$ man geo/how-ai-engines-source-content

PATTERN

AI engines follow a three-stage pipeline to go from a user query to a cited answer. Stage one is retrieval - the engine needs to find relevant content. This happens through a combination of pre-trained knowledge (what the model learned during training), real-time web search (Perplexity and ChatGPT with browsing both do this), and indexed content stores. Stage two is evaluation - once candidate sources are retrieved, the engine assesses them for relevance, authority, freshness, and extractability. Not every retrieved source makes it into the final answer. Stage three is generation - the engine synthesizes information from the top sources, generates a natural language answer, and decides which sources to cite inline or in footnotes. Each stage has different optimization levers. Retrieval depends on your technical infrastructure - can the AI find your content? Evaluation depends on your content quality and authority signals. Generation depends on how extractable and quotable your content is.

Different AI engines use different retrieval methods, but they converge on similar patterns. Perplexity runs a real-time web search for every query using its own search index plus Bing results. ChatGPT with browsing uses Bing search. Google Gemini and AI Overviews use Google's own search index. Claude does not browse the web in standard mode but has training data and can use tools. The retrieval step means that if your content is not indexed by search engines, it is invisible to most AI engines. All your SEO fundamentals - sitemaps, crawlability, robots.txt configuration - directly impact whether AI engines can even find your content. Beyond traditional search indexing, some AI engines also use RSS feeds, API endpoints, and direct content partnerships. Perplexity has been particularly aggressive about indexing RSS content quickly. This means your RSS feed is not just for blog subscribers anymore - it is a discovery channel for AI engines.

PATTERN

Once an AI engine has a set of candidate sources, it needs to decide which ones to actually cite. The evaluation criteria vary by engine but cluster around four factors. Authority: is this source recognized as an expert on this topic? Signals include domain authority, backlink profile, author credentials, and whether the source is frequently cited by other trusted sources. Relevance: does this source directly address the query? Sources that answer the question head-on beat sources that mention the topic tangentially. Freshness: was this content recently published or updated? AI engines strongly prefer recent content, especially for topics that evolve. Extractability: can the AI cleanly pull a specific claim, statistic, or answer from this source? Content that buries its key points in long paragraphs loses to content with clear, structured answer blocks. The weight of each factor varies by query type. For factual queries, authority and extractability dominate. For comparison queries, freshness and relevance matter more.

The generation stage is where your content either gets cited or gets left out. AI engines have different citation styles. Perplexity uses numbered footnotes with source links - it is the most citation-heavy of the major engines. ChatGPT with browsing includes inline links and sometimes source cards at the bottom. Google AI Overviews show source cards with site name, favicon, and title. The common thread is that AI engines cite sources at the claim level, not the page level. If your page has ten facts and the AI uses one of them, it cites your page for that specific claim. This means every individual claim on your page is a potential citation opportunity. Pages with more extractable, specific claims have more surface area for citations. Generic overview pages with no specific claims rarely get cited even if they rank well in traditional search. The engine has nothing specific to attribute to them.

ANTI-PATTERN

Some publishers have blocked AI crawlers like GPTBot and PerplexityBot in their robots.txt, hoping to force AI engines to license their content. While this is a legitimate business decision for large media companies with negotiating leverage, for most businesses it is self-defeating. Blocking AI crawlers does not prevent AI engines from knowing about your content - they still have training data and can access cached versions. But it does prevent them from citing you with a fresh, linked reference. You go from being a cited source to being uncredited background knowledge. For B2B companies, startups, and content-driven businesses, AI citations are free, high-intent distribution. Blocking that is like blocking Googlebot in 2005 because you did not want people finding your site through search.