AI citation study: is GitHub cited by AI?

GitHub is a primary source in the strict sense: its pages are written by the people who build the software. A README explains what a project does, a release page says what changed in each version, and the issue tracker shows which bugs are known and whether they are fixed. AI models treat these pages as evidence when they describe a tool.

Four kinds of GitHub pages tend to surface in answers:

• Repository home pages, whose README works as the official description of an open-source project.
• Releases and changelogs, used for questions such as "what is the latest version" or "when was this feature added".
• Issues and discussions, quoted when someone asks about a bug, a limitation or a workaround.
• Security advisories from the GitHub Advisory Database, which documents known vulnerabilities in open-source packages.

Organisation profiles matter as well. A company that keeps its SDKs, plugins or integrations in a public GitHub organisation gives AI models a verifiable trace that its product exists and is maintained.

Citations cluster around technical intents. Typical prompts include "open-source alternative to Jira", "how to install the Stripe CLI", "is this npm package still maintained", "best self-hosted password manager" or "how to patch the Log4j vulnerability". In each case the repository is the most direct proof an assistant can offer.

Questions about everyday businesses rarely reach GitHub. Nobody needs a repository to find "a reliable plumber in Lyon" or "the best accounting firm for a startup". As a result, GitHub citations concentrate on questions about software publishers, cybersecurity vendors and IT service companies, and stay marginal in other sectors.

Language plays a part too. Documentation on GitHub is overwhelmingly written in English, so English answers reach it far more naturally than French, German or Spanish ones, which tend to cite translated tutorials or local media instead.

Most of the gap comes from how each model retrieves information. Gemini grounds many answers in Google Search, where repositories rank well for project names, so a GitHub link can come along with the answer. ChatGPT combines what it learned in training with live results from ChatGPT search, and attaches a repository link mostly when the question names a specific project.

Perplexity searches the web for every question and lists its sources, yet for comparison queries it frequently relies on review articles and vendor sites rather than raw repositories. Grok adds posts from X, where developers announce releases and discuss bugs. Claude cites sources less often by default and tends to describe a project in its own words unless web search is turned on.

Part of GitHub's influence is invisible in citation counts. Public code and documentation hosted on GitHub feed the training data of many large models, so an assistant may explain how a library works without naming the page it learned from.

A citation here means that the answer gave a page on github.com as a source: a repository, an issue, an advisory or the GitHub documentation. Advice such as "host your code on GitHub" without a link is a recommendation of the service, a different signal that this rate does not capture.

Sites built with GitHub Pages live on github.io or on custom domains, so they are best read as websites in their own right. For a brand, the valuable case is one of its own repositories showing up in an answer about its category: a technical reference that an assistant can check, something marketing pages rarely achieve.