GEO metrics: what to measure in AI answers and how

GEO metrics measure how AI assistants show your brand. The core set: mention rate (share of answers that name you), citation share, AI share of voice, position in the answer, sentiment and accuracy, cited sources, AI referral traffic and conversions, prompt coverage and engine spread. Track them monthly on a fixed prompt set.

Below: each metric with a definition, a formula where one exists, free and paid ways to measure it, and the traps. Then a table with all nine and a monthly report template. It is part of our guide to LLM SEO. New to the term? Start with SEO vs GEO vs AEO.

What do you need before you measure anything?

Every GEO metric is counted from the same raw data: answers to questions your buyers ask. Collecting them on a schedule is called prompt tracking. Keep the setup fixed, or your numbers will move for reasons unrelated to your work.

1. Write 20 to 50 prompts. Mix category questions ("best help desk software for small teams"), comparisons ("X vs Y") and questions about your brand ("is X good for agencies").
2. Pick the engines: for example ChatGPT, Perplexity, Gemini, Claude, Copilot and Google AI Overviews.
3. Run each prompt at least three times per engine, on the same days each month.
4. Save every answer with its date, the engine, the brands named in order and the sources cited.
5. Pick 3 to 6 competitors to compare against, and keep the list fixed.

Step 3 is not optional. Prefer asked 80 questions three times on four engines (960 answers, collected 13 September 2026). Two runs of the same question shared on average 37% of their cited sites on ChatGPT and 35% on Gemini, against 68% on Claude and 96% on Perplexity. One run is a sample, not a result.

What are the main GEO metrics?

1. Mention rate: how often are you named?

Definition: the share of tracked answers that name your brand at all. Tools often call this visibility or an AI visibility score. Peec AI documents its version as "Visibility Score = (Number of responses mentioning your brand / Total responses) × 100" (Peec AI docs, read 1 November 2026).

• Formula: answers that name you ÷ all tracked answers × 100.
• Free: run the prompts by hand and mark yes or no per answer in a spreadsheet.
• With a tool: an AI visibility tracker runs the prompts and counts for you.
• Pitfall: tools calculate their scores in different ways, some with weighting. Two scores from two tools are not comparable.

2. Citation rate and citation share: are your pages used as sources?

Definition: citation rate is the share of answers that link to your domain. Citation share is your part of all citations in those answers. Bing defines its Citation Share as the "percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query" (Bing, 16 June 2026). Bing calls it an observational metric, not a ranking system.

• Formulas: citation rate = answers that cite your domain ÷ all answers × 100. Citation share = citations to your domain ÷ all citations × 100.
• Free: Bing Webmaster Tools has an AI Performance report with total citations, average cited pages, grounding queries and page-level citation activity, for Microsoft Copilot, AI summaries in Bing and some partners (Bing, 10 February 2026).
• With a tool: trackers count citations across the engines you pick.
• Pitfall: decide whether a URL cited twice in one answer counts once or twice, and keep that rule.

Expect a small number on category questions. In the Foundation and AirOps study of 5.1 million AI responses for 50 B2B brands, 2.24% of citations on unbranded questions went to the brand's own pages. The rest went to pages the brand does not own.

3. AI share of voice: how often are you named compared with competitors?

Definition: your brand mentions as a share of all brand mentions in the answers you track. Peec AI's formula is "SoV = (Number of times your brand was mentioned) /(Total number of brand mentions) × 100". Its example: 4 mentions for you and 12 for a competitor give you 25% (Peec AI docs).

• Formula: your mentions ÷ (your mentions + every tracked competitor's mentions) × 100.
• Free: count names per answer in your spreadsheet. Fix the competitor list first, because adding a competitor lowers your share.
• With a tool: some weight each prompt by search volume; Ahrefs Brand Radar uses an AI-adjusted search volume. Ahrefs notes that "two tools can report different AI SOV for the same brand without either being wrong" (Ahrefs, 22 October 2026).
• Pitfall: share is not presence. Peec's docs use the same 4 mentions in 10 chats for both metrics: a 25% share of voice, but a 40% visibility.

4. Position in the answer: are you named first or last?

Definition: where your brand appears in the order of brands an answer names. Peec AI describes position as "the average ranking across all responses where your brand appears", counting every brand it detects, not only your competitors (Peec AI docs).

• Formula: sum of your positions ÷ number of answers that name you. Lower is better.
• Free: note your place (1st, 2nd, 3rd) in each answer that names you.
• Pitfall: average only over answers that name you, and report it next to mention rate. A position of 1.5 in 3 answers out of 100 is not a strong result.

5. Sentiment and accuracy: what do the answers say about you?

Definition: sentiment is the tone of the answers that name you. Accuracy is whether the facts are right: what you sell, who it is for, your price, your integrations. Peec AI scores sentiment from 0 to 100 and says "most sentiment scores fall between 65 and 85" (Peec AI docs).

• Formula for accuracy: answers with no wrong fact about you ÷ answers that name you × 100.
• Free: write five facts the answer must get right, and check each answer against them. Mark tone as positive, neutral or negative.
• With a tool: most trackers score sentiment. Few check facts against your own list, so read a sample by hand.
• Pitfall: a neutral answer with a wrong price hurts more than a positive one. Report wrong facts as a separate list, with the source the answer used. A wrong fact is often a hallucination or an old third-party page.

6. Cited sources: which pages do the assistants read for your category?

Definition: the domains and URLs cited in answers to your prompts, sorted by how often they appear, with a column for whether each names you. Not a single number, but often the most useful part of the report.

• Formula: none. Count citations per domain and sort.
• Free: copy the source links from Perplexity, ChatGPT search and Gemini answers into your sheet.
• Pitfall: sources vary a lot by engine. Prefer found that 72.7% of the 1,329 sites cited in its 960 answers were cited by one engine only, and 2.2% by all four (Prefer). Keep the list per engine.

Lists are a big part of what you will find. When Prefer checked 188 answers in July 2026, 40% of the sources cited for buyer questions were independent lists. Vendors' own sites got 34%, review sites 3%. A cited best-of list that leaves you out is a work item.

7. AI referral traffic and conversions: does it bring visits and deals?

Definition: visits from links in AI answers, and the signups or demo requests they lead to. Google Analytics 4 has a default channel for it: "AI Assistant is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok" (Google Analytics Help, read 1 November 2026).

• Formulas: AI conversion rate = conversions from the AI Assistant channel ÷ visits from it × 100. Share of traffic = AI Assistant visits ÷ all visits × 100.
• Free: the Traffic acquisition report in Google Analytics 4, plus a "How did you hear about us?" field on your demo form.
• Pitfall: Google's Organic Search channel includes "Google's AI Overviews and AI Mode". Search Console also reports AI features traffic inside the Performance report, "within the 'Web' search type" (Google Search Central, last updated 2025-12-10). Google says this traffic is "included in the overall search traffic", so you cannot split Google's AI clicks out there.
• Pitfall: many answers end without a click (zero-click search), and an answer that cites a list sends the click to the list first.

Small volumes can still matter. Semrush found the average visitor from AI search is 4.4 times as valuable as the average traditional organic search visitor, based on conversion rate (July 2025; visitors from non-Google AI search such as ChatGPT; sample size not published). Track conversions, not only visits.

8. Prompt coverage: on how many buyer questions do you appear?

Definition: the share of your prompts where you are named in at least one run. Mention rate counts answers; prompt coverage counts questions. It shows whether you are missing whole topics or only appear by chance.

• Formula: prompts with at least one mention ÷ all prompts × 100. A stricter version counts only prompts where you are named in every run.
• Free: group your sheet by prompt and mark each one covered or not.
• Pitfall: coverage depends on which prompts you chose. Add new prompts as a separate group, so the old trend line stays comparable.

9. Engine spread: on which assistants do you show up?

Definition: your mention rate per engine, side by side, and the number of engines where you clear a threshold you set (for example, named in 20% of answers).

• Formula: engines where your mention rate is above your threshold ÷ engines tracked.
• Free: a pivot table of mention rate by engine.
• Pitfall: engines cite very different amounts. Prefer counted on average 19.48 sources per answer on Perplexity, 8.64 on Gemini, 4.58 on Claude and 3.05 on ChatGPT (Prefer). Compare each engine with itself over time, not with the others.

Which GEO metrics should you track? (all nine in one table)

If you track only three, pick mention rate, cited sources and AI-assisted conversions. The first tells you where you stand, the second what to work on, the third whether it pays.

What should a monthly GEO report contain?

One page, the same layout every month, so anyone can see what moved. Copy this template.

What are the common mistakes with GEO metrics?

• Trusting one run. Answers vary between runs, most on ChatGPT and Gemini in Prefer's data. Use three runs or more and report averages.
• Changing the prompt set. A new prompt changes every rate. Keep a fixed core set and test new prompts apart.
• Comparing tools. Each vendor counts and weights differently. Compare a tool with itself over time.
• Reading a citation as a click. A citation is not a visit. Bing's AI Performance report counts citations, not clicks, so check visits in Analytics.
• Missing Google's AI clicks. They sit in Organic Search in Analytics and in the Web search type in Search Console.

How do the metrics turn into work?

Most moves come from the cited sources list. If the assistants cite three list articles for your category and you are in one, the work is to get into the other two: pitch the writer, give them a reason to add you, or book a spot. Our AI visibility audit walks through that check step by step.

Rank Monthly is one route for the booking part. It places brands in list articles, sponsored articles and LinkedIn posts, and keeps one ledger with every placement, page visit, click and AI mention, month by month. It is not a full prompt tracker; see Rank Monthly vs AI visibility trackers. Nobody can promise an AI mention: the assistants decide. Pricing on request. Read more on the page for businesses.