How to Track Your GEO KPIs: Citations, Positions and Clicks

This page is for the B2B marketing or SEO manager who has to prove that their brand is cited by ChatGPT, Perplexity, Gemini and Google's AI Overviews. It sets out a single set of GEO KPIs, the source of each indicator and the decision it triggers.

Generative Engine Optimization (GEO) aims to make a brand visible and cited in the answers of generative engines; the fundamentals are set out in the definition of GEO and the overall approach in the GEO guide. Here, you decide what to track, how to record it and how to read a before / after without overpromising.

What GEO Measures, and What It Does Not

A generative engine publishes neither rankings nor query volumes: you measure your brand's presence in answers recorded under a stable protocol, then its visible effects on your site. GEO KPIs do not replace SEO metrics: they extend them. Indexing, positions, clicks and conversions remain the foundation, because a page that is not indexed has little chance of being picked up in an answer.

What Can Be Measured Reliably

The signals that are observable and comparable over time fall into five families:

• Presence in answers: brand named, domain cited or domain as source, on a fixed question set.
• Presence in AI Overviews: presence state recorded on the SERP of each tracked keyword.
• AI referral traffic: sessions from chatgpt.com, perplexity.ai or gemini.google.com, in GA4.
• Google impressions and clicks: Search Console, AI Overviews and AI Mode included.
• Conversions: key events from AI sessions and the associated pipeline.

What Cannot Be Measured: Your "Position" in ChatGPT

A generated answer varies with the model, the wording, the history and the session. Promising a "position" in ChatGPT means freezing a ranking that does not exist. Set out these limits from the scoping stage:

• No generative engine provides the volume of questions asked on your topic.
• A citation without a link leaves no trace in your analytics tools.
• Search Console separates neither AI Overviews, nor AI Mode, nor classic results.
• A search launched after an AI answer is attributed to Google or to direct.

The Single Set of GEO KPIs: Definition, Source, Decision

The reference set has six indicators: presence by keyword, citations by question, share of voice, referral traffic, Google impressions and conversions. Each has a stable definition, a single source and an attached decision; a KPI that triggers no action comes off the dashboard. The table below serves as the reference:

Keyword and Question: Two Units That Do Not Mix

A keyword carries a SERP and, where relevant, an AI Overview: on it you record one of the four presence states (brand named, domain cited, domain as source, absent). A question carries an answer requested from an assistant such as ChatGPT; it has no SERP. Never add the two together in the same average.

Presence in questions is read as movements, not as a stock. From one recording to the next, a citation is gained, lost or stable; a question added to the set has "no baseline", because it has no before.

AI Share of Voice: Calculate It on a Fixed Set

AI share of voice is the share of questions in the set where your brand is named. Calculate it engine by engine, then for two or three competitors on the same questions. Segmented by intent (TOFU, MOFU, BOFU), it shows where to focus editorial effort. A set that changes measures a change in scope, not genuine progress.

Building the Question Set That Serves as Your Reference

The question set is the measuring instrument: its quality determines the quality of every presence KPI. Apply these rules:

• By persona: write the questions each persona asks (marketing, IT, procurement), in their own words.
• By intent: cover discovery, comparison, objections and choosing a solution.
• Size: 30–50 high-value questions are enough to get started.
• Separate engines: record ChatGPT, Perplexity, Gemini and Copilot separately, with no overall average.
• Recording: note the presence state, the sources cited and how accurately your offer is described.
• Traceability: keep the date, the engine, the wording and the raw answer.

Reducing Noise in Answers

A single reading is fragile; a series of readings under a constant protocol is far less so. To stabilise your measurements:

• Keep the wording of each question stable and change only one parameter at a time.
• Repeat each question across several sessions and record presence across N attempts.
• Record on a fixed date, with no conversation history.
• Freeze the set between two reviews and classify new questions as "no baseline".

Before / After: Compare Each Page with Itself

The only comparison that holds compares a page with itself: same question set, same keywords, same engines, windows of the same length before and after the update. Comparing the site average from one month to the next mixes content, seasonality and scope. The fictional example below shows the expected reading after a page has been reworked:

Correlation, Never Causation

A before / after establishes a temporal correlation: the page changed, then its presence evolved. It does not prove that the update produced the citation, because the model, competing sources and the index change at the same time. Write "the page has been cited since its update", never "the update generated the citation". A change log (content updates, redesigns, campaigns) dates every break in trend.

The Pitfalls That Distort the Comparison

Before drawing conclusions, check that none of these biases distorts the reading:

• A mention without context inflates presence without producing a lead.
• The scope changed between the two periods: pages, markets or questions.
• Engines are mixed into a single average.
• A question not tracked before is counted as a gain.
• A rise in traffic comes from classic SEO, not from a new citation.

Linking KPIs to Business Results Without Overpromising

Part of GEO's value plays out without a click: the buyer comes back later through a brand search or directly. Attribution therefore combines the certain (sessions, measured leads) and the probabilistic (influence, assistance). Read it across three levels:

• Direct: conversions from AI referral traffic, measured in GA4.
• Assisted: changes in brand queries in Search Console and in direct traffic.
• Pipeline: lead quality, close rate and sales cycle length, by cohort.

Attribution Readings That Work in B2B

Last non-direct click attribution underestimates the influence of AI answers. An assisted reading tracks the most-cited pages (comparisons, proof points, documentation) within converting journeys. A before / after analysis by cohort isolates a group of updated pages and tracks their citations, AI traffic and conversions. Present the result as measured influence, never as causation.

Measuring Traffic from ChatGPT and AI Assistants in GA4

Assistants that display clickable sources send sessions identifiable by their source: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com. By default, Google Analytics 4 files them under referral traffic, mixed with other sites. Isolate them with this configuration:

• Create a custom channel group containing an "AI assistants" channel.
• Set the condition: session source matches the regular expression below.
• Place this channel before "Referral" in the channel evaluation order.
• Check that your key events (demo request, contact, sign-up) are active.
• Analyse the channel by landing page, engagement and key events.

A Regular Expression That Does Not Capture Bing

An expression containing "bing" files all of Bing's organic traffic under the AI channel and inflates the KPI with traffic unrelated to GEO. Restrict it to assistant domains: chatgpt\.com|openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com. Every quarter, add any new assistant domains that appear in your session sources.

What the AI Channel Does Not See

Clicks from an AI Overview or AI Mode arrive as google / organic: GA4 cannot separate them from the rest of Google Search. The AI channel also underestimates influence, because an answer without a click leaves no session. Read it as a floor.

Tracking AI Overviews and AI Mode in Search Console

AI Overviews (formerly SGE, renamed in May 2024) and AI Mode have been available in France since summer 2026. At the time of writing, Google Search Console counts impressions and clicks from these two surfaces in the Performance report, without distinguishing them from classic results or from each other. The KPI is therefore built page by page, before / after.

Being picked up as a source has measurable value: an average CTR uplift of +1.08% is observed when a site is cited as a source in an AI overview (Semrush, 2025). Track, keyword by keyword, the presence state in the AI Overview alongside the page's CTR. Market benchmarks are gathered in the GEO statistics.

To use Search Console without making it say what it does not:

• Filter the reworked pages and compare impressions, clicks and CTR before / after.
• Cross-reference with the presence state recorded in each keyword's AI Overview.
• Track the long, conversational queries that lead to your pages.
• Track brand queries, an indirect indicator of the influence of AI answers.

Reading the KPIs: From Finding to Action

A presence KPI is only worth the decision it triggers. When the brand is absent or loses citations, the following grid links the observed signal to the priority fix:

Deal first with lost citations on BOFU questions, then with absences on high-value keywords, and re-measure at the next recording.

Recording Cadence and Alert Thresholds

Technical disruptions need rapid detection; AI presence, which is more volatile, is read over series to avoid overreacting to a single reading. Set the cadence as follows:

Documented Alert Thresholds

Set thresholds from your own historical data rather than universal values, and trigger an alert on a series of readings, not on a single data point:

• Disruption: an indexing drop or error spike on a strategic directory.
• Erosion: AI share of voice declining over two or three consecutive readings.
• Accuracy: an answer that misdescribes your offer or confuses your brand.
• Business: a drop in key events from AI sessions.

Depending on Your Situation

Your starting point determines which page to read next:

• You need to present results to leadership: the GEO reporting template organises the sections, the reading by persona and the decisions.
• You are choosing a tracking tool: the guide to GEO tracking tools details the features to check.
• Your Google traffic is falling on informational queries: the article on AI Overviews and AI Mode explains how to get picked up.
• You need to quantify ROI and organise the team: the guide to rolling out GEO in your business covers roles, budget and roadmap.
• You do not yet have citable content: the GEO referencing method starts with a presence audit and the choice of questions.

To record these KPIs without a spreadsheet, Incremys' performance reporting module brings together, at the time of writing, positions, citations by question (gained, lost, stable, no baseline) and AI traffic, and compares each page with itself before / after its update.

FAQ on GEO KPIs

What Are the KPIs of GEO?

The KPIs of GEO are AI presence by keyword, citations by question, AI share of voice, AI referral traffic, Google impressions and clicks, and conversions. Each has a single source: recordings on a question set, GA4 or Search Console. A KPI that triggers no decision has no place on the dashboard.

How Do You Measure AI Share of Voice?

AI share of voice is measured as the share of questions in a fixed set where your brand is named in the answer. Calculate it engine by engine, on the same questions as your competitors, and segment it by intent. If the set changes, you are measuring a change in scope, not genuine progress.

How Do You See Traffic from ChatGPT in GA4?

ChatGPT traffic becomes visible in GA4 through a custom "AI assistants" channel defined on session source: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com. Place it before the referral channel and never include "bing" in the expression, or you will capture Bing's organic traffic. Then analyse landing pages and key events.

Does Search Console Measure AI Overviews?

Yes, Search Console counts impressions and clicks from AI Overviews and AI Mode in the Performance report, but without distinguishing them from classic results. The effect is therefore measured page by page, by comparing a period before and a period after an update, cross-referenced with the presence state recorded in the AI Overview.

Can You Prove That an Action Produced a Citation?

No, you can only establish a temporal correlation between an update and a citation. The model, competing sources and the index evolve in parallel. Good practice is to compare each page with itself before / after, on the same question set, and to date every change in a change log.

How Often Should You Measure?

The right frequency is weekly for critical questions and technical disruptions, monthly for citations, share of voice and conversions, and quarterly for reviewing the question set. Because AI answers vary from one session to the next, a trend is read across several consecutive readings, never from a single data point.

What Is the Difference Between a GEO KPI and a Classic SEO KPI?

An SEO KPI measures a page within a ranking: position, impressions, clicks. A GEO KPI measures a brand's presence in a generated answer: named, cited, as a source or absent. The two complement each other, because a page that is indexed and ranks well often remains a prerequisite for being picked up by generative engines.

What Should You Do If Your Brand Is Cited Without a Clickable Link?

A citation without a link is a visibility win, not an acquisition win: the "brand named" state counts towards share of voice but produces no session. Then strengthen the pages that can become clickable sources (guides, proof points, documentation) and, in your reporting, separate click-free visibility from web performance.