LLM Citations

LLM citations are the URLs that an AI assistant or answer engine attaches to a generated response to show where its information came from, displayed as inline numbers, footnotes, link chips or source cards. They are the most measurable output of GEO work, because a citation is a concrete link to a specific page on a specific domain.

Citations depend on retrieval. When a model answers from parametric knowledge learned in training, it usually cites nothing, because it has no record of which documents taught it a fact. When it runs a search or reads documents at answer time through retrieval-augmented generation, it can point to the pages it read. In practice, the live-search products (ChatGPT search, Google AI Overviews and AI Mode, Microsoft Copilot, Gemini, Perplexity and Grok) return citations on most informational answers, while API models without a search tool return few or none. Promptwatch publishes daily averages of sources per response by platform, and those numbers move as products change.

A citation is not the same as a mention. A mention means the answer names your brand in its text; a citation means the answer used your URL as a source. The two are recorded independently and respond to different work. A how-to article can be cited in an answer that recommends a competitor, and a well-known brand can be recommended without any of its pages being cited. Teams track both, along with metrics such as citation share, cited URL rate and citation rank (where in the source list a URL appears).

Citation accuracy is uneven. In a March 2025 study, the Tow Center for Digital Journalism gave eight AI search tools excerpts from news articles and asked them to identify the article, publisher and URL. Collectively the tools answered more than 60% of queries incorrectly, ranging from 37% for Perplexity to 94% for Grok 3, and several cited syndicated copies or broken URLs instead of the original. Citations can also shift abruptly when a platform changes its retrieval: Promptwatch data showed Reddit's share of ChatGPT search citations falling 86% in a single week in August 2026, as covered in our analysis of why ChatGPT stopped citing Reddit.

What earns citations follows from how retrieval works. The page has to be reachable by the platform's search crawler (OAI-SearchBot for ChatGPT, Bingbot for Copilot, Googlebot for AI Overviews) and rendered in HTML rather than client-side JavaScript. It has to rank or be retrieved for the sub-queries the model generates during query fan-out. And it has to state a specific, quotable answer, ideally with data or first-hand detail the model cannot find elsewhere. Third-party pages matter as much as owned ones: in many categories, review sites, forums and publications collect more citations than brand domains, so a citation source audit of who gets cited for your prompts is a sensible starting point.

Measurement comes from two directions. First-party reports are starting to appear, such as Bing Webmaster Tools' AI Performance report for Copilot and Google Search Console's generative AI performance report. They show your own citations but not competitors'. Prompt tracking on the live interfaces fills that gap by recording every cited URL for a fixed prompt set over time, so teams can see which pages win citations, which competitors take them, and how that changes after content updates or platform changes.