What is an LLM citation? A guide for business owners

LLM citations are the clickable source links that AI assistants such as ChatGPT, Perplexity, Claude, Copilot and Google's AI Overviews attach to their answers, crediting a specific web page for part of what they just said. If one of those links points at your website, you have been cited. This guide explains what LLM stands for, how each tool shows its sources, how the pages behind an answer get chosen, why the same question can produce different sources on different days, and what a business can realistically do to be picked.

It is for business owners and marketers who keep hearing they should "get cited by ChatGPT" and want a plain explanation first. The claims about how the tools work come from what Google, OpenAI, Microsoft, Perplexity and Anthropic publish about their own systems, checked in October 2026, plus the robots.txt standard, the original 2020 research paper on retrieval and one independent accuracy study.

What is an LLM citation?

LLM stands for large language model, the kind of AI system behind ChatGPT, Claude, Gemini, Copilot and Perplexity. A language model learns from an enormous amount of text during training and uses that to write fluent answers, but training alone cannot show where a fact came from.

In one sentence: an LLM citation is a visible, clickable reference in an AI-generated answer that credits a specific web page as a source for part of that answer. It usually appears as an inline marker, a numbered footnote or a source card, and selecting it opens the page. A web citation normally appears only when the assistant has looked something up live, which is why some answers carry sources and others do not.

What an LLM citation looks like in each tool

Each product shows sources a little differently, and the designs keep changing. As of October 2026:

• ChatGPT. Responses that use web search may include citations; select one to open its source, and a Sources button below the answer lists the cited pages and other relevant links. ChatGPT search launched on 31 October 2024 and has been open to everyone in supported regions since 5 February 2025.
• Perplexity. It searches the web in real time and says every answer includes numbered citations linking to the original sources.
• Claude. Anthropic says that when Claude brings web information into a response it provides direct citations so you can fact-check the sources. Web search has been available on all Claude plans worldwide since 27 May 2025.
• Google AI Overviews and AI Mode. Both show links to supporting websites beside the AI-written text, and Google has since moved inline links higher in AI Mode and AI Overviews.

Citations, mentions and rankings are not the same thing

Three words get mixed up in conversations about AI search, and they describe different outcomes.

• A citation is a link to your page inside the answer.
• A mention is your brand name in the answer text, with or without a link.
• A ranking is a position in a list of links on a classic results page.

Citations and mentions often arrive together, and the split between them is an industry convention rather than a platform definition. The comparison with ranking matters more. Microsoft describes assistants such as Copilot breaking pages into smaller pieces, judging each for authority and relevance, and assembling an answer from pieces that often come from several sources; in Microsoft's words, visibility "is no longer defined by ranking position alone". A citation, then, says less about your page as a whole and more about one passage on it.

How AI assistants find the pages they cite

An assistant has two kinds of knowledge: what the model absorbed during training, with no links attached, and what it retrieves live by searching and reading pages while it writes. Citations come from the second kind.

The technique that joins the two is called retrieval-augmented generation (RAG), described in a 2020 research paper by Patrick Lewis and colleagues. Its authors noted that showing the provenance of a model's output was an open problem; retrieval is what makes a citation possible, because the answer can point back to the document it pulled in.

• ChatGPT may search automatically when a question would benefit from current information, or you can switch Search on yourself. OpenAI says that when ChatGPT uses outside search providers it typically rewrites your question into one or more targeted queries, reviews the first results and may then send narrower follow-ups; its own example turns a question about cancer drugs that target CCR8 into "CCR8 immunotherapy drug development 2025". Results come from third-party search providers, including Microsoft, plus content supplied directly by partners.
• Google says AI Overviews and AI Mode may use what it calls query fan-out, issuing multiple related searches across subtopics and data sources, and that its models identify more supporting pages while the response is being written, so these features can show a wider set of links than a classic search.

The practical point: your page competes to answer one of the smaller, more specific searches the assistant runs, rather than the broad question the person typed.

Why the sources change from one answer to the next

Ask the same question twice and you may get different sources. That is normal, and the reasons are documented.

• The question is rewritten, and phrasing, location and memory shape the rewrite. OpenAI says ChatGPT may use a general location from your IP address, turning "good restaurants near me" into "top restaurants San Francisco", and with memory on it may fold in saved facts such as a dietary preference.
• Different features use different models. Google says its two AI features do not necessarily run on the same model or method, so AI Mode and AI Overviews can give different answers with different links, and AI Overviews appear only when Google's systems judge them additive, so often they do not appear at all.

A single screenshot therefore proves very little; track a set of questions over time.

Why LLM citations matter for a business

A citation is a route back to your website from an answer that might otherwise have ended the search. OpenAI framed ChatGPT search at launch as giving content owners "new opportunities to reach a broader audience".

The clicks that do come through look worthwhile: Google says clicks from a results page with an AI Overview are higher quality, meaning visitors are more likely to spend more time on the site, and it suggests judging those visits by conversions rather than by clicks alone.

There is a quieter benefit too: when an assistant cites you, your business name sits beside the answer at the moment someone is researching a purchase, whether or not they click.

Are LLM citations always accurate?

No, and the companies say so. OpenAI's help page states that search results and citations "can be incomplete, outdated, or incorrect", and tells users to open the cited source and check that it supports the answer.

An independent test backs that up. In early 2025 the Tow Center for Digital Journalism gave eight AI search tools excerpts from 200 news articles and asked each to identify the article, publisher, date and URL, 1,600 queries in all. Collectively the tools answered more than 60 percent of queries incorrectly; Perplexity got 37 percent wrong and Grok 3, then in beta, got 94 percent wrong. The tools sometimes credited syndicated copies instead of the original publisher, and more than half of the responses from Gemini and Grok 3 pointed at fabricated or broken URLs. Treat the percentages as a snapshot of news attribution in early 2025; the types of failure are what matter.

For a business the lesson is practical: your facts can be paraphrased wrongly, and your content can be credited to a copy on someone else's site. Publish clear, unambiguous facts on your own website, keep them current, and make the original easy for crawlers to find.

What makes a page more likely to be cited

No platform publishes a formula; Microsoft says outright there is no secret strategy for being selected. What they do publish is consistent enough to act on.

Google: be indexed, be snippet-eligible, be useful. A page must be indexed and eligible to appear in Google Search with a snippet to be a supporting link in AI Overviews or AI Mode, and the site must not be excluded through Search Console's Search generative AI control. Beyond that, Google says no specific optimisation is needed; it points to unique, non-commodity content, an unblocked Googlebot, a normal HTTP 200 status and important content in text form.

Microsoft: fresh, authoritative, structured and semantically clear. Its guidance is the most specific about structure. It recommends:

• A page title, description and H1 that agree with each other and say plainly what the page delivers.
• Descriptive headings that mark where one idea ends and the next begins, so a section can be lifted on its own.
• Direct question-and-answer pairs with concise, self-contained answers, which Microsoft says assistants can often lift word for word, and lists or tables for steps and comparisons.
• Specific, measurable claims: Microsoft's example is that "42 dB dishwasher designed for open-concept kitchens" should replace a plain "quiet dishwasher".

Microsoft also lists what hurts: long walls of text, answers hidden in tabs or expandable menus that may not be rendered, core information kept only in PDFs, and key facts that exist only in images.

Structured data helps machines read what is already on the page. Microsoft describes schema markup as code that labels content as a product, review, FAQ or event, and Google's rule is that it must match the visible text. Neither company presents markup as a shortcut to being cited.

Put together, the page most likely to be cited reads like a good answer sheet: a clear question in the heading, a direct answer underneath, specific numbers, and nothing important hidden. Doing that across a whole site, tuned to how ChatGPT retrieves and ranks pages, is what our page on SEO for ChatGPT in Australia sets out.

Let the right bots in: robots.txt and AI crawlers

A crawler that cannot reach your page cannot cite it, so this is the first technical check. robots.txt is the text file at the root of your site that tells crawlers which paths they may fetch. The standard behind it, RFC 9309, says its rules are requests that crawlers are asked to honour and are not a form of access authorisation.

Each company runs more than one bot, and the one that matters for citations is usually not the one people block.

• OpenAI. OAI-SearchBot surfaces websites in ChatGPT search; OpenAI says a site that blocks it is left out of ChatGPT search answers, although a plain navigational link to that site can still show up. GPTBot is separate and crawls content that may be used for training, and the two settings are independent, so you can allow OAI-SearchBot and disallow GPTBot in the same file. OpenAI recommends allowing OAI-SearchBot and its published IP ranges, and says it can take about 24 hours after a robots.txt change for its search systems to adjust.
• Anthropic. ClaudeBot collects content that may contribute to model training; Claude-User fetches pages when someone asks Claude a question; Claude-SearchBot indexes content to improve search quality. Anthropic says disabling either of the last two may reduce your visibility in Claude's search answers.
• Google. AI Overviews and AI Mode are part of Search, so the robots.txt rules for Googlebot are the main control, along with the nosnippet, data-nosnippet, max-snippet and noindex settings; more restrictive settings limit how your content is featured. Search Console also has a Search generative AI setting (Settings > Search generative AI), on by default, that can exclude a whole site from AI Overviews and AI Mode without affecting normal results. Google-Extended is a separate token for Gemini training and grounding; Google says it does not affect inclusion in Google Search and is not a ranking signal, so blocking it does not take you out of AI Overviews.

Two things trip people up. A CDN, firewall or host can block these bots regardless of robots.txt, and both Google and OpenAI tell site owners to check that layer. And if ChatGPT never cites you, a blocked OAI-SearchBot is the first thing to rule out, which is also where ChatGPT SEO for US businesses starts. Allow time after any change: Google says recrawling can take several days to several months.

How to check whether you are being cited

No single report lists every AI citation of your site, so measurement combines four sources.

1. Search Console, for Google. The Generative AI performance report shows how often links to your pages appeared in AI Overviews and AI Mode, by page, country, device and date. It shows impressions only; clicks from those features are still counted inside the Web search type of the main Performance report and are not separated out.
2. Bing Webmaster Tools, for Copilot. Its AI Performance report shows how often your pages are cited as sources in Copilot and Bing's AI summaries, which URLs are cited and a sample of the queries behind them. It does not show clicks.
3. Referral traffic, for the rest. In GA4, visits from recognised assistants such as ChatGPT, Gemini and Claude are grouped automatically in the AI Assistant channel (added in May 2026); check it, and add any AI domains it misses with a custom channel group. Check the setup before you trust a trend: one ChatGPT referral drop in GA4 in October 2026 appears to have been mainly a missing tracking tag, which sent those visits to Direct, rather than lost visitors.
4. A prompt tracker. Write down 20 to 30 questions your buyers ask before choosing a business like yours, run them in ChatGPT, Perplexity, Claude, Copilot and Google's AI Mode on a fixed schedule, and record which pages are cited, including competitors'.

For a ready-made worksheet for that fourth step, the GEO Playbook includes a prompt tracker and a 90-day plan and explains how generative engines choose their sources.

Common mistakes and myths

• Myth: you need a special AI file or markup. Google says you do not need new machine-readable files, AI text files such as llms.txt, or special markup to appear in its AI features, because Google Search does not use them, and none of the other platform documentation used for this guide describes a special AI-only file as a route to citations.
• Myth: you can pay to be cited. OpenAI says ChatGPT ranks search results on multiple factors and that placement is not guaranteed; ads in ChatGPT are a separate product and an ad is not a citation.
• Myth: a citation is a backlink. A backlink sits on another website; a citation sits inside an AI answer and is chosen fresh each time the question is asked.
• Mistake: blocking every AI bot by default. GPTBot is for training; OAI-SearchBot decides whether you appear in ChatGPT search answers. Google-Extended governs Gemini training and grounding, and Google says it does not affect Google Search. Decide separately about training crawlers and search crawlers.
• Mistake: hiding the answer. Key facts in tabs, PDFs or images are exactly what Microsoft lists as hurting AI visibility, and Google has explained how Search indexes and ranks PDF files and why most business documents do better as web pages.
• Mistake: judging success by one prompt on one day. Sources vary with phrasing, location, memory and model, so track a set of prompts over time.

How to get started this month

1. Check robots.txt and your CDN. Confirm Googlebot, OAI-SearchBot and Claude-SearchBot are allowed and that your firewall or CDN is not blocking them. Check that Search Console's Search generative AI setting is on Include. Decide separately about GPTBot and ClaudeBot.
2. Confirm your key pages are indexed in Search Console, then write 20 to 30 of your buyers' questions in full sentences.
3. Rewrite one page answer-first: a one- or two-sentence answer under a descriptive heading, specific numbers, a short FAQ, and nothing important left only in images or PDFs.
4. Run a baseline. Ask your questions in each assistant, record who is cited, and repeat monthly. Group the AI domains in your analytics and watch enquiries as well as clicks.

All of this is ordinary work: making a page clear for a person, done knowing a machine will read it one passage at a time. For the wider subject, our AI SEO guide covers the foundations and our guide to GEO explains how it fits alongside SEO and AEO. Start with the robots.txt check; it takes an hour and settles whether the search bots can see you at all.