AI citation share: what it is and how to measure it

When ChatGPT, Gemini, Perplexity or Copilot answers a question with web search, it lists sources. Citation share asks a simple question about those sources: out of all of them, across the questions that matter to you, how many are yours? It is the AI version of share of voice, and it is the closest thing to a ranking metric for AI answers.

1. What citation share means

A citation is a link to a page that an AI engine shows as a source for its answer. Citation share is your part of all those links, for a defined set of questions. Three things define the number, and it means nothing without them:

• The prompt set. Which questions you ask. A different set gives a different share.
• The engine. ChatGPT, Gemini and Perplexity cite different sources for the same question.
• The period. Answers change over time, so a share belongs to a week or a month.

2. The formula

Pick one counting rule and keep it. The simplest is to count each domain at most once per answer, so a page cited three times in one answer counts once. Whatever you choose, apply it to the numerator and the denominator the same way, or the share will drift for reasons that have nothing to do with your site.

3. How to measure it by hand

You need a spreadsheet and about an hour per engine per run.

1. Write at least 10 questions your customers actually ask. Use the problem, not your brand name.
2. Freeze the list. Changing questions between runs changes the denominator and breaks the trend.
3. Ask each question with web search on, in a fresh session (a new chat or a logged-out window), so your history does not steer the answer.
4. Log every cited domain for every answer, one row per citation: date, engine, question, domain.
5. Count your domain, count all domains, divide. Do this per engine.
6. Repeat the whole panel a few times in the same period and report the range.

The same panel also shows who is cited instead of you. Those domains are your real competitors in AI answers, and they are often not the ones you compete with in Google. For a worked version of this panel for one engine, see is my site cited by ChatGPT.

4. Why the number is noisy

AI answers are not fixed. The same question asked twice can cite different sources, and each engine uses its own search and ranking. That has three consequences:

• Use at least 10 prompts. With a handful of questions, one answer changing moves the share a lot.
• Repeat runs within a period and report a range, as in the example above.
• Never add engines together. A share in ChatGPT and a share in Gemini are two numbers, not one.

A change inside your normal range between runs is noise. A trend that holds across several periods, on a frozen panel, is a signal.

5. How it differs from brand mentions and AI referral traffic

Three metrics get mixed up. They measure different things:

• Citation share counts your pages listed as sources, relative to everyone else’s. It needs a prompt panel because it has a denominator.
• Brand mentions count your name appearing in the answer text. An engine can name you without linking you, or link you without naming you.
• AI referral traffic counts people who clicked a citation. In GA4 it shows up as session source chatgpt.com, perplexity.ai and similar. It proves a real citation, but it misses every citation nobody clicked and it cannot see your competitors.

Use referral traffic as proof that citations happen and turn into visits. Use citation share to see how much of the answer space you hold and whether that is growing.

6. Tools that measure it

A number of AI-visibility tools run prompt panels for you and report share of voice or citation share across engines. They differ in which engines they query, how they sample and how they count. We keep a side-by-side comparison of AI-visibility tools. When you compare tools, ask each one which prompts it runs, how often, and how it counts a citation, because those choices set the number.

7. Where CitationDesk fits

Our free Citation Readiness Score measures readiness, not share. It audits one URL for what an AI engine needs before it can quote a page: crawler access, content that can be read without running JavaScript, a clear direct answer, structured data and a recognisable entity. A page that fails those checks is unlikely to earn citations, so it is a sensible first step before you start counting.

CitationDesk Pro (€29/month) runs the questions you choose every week against Gemini grounding and OpenAI web search, and records which domains are cited. Perplexity is being added and is not in the reports yet. That replaces the hand-run panel above for those two engines.

Questions people ask

See also tracking citations across four LLMs and is my site cited by ChatGPT.