What are AI citations? How to check yours and why pages earn them

What is an AI citation?

An AI citation is a link an AI assistant shows to a web page behind its answer. If someone asks ChatGPT, Gemini, Claude or Perplexity a question and your page appears as a numbered marker, a source link or an entry in a sources list, your website has been cited. This page is about a website being cited by an assistant, not about citing AI in a paper, which style guides such as APA cover.

You will also see the term LLM citations. LLM stands for large language model, the kind of system behind these assistants, and the term means the same thing.

For a business, a citation is useful evidence, within limits. It shows that an assistant’s search reached a page of yours and treated it as relevant to that question. It does not show that anyone clicked, that your page is the best source, or that the sentence next to the link says what your page says.

AI citations vs mentions: what is the difference?

A citation is a link. A mention is a name. An assistant can give either one without the other, so each needs its own check.

None of the help pages or reports read for this guide gives a click rate for citations.

Semrush measured the gap in a study with Kevin Indig, published June 9, 2026. It logged 3,981 appearances of a domain in answers to 115 prompts, across 14 countries, on ChatGPT, Gemini, Google AI Overviews and Google AI Mode. In 61.7% of them the domain was a source link while the brand name stayed out of the answer, which the study calls a ghost citation. In 25.1% the brand was named with no link, and in 13.2% it got both.

The split changed by assistant. When a domain appeared in a ChatGPT answer, 87% of those appearances had a citation and 20.7% a mention. For Gemini it ran the other way: 21.4% and 83.7%.

That is a small set of prompts, Claude and Perplexity were not in it, and Semrush sells a tool that tracks these numbers. The safe reading is the narrow one: being cited and being named are different outcomes, so a report of one says little about the other.

Where does each assistant show its sources?

Each company describes this on its own help page. All four pages were read on November 4, 2026, and they change often.

No search, no source. An answer written from what the model learned in training has no page behind it to link to, so it cites nothing. That is how an assistant can name your business with no link at all, and why one question can show sources on one try and none on the next. Perplexity is the exception in its own description, with a search behind each answer. Our guide to Perplexity SEO covers that assistant on its own.

A list of sources is wider than a citation. OpenAI says its Sources list holds cited sources and other relevant links. Google describes its panel as sources and related content. So a link in a panel does not prove that a sentence came from that page. When you check, note whether your link sits beside a claim in the text or only in the list.

How do you check your AI citations by hand?

You need a spreadsheet and the four assistants. Work in this order:

• Write the questions. Ten buyer questions without your name, in the customer’s words, plus two that do name you. The first kind shows whether you get found. The second shows which pages an assistant reads about you.
• Make the search run where you can. In ChatGPT, open the tools menu and select Search. In Claude, add “Search the web” to the question, or turn on Web search from the + menu if your account has it. Gemini’s help page describes no switch for this, so record whether sources appear. Perplexity describes a search on every answer.
• Open every source. Select the citations in the text, then the Sources button where there is one. Look for your domain, and for pages about you on other sites.
• Open each page cited about you. Find the sentence that supports the claim.
• Keep the setup the same, and ask more than once. OpenAI says ChatGPT can use your approximate location and saved memories when it rewrites a search. Run each question at least three times, on different days: that is our suggestion, not any company’s rule.

Record each answer in a sheet like this one. The rows are invented, to show how to fill it in.

Then read the sheet by pattern:

• Cited, not named. Your page feeds the answer, but the reader may never learn who you are. Check whether the passage used carries your name at all. Adding it is a reasonable change, though no study we read tests it for a single site.
• Named, not cited. The name came from memory or from other sites. If sources are listed, read those pages first.
• Someone else cited about you. A directory or review page is speaking for you. Check its facts.

Your analytics adds one small piece: it sees only the citations someone clicked, as visits from domains such as chatgpt.com.

What free citation data do Microsoft and Google give you?

Two companies give site owners first-party data free of charge, through Bing Webmaster Tools and Google Search Console. Each asks you to verify that you own the site.

The Bing report is the one that counts citations by name and shows which of your pages were cited. On June 16, 2026 Microsoft added, in preview, query intents, topics, a comparison with an earlier period, and Citation Share, your part of all citations shown for a query.

Google’s report counts links shown, not citations per claim, and it may not appear for a site with too few impressions.

Do not read Microsoft’s number as ChatGPT’s. OpenAI’s help page links Microsoft’s privacy statement among its search providers, but Microsoft’s posts do not name ChatGPT as a partner.

None of the help pages read for this guide describes a citation report for site owners from OpenAI, Anthropic or Perplexity.

What does an AI citations checker do?

The name covers two different jobs. A citation tracker puts a set of questions to the assistants on a schedule and records what comes back: an outcome you cannot control directly. A site checker reads your pages to see whether bots can get in and whether there is text to quote: causes you can fix. One does not replace the other.

If you try a tracker, these questions tell you how to read its numbers:

• Which assistants, and through what? An answer through an API or an automated browser can differ from what a person sees, because the app can bring in location and memory.
• Is web search on, and how many runs per question? With search off an assistant has nothing to cite, and one run is one draw from many possible answers.
• Does it separate citations from mentions? A blended score hides the gap.
• Can you read the raw answer and its sources? Without them, you cannot check the claim.

Our guide to AI brand monitoring covers following your name across assistants over time.

What is known about why a page gets cited?

Less than you might expect, because none of the four companies publishes its ranking rules. Here is what each one says, as read on November 4, 2026:

• OpenAI says ChatGPT ranks results using multiple factors aimed at relevant, reliable information, without listing them, and that placement is not guaranteed.
• Google says, in a guide last updated July 10, 2026, that its AI features rest on its core Search ranking and quality systems.
• Anthropic says Claude processes multiple sources to find relevant content, and gives no criteria.
• Perplexity says it gathers from authoritative sources such as articles, websites and journals, without defining authoritative.

The rest is outside observation. These are studies we could read at their own pages:

Three cautious readings come out of these.

• Ranking for the exact question is not the whole story. Many cited pages do not rank for the words typed, possibly because assistants search in their own words: OpenAI says ChatGPT typically rewrites a question into targeted queries when it uses search providers. Rankings still count where Google says they do, in its own AI features.
• Large domains take a visible share, and it moves. A fall from close to 60% to around 10% within weeks shows how little one snapshot tells you.
• Specifics may help a page once it is found. That comes from a lab setup on a 2023 model: a direction, not a guarantee.

We found nothing published on the weight each assistant gives to a page’s content, domain or age, on how often a citation gets clicked, on how soon a change earns or loses one, or on what an ordinary small-business site can expect. Treat any exact formula as a claim to check. The parts within your control are in our guide on how to get cited by ChatGPT.

Does the cited page always support the answer?

No. A link next to a sentence does not prove the page says it. OpenAI’s help page warns that search results and citations “can be incomplete, outdated, or incorrect”, and tells users to open a cited source and check it.

Two studies have measured versions of this problem.

Nelson Liu, Tianyi Zhang and Percy Liang had people audit answers from Bing Chat, NeevaAI, Perplexity and YouChat in “Evaluating Verifiability in Generative Search Engines”, posted in April 2023. On average, 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported the sentence they were attached to. Those were 2023 products, so the figures do not measure today’s versions.

The Tow Center for Digital Journalism published “AI Search Has a Citation Problem” in the Columbia Journalism Review on March 6, 2025. It gave eight search tools, including ChatGPT Search, Gemini and Perplexity, excerpts from 200 news articles and asked each for the headline, publisher, date and URL, 1,600 queries in all. Together they answered more than 60% incorrectly, with Perplexity at 37%. More than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs, and some tools linked to syndicated copies instead of the original. Each excerpt ran once, the authors say the task may not reflect typical use, and it covered news only. Claude was not tested.

For a business, this means a citation can attach your name to something you never said, or send the credit for your text to a copy on another site. So sort every page cited about you into one of these:

• Supported. The page says it, and it is current.
• Out of date. An old price, old hours. Update that page and any older page that repeats it.
• Not supported. The page does not say it. Note the answer, date and assistant, and if your page is vague on the point, state the fact plainly.
• A copy. Another site republished your text and got the link. Check that your original is indexed, and ask the other site to credit it.

A citation count measures none of this: Microsoft’s report shows how often pages appear as sources, not whether they were represented correctly.

Where the figures come from

The Ahrefs and Semrush figures are each company’s own study on its own data, and both sell tools that track AI citations. None is a census of the market. All sources were read at their own pages on November 4, 2026.

• Ahrefs: “Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt”.
• Semrush: “Why 62% of AI citations don’t lead to brand mentions” and “The Most-Cited Domains in AI: A 3-Month Study”.
• Research: Aggarwal and others, “GEO: Generative Engine Optimization”, arXiv 2311.09735; Liu, Zhang and Liang, arXiv 2304.09848; the Tow Center study.
• Company pages: the web search and sources help pages of OpenAI, Google, Anthropic and Perplexity, Search Console help, Google’s generative AI guide, and Microsoft’s Bing posts of February 10 and June 16, 2026.

What the MWY checker does and does not do about citations

The free AI visibility checker does not count citations, because it asks no assistant anything. It reads your site as raw HTML without running JavaScript, which is how outside tests have mostly found AI bots reading pages, and scores 24 items in three groups: can AI read it, does AI know who you are, and can AI quote it. That covers the conditions on your side of a citation: whether robots.txt allows the search bots, whether a firewall turns them away, whether your text is in the raw HTML, and whether your pages state answers, prices and dates. You get the score and the 5 items that matter most in a few seconds, with no sign-up.

Pro, which opens soon, adds the answers themselves. Every week, 20 buyer questions go to ChatGPT, Gemini, Claude and Perplexity through their APIs with web search on, and the report shows which pages each answer cites and what changed since the week before. Answers through an API can differ from what one person sees in the app.

The checker cannot tell you why an assistant chose another page, and it does not test whether a cited page supports the sentence beside it: that step stays with you. It does not promise that you will be cited, because nobody outside the AI companies controls that.