll it up by month into four numbers per engine and per prompt class: mention rate, citation rate, share of answer and the characterization mix. Chart the trend over several months rather than reading a single month. A one-month change in a 20-prompt panel is rarely a signal.

Measuring AI referrals in analytics

The fifth metric comes from your analytics, not from the panel. The idea is simple: identify visits whose source is an AI assistant, group them, and watch the trend.

In Google Analytics 4, the traffic acquisition report breaks sessions out by source and medium. Google's help documentation describes the Referral channel as visits from non-ad links on other sites and apps, and describes the default channel group rules that classify traffic. It also says you can create custom channel groups for your own categories. A practical setup is a custom channel group that collects the assistants' domains you see in your own source list into one AI assistants channel, so the count is visible at a glance. Build the list from what actually appears in your data rather than from a list someone else published, and revisit it quarterly because assistants and their domains change.

Three cautions. First, the count is a floor. Some visits can arrive without referrer information and be filed as direct, and many people read an answer, remember your name and search for you later, which shows up as branded search. Second, Google reports clicks from AI Overviews and AI Mode as part of overall Search traffic: its documentation says sites appearing in these features are included in the overall search traffic in Search Console, in the Web search type. On that page we found no separate AI-only view, so check your own property before assuming you can isolate those clicks. Third, a referral shows that someone clicked, not that your brand was described well.

Server logs add a supporting signal. OpenAI's crawler documentation says OAI-SearchBot is used to surface websites in ChatGPT's search features, and that ChatGPT-User is triggered by user requests in ChatGPT and Custom GPTs. Seeing those user agents fetch a page tells you the page was reachable, and it is a useful health check on crawler access. It does not tell you the page was cited in an answer, so do not report it as a citation.

A last practical step is a how did you hear about us field on lead forms, with an option for AI assistants. It will catch the people analytics cannot attribute.

Conversion rate and CPA for AI traffic

People often ask what the conversion rate or cost per acquisition for AI traffic should be compared with Google. We do not quote a figure, because we know of no published benchmark we can verify, and your own funnel differs from anyone else's. Compute yours.

Conversion rate. In analytics, take sessions in your AI assistants channel and the key events you already track, such as a form submission, a call or a booked meeting. Conversion rate is conversions from that channel divided by sessions from that channel over the same period. Put organic search and direct beside it for the same period, using the same definition of a conversion. Report the raw counts next to the rate. If the channel has a few dozen sessions, a rate is unstable, and one extra conversion will swing it.

Cost per acquisition. CPA is cost divided by conversions. The hard part is the numerator: AI visibility has no media buy. Add up the fully loaded cost of the program for the period, meaning staff time, any tools and any agency fees, then divide by conversions from the AI channel. Use a long enough window that a lead recorded this month can be matched to work done earlier. State plainly that the number understates the program's effect, since it excludes assisted visits that arrive as branded search or direct.

What not to do. Do not borrow another company's rate or a vendor's industry average as a target. Do not compare a three-month-old channel with an organic channel that has years of history. Do not report CPA without the conversion count underneath it.

KPIs for GEO vs SEO

AI search measurement sits beside SEO measurement and does not replace it. Google says the same fundamentals apply to its AI features, so the pages you measure in SEO are the pages that feed AI answers. Here is how the questions line up.

• Are we present?

  SEO KPI

  Indexed pages, impressions

  GEO and AEO KPI

  Mention rate

  Shared caution

  Presence is not the same as being chosen

• Are we chosen?

  SEO KPI

  Average position, click-through rate

  GEO and AEO KPI

  Citation rate, share of answer

  Shared caution

  Position is a list concept; share is a frequency

• Is what is said correct?

  SEO KPI

  Snippet text, brand search results

  GEO and AEO KPI

  Characterization mix

  Shared caution

  Needs a human coder and a fixed scheme

• Does it send traffic?

  SEO KPI

  Organic sessions

  GEO and AEO KPI

  AI referral sessions

  Shared caution

  Both undercount influence that arrives later as branded search

• Does it convert?

  SEO KPI

  Organic conversions and CPA

  GEO and AEO KPI

  AI-referral conversions and CPA

  Shared caution

  Small volumes make rates unstable

• Is it stable?

  SEO KPI

  Ranking trend over weeks

  GEO and AEO KPI

  Panel trend over months

  Shared caution

  Expect run-to-run variation in any single month

For a deeper look at how the two disciplines relate, see our guide to generative engine optimization.

Measuring ROI without overclaiming

ROI for this work is easy to calculate badly. A defensible version has three parts: the value of conversions you can attribute to the AI channel, a note on the influence you cannot attribute, and the cost of the program. Report the first and third as numbers and the second as a caveat, and resist adding an estimated multiplier for unattributed influence. If the attributed number alone does not justify the work, say so. If it does, you have a conservative case.

Review the panel monthly and make decisions quarterly. One bad month should send you to the recorded answers and their sources, not to a new strategy.

When a tool is worth it

A spreadsheet and an hour or two a month is enough for a 20-prompt panel. Consider a tool when the panel outgrows what you will finish by hand or several people must code answers consistently, and check that it stores full answer text and every source URL, shows the run count behind each number, and exports your data.

If you want the measurement done for you, our AI search optimization work includes it, and the AI visibility audit page describes the first step. We have worked on online reputation since 2013, and we cannot control what any engine says or cites.