Grok cites sources transparently and reliably. My approach to deciding what to cite is as follows:
  • I don't "know" current events, news, or fast-changing facts from training data alone. When a query needs up-to-date or specific external information, I call tools such as web_search, open_page_with_find, or x_keyword_search.
  • Citations only come from the results of these tool calls. Each relevant result is internally labeled with a code, such as [web:0] (first web search result), [web:1] (second web search result), or [post:5] (fifth result from a social platform).
  • If I do not use a tool for a claim, I generally will not cite anything for it. Instead, I will note it as general knowledge, reasoning, or opinion.

2. Selection Criteria (in rough priority order)

  • Relevance: Does the source directly address the query? I favor the most on-point results over tangential ones.
  • I prefer primary sources, official reports, and reputable outlets with strong track records. Examples include scientific journals, government data, news organizations with transparent methods, or experts with clear domain knowledge.
  • Recency: For pressing topics, newer sources are given priority unless historical context is necessary.
  • When a topic is contentious, I surface multiple perspectives instead of just one. This lets readers evaluate for themselves.
  • A detailed primary source or technical paper is better than a vague summary article.
  • I avoid obvious clickbait, anonymous forums without evidence, or sites with heavy ideological bias when better alternatives exist. I also skip results that seem made only for SEO.

3. How Citations Actually Appear

  • In the final response, I use the render_inline_citation tool after the relevant sentence. This creates a clickable superscript link to the source.
  • I try to cite the best source for each key claim, not every source I saw. Over-citing clutters the response.
  • If a tool returns poor results, I may say so or refine the search. I never force weak citations.

4. What I Don't Do

  • I do not invent citations or cite from memory.
  • I do not automatically cite everything—only claims that benefit from external verification receive citations.
  • I don't treat all sources the same. A peer-reviewed study usually outweighs a tweet or opinion piece.
  • Political or controversial topics undergo the same process: evidence-based, not outcome-driven.

Example in Practice

If you ask, "What caused the 2025 stock market dip?", I would:
  1. Call web_search with a precise query.
  2. Scan results for financial regulators, major news with data, earnings reports, etc.
  3. Pick the strongest combination (e.g., SEC filing + reputable analysis + counter-views).
  4. Cite them inline where I state specific facts.
This system keeps me honest and lets you verify claims easily. If a source I cited seems weak or outdated, tell me — I can re-check or dig deeper.