AEO metrics and KPIs: definitions

The main KPIs for AEO and GEO are the mention rate, which is the share of AI answers to buyer questions that name your company, the citation rate, share of voice, and the accuracy of what the answers say about you. Each is counted separately for each AI assistant, over many repeated runs. Around them sit the counts Google and Microsoft report to site owners, AI referral visits in analytics, and what buyers say when asked how they found you.

This page defines each metric, gives its formula and data source, and states what it cannot show. It is a companion to how to measure AI visibility, which covers how many prompts and runs to collect and how to calculate error ranges, so those steps are not repeated here. Where a platform or a vendor defines a metric differently from AEO HQ, the difference is noted. Sources were checked on September 27, 2026.

Scope and definitions

This page covers assistants that answer questions in sentences and sometimes cite web pages: ChatGPT, Google's AI Overviews and AI Mode, Gemini, Claude, Perplexity, and Microsoft Copilot. It does not cover ordinary search rankings, except where Google or Microsoft report AI features in their own tools.

• A metric is a defined count or rate. A KPI (key performance indicator) is a metric chosen to judge progress toward a stated goal. The same metric can be a KPI for one team and background for another.
• A prompt is a question typed into an assistant. A run is one answer to one prompt, collected once. A prompt panel is a fixed set of prompts run on a schedule.
• An unbranded prompt asks about a need or a category without naming a company. A branded prompt names the company, such as "What does [company] charge?"
• A mention is the company's name in an answer. A citation is a link to one of the company's pages shown with the answer.
• A window is the period a figure covers, such as four weeks.

The metrics at a glance

The sections below take each metric in turn. Figures marked "AEO HQ calculation" are ours.

Metrics from a prompt panel

The first eight metrics come from a prompt panel: fixed prompts run on each assistant, with every answer saved. The measurement guide linked above explains how to build the panel and how many runs to collect, and AEO HQ's recommended measurement design gives default panel sizes.

A worked example

The table below is hypothetical data, made up to show the arithmetic. It describes no real company or result. It has three prompts, two runs of each, one tracked company, and two competitors, X and Y.

From this log (AEO HQ calculation):

• Mention rate: 3 of 6 runs, 50.0%.
• Citation rate: 1 of 6 runs, 16.7%. Counted only among the four answers that cited any source, it would be 1 of 4, 25.0%.
• Prompt coverage: 2 of 3 prompts, 66.7%.
• Share of voice: the company was named 3 times and the competitors 6 times (X four times, Y twice), so 3 of 9 mentions, 33.3%.

Six runs are far too few to act on: the 95% Wilson interval, an error range suited to small samples, runs from 18.8% to 81.2% for 3 of 6 (AEO HQ calculation; the method is in AEO HQ's statistics section).

Mention rate

Definition. The share of runs whose answer names the company. AEO HQ counts a mention when the answer names the company as an option or describes it, using matching rules fixed before the results are read, such as the exact company name, its common variants, and its domain.

How to compute it. Mention rate = runs that name the company ÷ all runs, for one assistant and one window. Keep every run in the denominator, including answers where the assistant did not search.

Why it is the usual primary KPI. Whether a brand is named holds steadier than where it appears. In a study in which 600 volunteers ran 12 prompts a combined 2,961 times, it took about 1 in 1,000 runs to see two brand lists in the same order (opens in a new tab), and the author concluded that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric" (opens in a new tab) (vendor study; a co-investigator works for a tracking vendor). In a panel of 102,025 answers, 77.5% of brand, prompt, and engine combinations were always or never mentioned (opens in a new tab) (preprint; the author co-founded, and holds equity in, the platform studied).

Limits.

• It describes the prompts, not all buyers. People word the same need very differently: 142 prompts written by people for one intent had a mean semantic similarity of 0.081 (opens in a new tab). A rate holds for the panel it was measured on.
• It needs an error range and a per-assistant breakdown. Report each assistant separately, with the number of runs and an interval.
• Vendors count mentions differently. HubSpot writes that "Mentions tell you how often your brand appears in AI-generated answers without a link" (opens in a new tab) and treats linked appearances as citations. AEO HQ counts a mention whether or not a link is shown. Ask how a tool counts before comparing its numbers with anyone else's.

Citation rate

Definition. The share of runs whose answer links to at least one of the company's pages.

How to compute it. Citation rate = runs that cite the company ÷ all runs. Group URLs by domain before counting, and decide in advance how redirects, parameters, and subdomains are handled (AEO HQ recommendation).

Variants. Other counts share the name. Some tools report the share of all citations that point to the company, which divides by citations rather than by runs. HubSpot's documentation records a citation when an answer includes a link to a webpage or "a domain name embedded in its answer" (opens in a new tab), so a domain written as text also counts. Bing's first-party counts are described below.

Limits.

• Many answers cite nothing. In one panel, ChatGPT activated web search only for specific queries, leaving 57.8% of its runs with zero citations (opens in a new tab) (preprint; German-language commercial prompts, January–March 2026; one author is affiliated with the company whose data export was used). A rate calculated only among answers that cite something leaves those runs out and overstates visibility. A 2026 review of 45 studies puts it plainly: "Outputs without search, without citations, or with errors are outcomes, not data to be discarded" (opens in a new tab) (preprint).
• Being cited is not being recommended. In a published test of 48 AI answers to questions about which agency to hire, the tester found that one agency, Searchbloom, was the source AI reached for most on the question in its wider tracking, yet was named as an agency once in 48 answers (opens in a new tab) (practitioner test by an agency). Report mention and citation rates side by side.

Prompt coverage

Definition. The share of tracked prompts for which the company was named at least once in a window. HubSpot uses it as one part of its brand visibility score: "Prompt coverage: the percentage of prompts where your brand appears at least once." (opens in a new tab)

How to compute it. Prompt coverage = prompts with at least one mention in the window ÷ prompts tracked.

Limits. Coverage rises with the number of runs even when nothing else changes. If a company has a fixed 10% chance of being named in each answer to a prompt, the chance it is named at least once is 10.0% after one run, 27.1% after three, 52.2% after seven, and 95.8% after 30 daily runs (AEO HQ calculation: 1 minus 0.9 raised to the number of runs, treating runs as independent). Real runs of one prompt tend to repeat each other, as the 77.5% figure above shows, so real coverage climbs more slowly. Compare coverage only between periods, tools, or competitors measured with the same number of runs.

Share of voice

Definition. Share of voice is the company's mentions as a share of all mentions of a named set of companies, the company included. HubSpot's documentation defines it as "the proportion of total brand mentions that belong to your brand across tracked prompts" (opens in a new tab).

How to compute it. Share of voice = company mentions ÷ mentions of every company in the set, counted over the same runs.

Limits. It changes whenever the competitor set changes, so name the set beside the figure and keep it fixed. Two tools can report different shares for the same company if one counts a company once per answer and the other counts every time it is named. It measures how often, not how favorably. Bing's Citation Share, described below, is a different metric.

Position

Definition. Where the company first appears in an answer that lists several options, such as first or third. Several tracking tools report it as a rank or an average position; AEO and AI visibility tools compared notes which.

Limits. Order is the least stable part of an answer: two brand lists in the same order turned up about once in 1,000 runs in the study above, and its author called tracking "ranking position" in AI tools "foolhardy" (opens in a new tab). AEO HQ does not recommend position as a KPI. If you report it, report the share of runs in which the company was named first, with the number of runs.

Sentiment

Definition. How favorably an answer describes the company, usually graded positive, neutral, or negative by a language model or a person. HubSpot's sentiment score shows "how positively or negatively your brand is described in answer engine responses" (opens in a new tab).

Limits. Sentiment moves far more than mentions do: whether a brand was framed positively or negatively flipped about 6.7 times more often than whether it was mentioned (opens in a new tab) (preprint). A change in one window is weak evidence of a real change. AEO HQ's recommended design reports sentiment as description only, not as a KPI (what the design reports).

Accuracy

Definition. The share of answers to branded prompts that state the company's facts correctly, graded by a person against a written fact sheet: what the company sells, its prices, who it serves, its locations, and its founders.

How to compute it. Accuracy = answers graded correct ÷ answers graded. Record every error and the page the answer cited for it, because that page is where the fix starts.

Limits. It needs a person and a current fact sheet, and it covers only the branded prompts you ask. The evidence that assistants misstate facts is covered under "Counting mentions without checking what the answer says" in AEO antipatterns.

Composite visibility scores

Definition. A single number that a vendor builds from several measures. HubSpot's free AI Search Grader, for example, publishes its weights: sentiment up to 40 points, presence quality 20, brand recognition 20, share of voice 10, and market competition 10 (opens in a new tab), scored from what ChatGPT, Perplexity, and Gemini say about a brand "based on their training data" (opens in a new tab).

Limits. A score hides which part moved. The 2026 review of 45 studies says a single score is defensible only when its weights correspond to an explicit objective, and that combining a mention, an accurate citation, and a conversion without a model of their value "merely obscures normative choices" (opens in a new tab). AEO HQ's own automated audit reports a readiness score that a language model writes without a fixed formula, and the methodology page says so. Our recommendation: report the parts separately, and if you use a vendor's score, get its formula and weights in writing.

Metrics from Google's and Microsoft's own reports

Google and Microsoft give site owners first-party counts of their pages in AI answers. OpenAI's publisher FAQ points publishers to analytics platforms such as Google Analytics for tracking referral traffic from ChatGPT (opens in a new tab) and describes no citation report.

AI feature impressions (Google Search Console)

Definition. Search Console's generative AI performance report counts impressions in AI Overviews and AI Mode. Google defines them this way: "Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search." (opens in a new tab)

How it is counted. If two results from the same site appeared in one generative AI feature, they count as a single impression in the chart total, and the table can group data by page, country, date, or device (opens in a new tab). The report has been available to all websites worldwide since August 31, 2026 (opens in a new tab).

Limits. It counts impressions, not clicks: clicks from AI Overviews and AI Mode are counted in the Performance report under the Web search type and are not shown separately (opens in a new tab). A site may not see the report until it has received enough impressions in these features (opens in a new tab). It covers Google only, and it cannot see an answer that names you without linking to you.

Citations and cited pages (Bing Webmaster Tools)

Definition. Bing's AI Performance report counts citations of your pages across Microsoft Copilot, AI-generated summaries in Bing, and "select partner integrations" (opens in a new tab). Its main figures, in Microsoft's words:

• Total Citations: "the total number of citations that are displayed as sources in AI-generated answers during the selected time frame" (opens in a new tab).
• Average Cited Pages: "the average number of unique pages from your site that are displayed as sources in AI-generated answers per day" (opens in a new tab).
• Grounding queries: "the key phrases the AI used when retrieving content that was referenced in AI-generated answers" (opens in a new tab), shown as a sample.

Limits. The measures Microsoft lists count citations; none counts clicks (opens in a new tab), and Microsoft says its page-level counts reflect how often pages are cited, "not page importance, ranking, or placement" (opens in a new tab). It covers only Microsoft's own products and the partners it does not name.

Citation Share (Bing Webmaster Tools)

Definition. Added in June 2026, Citation Share is "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query" (opens in a new tab).

Limits. Microsoft calls it an observational metric, "not a ranking system or a competitive scoreboard," that "does not expose competitor domains, represent traffic share, or assign quality scores to content" (opens in a new tab). It differs from share of voice in three ways: it counts citations rather than mentions, it is calculated for one grounding query at a time, and it compares you with all cited sites rather than a named set of competitors.

Traffic and outcome metrics

AI referral sessions

Definition. Visits that begin with a click on a link in an assistant's answer. Google Analytics 4's AI Assistant channel counts visits from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok, and excludes Google's AI Overviews and AI Mode (opens in a new tab), which it counts as organic search. Microsoft Clarity's AIPlatform channel captures sessions that begin with a click on "a free, organic link within an AI platform's chat experience," and a separate PaidAIPlatform channel captures clicks on paid ads within those platforms (opens in a new tab).

How to compute it. Count sessions in a channel defined by the assistants' web addresses; the setup is in how to track AI referral traffic in GA4. Share of traffic = AI referral sessions ÷ all sessions.

Limits.

• It is small. In a February 2025 study of 3,000 sites, 0.17% of the average site's visitors came from AI chatbots (opens in a new tab) (vendor study).
• It undercounts. Traffic referred by Claude's native app does not include a Referer header (opens in a new tab) (network measurement), so such visits are recorded as direct.
• Most answers produce no click. Google users who saw an AI summary clicked a link in the summary itself in 1% of visits (opens in a new tab) (browsing data from 900 U.S. adults, March 2025).

AI key events and conversion rate

Definition. Key events are actions you mark as important in Google Analytics 4, such as a form submission or a purchase. GA4's session key event rate is "the percentage of sessions in which a user triggered a key event" (opens in a new tab).

How to compute it. Count key events in AI referral sessions. AI session key event rate = AI referral sessions with a key event ÷ AI referral sessions. Compare it with other channels only when each has enough sessions to be meaningful.

Limits. Whether AI visitors convert better is contested. Across 973 e-commerce sites, ChatGPT referrals converted above paid social but below all other traditional channels (opens in a new tab) (peer-reviewed; the first author is employed by the company that supplied the data). One software company, by contrast, reported that AI search sent 0.5% of its visitors and 12.1% of its sign-ups over 30 days (opens in a new tab) (single site; the company sells SEO software). Counts based on clicks miss buyers who read an answer and arrive later another way, which the next metric addresses.

Self-reported AI source

Definition. The share of new leads or customers who name an AI assistant when asked how they first heard of the company.

How to compute it. Answers naming an AI tool ÷ all answers to the question. Report the response rate beside it, because the question is often optional.

Example. In one agency's records, 189 of 213 leads who answered the question named an AI tool, but first-touch attribution credited AI with only 28 of those 189; 95 were filed under Organic Search and 59 under Direct (opens in a new tab) (single firm; the agency sells AEO services). The agency treats self-report as a lower bound, because only leads who completed an optional field are counted (opens in a new tab).

Limits. People misremember, skip the question, or name the last thing they saw. It does not say which assistant, prompt, or answer they saw unless you ask.

AI-influenced pipeline

Definition. Deals and revenue in the CRM from contacts whose self-reported or recorded source is an AI assistant.

How to compute it. Sum pipeline or revenue for those contacts in a window. Keep contacts identified by self-report apart from those identified by click-based source, because the two methods disagree.

Limits. It shows association, not effect. The assistants' own growth raises AI numbers with no work on your side: in the only controlled field study we found, total ChatGPT referrals to one site grew 5.7 times while its untreated pages grew 3.5 times (opens in a new tab) (preprint; the authors work for the company that owns the site). To judge whether work paid off, compare changed pages or periods with unchanged ones, as that study did. AEO examples collects documented cases, including this one.

Access metrics: crawler and fetcher requests

Definition. Requests to your site from AI companies' agents, counted by agent and URL in server logs. Two kinds matter (crawler vs. fetcher): search crawlers, which build the index an assistant searches, and user-initiated fetchers, which visit a page while an assistant answers a person. OpenAI says ChatGPT-User visits pages when a user's question calls for it and is not used to crawl the web automatically (opens in a new tab); Anthropic says Claude-User accesses websites when people ask Claude questions (opens in a new tab); and Perplexity says Perplexity-User visits a page when a user asks Perplexity a question (opens in a new tab). Each company's agents are listed in the glossary under OpenAI crawlers, Anthropic crawlers, and Perplexity crawlers.

How to compute it. Requests per agent, per URL, per window, and the share answered with HTTP 200. Check that a request claiming to come from a crawler really does: OpenAI publishes IP address lists for its crawlers and fetchers (opens in a new tab), as do Anthropic (opens in a new tab) and Perplexity (opens in a new tab).

Limits. A fetch shows that an answer retrieved your page, not that it cited or recommended you. None of the major AI crawlers rendered JavaScript (opens in a new tab) in Vercel's December 2024 data, so these requests never appear in analytics tools that run in the browser. Some requests cannot be counted by name: Brave's crawler does not advertise a differentiated user agent (opens in a new tab).

Choosing KPIs

These are AEO HQ's recommendations. Pick one KPI for each question the business needs answered, and keep the rest as supporting metrics.

Five rules keep the numbers honest:

1. Report each assistant separately. A pooled figure describes none of them.
2. Attach the method to every figure: the prompts, the number of runs, the window, and the interval.
3. Mark events on the timeline, such as model releases, panel changes, and site changes.
4. Do not combine the KPIs into one score.
5. Judge change against a comparison, such as unchanged pages or an earlier period measured the same way.

AEO metrics compared with SEO metrics

What the evidence shows and does not show

Antipatterns

An antipattern is a practice that looks helpful but fails or backfires. AEO antipatterns covers four measurement mistakes in depth. Seven more concern the metrics themselves:

1. Comparing prompt coverage across different numbers of runs. Coverage rises with runs alone, as the calculation above shows. Instead, compare at the same number of runs.
2. Dividing citations by cited answers only. It drops the runs where the assistant did not search (opens in a new tab). Instead, divide by all runs.
3. Treating a citation as a recommendation. A site can be read often and named rarely (opens in a new tab). Instead, report mention and citation rates separately.
4. Tracking position as a KPI. Order almost never repeats (opens in a new tab). Instead, use the mention rate.
5. Reporting a score without its weights. A score needs weights tied to a stated objective (opens in a new tab). Instead, report the parts, or publish the formula.
6. Crediting all growth to the work. Untreated pages on one site grew 3.5 times (opens in a new tab) over the same period. Instead, compare with unchanged pages or periods.
7. Counting fetches as citations. A fetch is retrieval, not a recommendation. Instead, report log data as an access metric.

Checklist

FAQ

What are the most important AEO metrics to track?

Start with the mention rate on unbranded buyer prompts for each assistant, then the citation rate, share of voice, and accuracy on branded prompts. Add Search Console's AI feature impressions, Bing's citations, AI referral sessions, and a "How did you hear about us?" question. This set is AEO HQ's recommendation; no standard list exists.

How do I track AEO?

Use four instruments: a prompt panel run on each assistant, the reports Google and Microsoft give site owners, analytics with a channel for AI assistants, and a question on your forms about how buyers found you. The measurement guide linked at the top of this page sets out the steps.

What is a good mention rate?

No published benchmark exists, because the rate depends on the category, the prompts, and the assistant. As context only, one panel found that first answers named household brands 73% of the time, mid-market brands 44%, and small brands 11% (opens in a new tab) (preprint; the author co-founded the platform studied). Track your own rate against your own earlier windows and your named competitors.

How are AEO KPIs different from SEO KPIs?

SEO KPIs count positions, impressions, and clicks in a results page. AEO KPIs count how often answers name, cite, and correctly describe a company across repeated runs, because answers change from run to run. The comparison table above maps each SEO metric to its closest AEO metric.

Which AEO metrics show ROI?

No metric shows return on investment by itself. Combine self-reported AI source and AI-influenced pipeline with a comparison against unchanged pages or earlier periods, because the assistants' own growth raises AI numbers for everyone. The field study above found untreated pages growing 3.5 times (opens in a new tab) over the same period.

How often should AEO metrics be reported?

AEO HQ recommends rolling four-week windows, reported for each assistant. HubSpot gives similar advice for its own tool: "Because responses change over time, review performance across multiple days or weeks before evaluating trends." (opens in a new tab)

What is the difference between share of voice and Citation Share?

Share of voice counts your mentions against a named set of competitors across your prompt panel. Bing's Citation Share counts your citations against all citations for one grounding query in Microsoft's AI answers, and does not expose competitor domains (opens in a new tab).

Can Google Analytics measure AEO on its own?

No. Analytics records visits, not answers. It misses answers that produce no click and visits that arrive without a referrer, and it cannot separate clicks from AI Overviews and AI Mode, which GA4 counts as organic search (opens in a new tab).

Next steps

AEO HQ's Instant AEO Audit ($499) is an automated check of a site's crawler access, key pages, and schema, with a five-question sample from one AI model. It does not calculate the rates on this page; the methodology page lists what it measures and what it does not.

Change log

• September 28, 2026: First published.