AI Citation Tracking: How to See Which Sources AI Answers Cite
How to track AI citations: what counts as one, the free reports from Bing, OpenAI and Google, a manual protocol, the numbers to log, and what to do next.
By Citlyze Research. Published . 5 min read.
On this page
AI citation tracking means recording which web pages AI answers link to as sources, for the questions your buyers ask, over repeated runs. It tells you two things a mention count can't: whether your own pages feed the answers, and which third-party pages shape what engines say about your category. You can track citations for free through the reports Bing, OpenAI and Google now publish, by sampling answers yourself, or with a tool that stores every answer's source list.
This guide covers what counts as a citation on each engine, the free first-party data, a manual protocol, the numbers worth logging, and how to turn a citation list into work.
What counts as a citation
A citation is a link an AI answer points to as its source. It is a different event from a mention. An answer can name your brand without linking you, link your page without naming you, or cite a review site that recommends a competitor. Track the two separately and the gap between them becomes useful: named but not cited usually means the engine trusts someone else's page for the facts about you.
Engines show citations differently:
| Engine | When it cites | Where the sources appear |
|---|---|---|
| Perplexity | On almost every answer | Numbered sources beside and inside the answer |
| ChatGPT | When it searches the web | Inline links and a sources list |
| Claude | When it searches the web | Citations in every search answer, per Anthropic |
| Copilot | On most answers | Linked sources under the answer |
| Google AI Overviews and AI Mode | When the feature appears | Source cards and links beside the answer |
| Gemini | When it grounds an answer in search | Source links under the answer |
The "when it searches" condition matters. In our 200-prompt study, ChatGPT cited at least one source in only 39% of answers in August and 40% in September, so the rest of its answers had no citations to track at all. The Q3 citation sources study has the full numbers.
Free first-party data
Three engines now report part of the citation picture for your own site:
- Bing Webmaster Tools, for Copilot. The AI Performance report counts how often Copilot cited each of your URLs and shows samples of the grounding queries that retrieved them. It is the most complete free citation report any engine offers. Copilot rank tracking walks through it.
- ChatGPT Business Hub. After you claim your business and pass an identity check, OpenAI shows 28 days of mentions, clicks and topics. For an online business, a mention counts only when ChatGPT references your website, which makes it close to a citation count. The Business Hub guide covers the definitions.
- Google Search Console. The Generative AI performance report counts your pages' impressions in AI Overviews and AI Mode. It shows that you appeared, not for which question or next to whom. The report guide covers what it leaves out.
All three show your own site only. None tells you which competitor or publisher got cited instead of you, and that is usually the more useful half.
A manual protocol
Sampling answers by hand fills the competitor gap and costs an afternoon a week:
- Take 15 to 25 buying questions, written the way buyers ask them. Finding buyer prompts shows how to build the set.
- Run each one three to five times per engine in fresh sessions. Citations vary between runs more than mentions do, so a single run tells you little.
- For every answer, log each cited URL, its domain, and its position in the source list.
- Label each domain: yours, a competitor's, or a third party (publisher, review site, community, marketplace).
- Note what each citation supports. A review site cited for your pricing is a different problem from one cited for your competitor's features.
Keep the prompt wording fixed between weeks. Changing a question resets its history, and you won't be able to tell a real shift from a reworded prompt.
The numbers to log
Four numbers cover most decisions:
- Answer-level citation rate. The share of answers that cite at least one of your URLs. It answers "how often are we part of the evidence?"
- Share of citations. Your citations divided by all citations in those answers. It falls when engines cite more sources per answer, even if nothing about you changed, so read it beside the rate. What makes a good AI citation rate explains both and gives reference ranges.
- The cited-domain leaderboard. Every domain ranked by how many answers cite it, per engine. This is your map of whose pages decide the answers.
- The owned versus third-party split. In our Q3 data, publishers and independent review sites took 41% to 54% of citations, and category vendors' own sites 12% to 14% on ChatGPT and 5% to 6% on Perplexity. Expect most of your category's citations to go to pages you don't own.
Report each with the number of answers behind it. A rate from 100 answers and a rate from 10 are different claims.
From citation list to work
The leaderboard turns into two lists:
Pages to fix. Your own URLs that get cited for the wrong fact, or that should be cited for a question and aren't. Common causes: the answer sits low on the page, the page can't be read without JavaScript, or a competitor's comparison page answers the question more directly. The GEO audit guide covers the technical side.
Sources to win. Third-party pages that engines cite repeatedly for your category and that leave you out or describe you wrongly. Those are your outreach, PR and review targets, in order of how many answers cite them. An AI visibility gap analysis turns the list into a plan.
Then re-measure. Engines change their source mix without warning: between our August and September runs, Reddit fell from 13.5% of Perplexity answers to 1.5%. A citation strategy built on one month's leaderboard can go stale within weeks, so check the list every month.
Tracking it with a tool
Manual tracking stops scaling past about 25 prompts on two or three engines. A tool runs the prompts on a schedule, stores every answer's source list, labels domains, and trends the numbers. When you compare tools, check that they store the full cited URL list for each answer, not just whether your domain appeared; without it you can see the rate but not the reason. The best AI visibility tools roundup compares the field.
Citlyze's citation tracking keeps the cited domains and URLs of every answer it collects for your prompts, across ChatGPT, Gemini, Perplexity, DeepSeek, Google's AI Overviews and AI Mode, and Copilot, split into owned, competitor and third-party, with the run count beside every rate.
Keep reading

Measurement
How to Fix Negative Brand Sentiment in AI Answers
AI models repeat old narratives. How to audit what ChatGPT and Gemini say about your brand, find the sources feeding the negativity, and fix them.
, 7 min read

Measurement
What Is AI Visibility? Metrics, Measurement and What Good Looks Like
AI visibility is how often AI systems mention, cite, and recommend your brand. The five metrics that define it and how each one is measured.
, 11 min read

Citations
What's a Good AI Citation Rate? How to Benchmark Honestly
How to calculate an AI citation rate, reference ranges from our 200-prompt study, why industry benchmarks mislead, and how to build a baseline.
, 5 min read