If you run a business or lead its marketing, you need to know which websites AI tools cite when they answer your buyers’ questions, and whether yours is one of them.
An AI citation audit answers that. You write a fixed set of prompts, run them on the same AI platforms, record every source, and compare the log month to month.
The audit measures how often your brand appears in AI answers, shows where your AI search visibility has gaps, and helps you decide which content to fix first. You do not need paid tools.
You need a spreadsheet, a few AI assistants, your Search Console account and a clear list of the questions your customers ask. The routine builds on the twelve questions in Am I visible in ChatGPT?, which checks whether you can be found. This one records who actually gets named.
The phrase “citation audit” also has an academic meaning: a systematic check that the references in a research paper are real and accurate.
One 2025 preprint describes an AI agent that checks every reference against databases such as Semantic Scholar, Google Scholar and CrossRef, and flags fabricated references, retracted articles, orphan citations and predatory journals (arXiv, 2025).
This guide covers the marketing meaning, which is about the sources AI systems cite when they answer questions about your category.
Most owners test AI visibility once, see one answer and draw a conclusion. A single answer proves very little.
A logged routine shows citation patterns you can act on, and it shows the competitors and third party sources that AI systems prefer to your own pages. The practical implication is simple: you learn what to fix before you spend on more content.
Mentions and citations are different results, and an audit records both
Two things can happen when an AI tool answers a question in your category. It can name your brand in the text, which is a brand mention. It can link to one of your pages as a source, which is a citation. Either can happen without the other.
That difference decides what you fix. A brand that is mentioned but never cited is known, yet its pages are not the evidence the system reaches for. A brand that is cited but never mentioned has useful pages and a weak name. Both cases need different work.
Citations also tell you something a mention cannot. They show the sources AI systems rely on for your topic. If a review site, a trade publication or a competitor’s guide appears again and again, that page is part of how your category is explained to buyers.
| Result | What you see in the answer | What it tells you |
|---|---|---|
| Mentioned and cited | Your name appears and your page is linked | Your pages and your name are working together |
| Mentioned, not cited | Your name appears, but links go elsewhere | You are known, but other pages are the evidence |
| Cited, not mentioned | Your page is linked, your name is absent | Your content is useful, your brand is not attached to it |
| Neither | Competitors and independent sources fill the answer | You are not yet part of the answer set |
Traditional search reports cannot show citation patterns, so the audit adds a second view
Google rankings and organic traffic tell you how a page performs in traditional search. They do not tell you whether an AI engine used that page as proof for an answer. A page can rank well and never be cited, and a page with modest rankings can be cited often.
That is why measuring AI visibility needs its own method. An audit identifies three things: where you appear, which sources are cited instead of you, and which pages need work.
Traditional SEO reporting focuses on traffic. The audit asks whether your authority and relevance are recognized inside the answer. Traditional SEO still matters, because the same technical SEO and useful material feed both views.
The methodology below is deliberately simple. It produces a full report you can read in one sitting: a prompt set, a log, a few rates, a list of cited domains and a fix list.
Six steps turn the audit into a routine you can repeat every month
Run the steps in order the first time. After that, steps one and two are already done, and each later month starts at step three.
1. Write a fixed prompt set of 15 to 20 buyer questions
Write the questions a real buyer would ask, in their words, without your brand name in them. Group them into four types so the log shows where you are strong:
- Discovery: “Who are the best [service] providers in [city]?”
- Comparison prompts: “[Option A] or [option B] for a business like mine?”
- Problem: “How do I fix [the problem you solve]?”
- Brand: “What is [your company] known for?”
Keep the exact prompt wording and never edit it later. A changed exact prompt changes the result. If you need a new question, add it with a new ID and keep the old ones. A fixed set is what makes this month’s log comparable to last month’s.
Number each prompt, such as D1 for discovery or B2 for brand. The ID is what you will record in the log.
2. Fix the test conditions before the first run
AI answers shift with who is asking and where. OpenAI states that ChatGPT may use an approximate location from your IP address, and that saved memories can influence how it rewrites a search query (OpenAI Help Center, 2026). Your own account can therefore see a different answer than a stranger does.
So choose one test setup and write it down. Use a logged out window or a clean account with memory switched off, one location, and a new chat for every prompt so that earlier answers do not shape later ones.
Pick each AI engine where your buyers look. For most businesses that means ChatGPT, Google AI Overviews and AI Mode, Perplexity and Gemini. Google’s AI Overviews appear on some searches and not others, so note when one does not show.
Record the platform, date and account type in the first row of your sheet.
3. Run every prompt three times on every platform
One run is a guess. SparkToro and Gumshoe.ai asked 600 volunteers to run 12 prompts through ChatGPT, Claude and Google’s AI, 2,961 runs in total. They found a less than 1 in 100 chance that ChatGPT or Google’s AI would give the same list of brands in any two responses (SparkToro, 2026).
The same study found that appearing often across many runs was still a useful signal of how strongly a system associates a brand with a topic. Rank order was not. In one example, one agency appeared in 85 of 95 Google AI responses to a prompt.
Fishkin disclosed that Gumshoe.ai sells AI tracking, so read the study as one careful data point and not as a law. Its practical lesson holds either way: repeat the prompt and count appearances.
Three runs per platform is a coarse sample, not the dozens the study used. Treat results as direction. Re run the set each month and look for patterns that repeat across prompts and across months.
4. Log every answer: mention, citation, cited domains and accuracy
Open a sheet with one row per run. These columns are enough:
- Date, platform, prompt ID and run number
- Brand mentioned (yes or no)
- Your page cited (yes or no, and which URL)
- Every cited domain, in the order shown
- Competitors named
- Description accurate (yes, partly, no)
Click through to the sources that are cited. Perplexity states that each answer includes numbered citations linking to the original sources (Perplexity Help Center, 2026). ChatGPT shows citations and a Sources list when it searches the web.
Check that each cited page really supports the claim next to it. OpenAI itself warns that citations can be incomplete, outdated or incorrect, so a link in an answer is a lead to verify, not proof.
5. Read the log for citation rate, citation share and cited domains
At the end of the runs, count three numbers.
Mention rate is the share of runs where your brand was named. Citation rate is the share of runs where a page of yours was linked. Citation share is your links as a share of all links shown across the set. Together they act as a simple AI citation score you can track each month.
Then list every cited domain and every one of the cited URLs, and count how often each appears. This is the most useful part of the audit.
The top cited domains are the sources AI systems rely on for your topic, and you will often find directories, review sites, trade media and competitor guides among them.
Forums and community sites can appear too, so note them when they do. Also note the type of page cited, such as a long educational guide, a service page or a directory entry. Over several months that shows which formats the answers in your category favor.
Sort the log by prompt type. A brand prompt that returns a wrong description is a different problem from a discovery prompt where you never appear.
Here is how a result looks, using made up numbers for a fictional bookkeeping firm.
| Prompt type | Runs | Mentioned | Cited | Most cited domains (illustration only) |
|---|---|---|---|---|
| Discovery | 12 | 2 | 0 | review site, accounting association, competitor guide |
| Comparison | 12 | 3 | 1 | competitor guide, software vendor page |
| Problem | 12 | 0 | 0 | government guidance, trade publication |
| Brand | 6 | 6 | 4 | own website, directory listing |
In this invented example the firm is known when asked by name, and absent when a stranger asks a category question. That points to third party evidence and plain answers on its own service pages, not to more brand posts.
6. Add the platform data you own, then write a short fix list
Your log shows what AI tools said. Your own data shows what they fetched.
Microsoft introduced AI Performance in Bing Webmaster Tools in February 2026 as a public preview. It reports total citations in Microsoft Copilot, AI generated summaries in Bing and select partner integrations.
It also shows average cited pages, the URLs cited most, and grounding queries, the phrases the AI used to retrieve your content (Bing Webmaster Blog, 2026). Microsoft notes that grounding queries are a sample of overall activity.
Google does not split AI features out. It states that sites appearing in AI Overviews and AI Mode are included in overall search traffic in Search Console, in the Performance report under the Web search type (Google Search Central, 2026).
Use Search Console to watch impressions and clicks on the pages your log shows being cited.
Then write the fix list: at most five items, each tied to a row in the log. For example, “Discovery prompts D1 to D4 never cite us; rewrite the service page to answer D1 in its first two lines.” Our guide to setting up Search Console, GA4 and AI search tracking covers the setup.
Share the fix list with your leadership teams on one page: the three rates, the top cited domains and the key takeaways.
If comparison prompts show competitors named instead of you, comparison pages that state your differences plainly are a sensible first fix. Re run the same prompts next month and compare the three numbers against this month’s.
Pages that are never cited usually fail one of five citation readiness checks
When your log shows a gap, check the page that should have been cited. These AI citation audit checks look at your owned content first.
Google states that to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, with no additional technical requirements. It also says important content should be available in text and that structured data must match the visible text.
That gives you the first checks for AI citation readiness:
- Is the page indexed, and free of a stray noindex or nosnippet tag?
- Are the key facts in plain text, not only in images, tabs or PDFs?
- Does the page open with a direct answer to the prompt?
- Do your name, services and description match on your site, LinkedIn and business listings? That entity consistency makes the entity clarity retrieval systems need.
- Does anyone independent say the same about you? Independent corroboration backs up what your own pages claim.
Microsoft gives matching advice for pages cited in Copilot and Bing. It recommends clear headings, tables and FAQ sections, claims backed by examples and data, regular updates for topical freshness so AI systems reference the current version, and the same description of each entity across text, images and video.
It also notes that pages cited for specific grounding queries often show a clear subject focus, so depth and specificity on a narrow question can reinforce authority (Bing Webmaster Blog, 2026). IndexNow is its suggested way to tell search engines when content changes.
The last two checks above are AI citation signals you build outside your own site. If the cited domains in your log are review sites and trade publications, your next move is earning a place on them.
That is public relations, digital PR and earned media work: real case material, original data or original research, and clear definitions that editors want to quote.
Make sure the visible content on the page says the same thing as your structured data. Google also states there is no special schema.org markup or AI text file needed, so fix the plain text first.
The wider causes are covered in what AI visibility is and in brand visibility in AI search. The pillar guide, What Is GEO?, puts the audit in the context of generative engine optimization.
Read the results with care, because AI answers vary and no audit is exact
Treat a manual audit as a map, not a measurement. Three limits matter.
First, answers are not stable. The SparkToro study found that the order and the length of recommendation lists changed from run to run, which is why this routine counts appearances and ignores rank position.
Second, your setup shapes the answer. Location, account memory and platform version all change what you see, so keep conditions fixed and note any change in the log.
Third, a mention is not a lead. Compare your results with inquiries, branded search and the number of times customers say they found you through an AI tool.
The monthly numbers in How to Measure Search and AI Visibility show where AI citation fits among the rest of your reporting, including share of voice against competitors.
Review the third party content that shows up in your log, because it is what the answers draw on. Paid AI visibility tools automate the same job at larger scale, with their own prompts and scoring. A manual audit is a sound way to learn what the tools are measuring before you pay for one.
Take the next step with the ebook and the weekly letter
If the audit shows you are absent, the cause is usually a page an engine cannot quote, or too little independent evidence. When AI Answers First is our free ebook and workbook on being the business an AI tool names, including how to measure it without buying certainty.
For one argument and one test a week, subscribe to Ground Truth, our newsletter, sent every Wednesday. If you want a second pair of eyes on your distribution and search work, see Distribution Strategy.
FAQ
What is an AI citation audit?
An AI citation audit is a structured check of which sources AI tools link to when they answer questions in your category. You run a fixed prompt set, log whether your brand is mentioned and whether your pages are cited, and compare the results over time.
How often should I run an AI citation audit?
Monthly is a good rhythm. Answers change from run to run, so a single check says little. A monthly routine with the same prompts, platforms and test conditions shows whether your citation rate is rising or falling.
How do I check whether AI cites my website?
Ask your buyer questions in ChatGPT, Perplexity, Gemini and Google AI Overviews, open the cited sources, and record whether a page of yours is linked. Bing Webmaster Tools also reports citations in Copilot and Bing, and Search Console includes AI feature traffic in the Web search type.
What is the difference between an AI mention and an AI citation?
A mention is your brand name appearing in the answer text. A citation is a link to your page as a source. You can be mentioned without being cited, and cited without being mentioned, so the audit records both.
Can I audit AI citations without paid tools?
Yes. A spreadsheet, a fixed list of prompts, a few AI platforms, Search Console and Bing Webmaster Tools are enough to run a useful audit. Paid tools add scale and automation, but they use their own prompts and scoring, so check what they measure.