GEO Guides
How ChatGPT, Perplexity, and other AI engines actually decide what to retrieve, trust, and cite in an answer.
Getting cited by an AI engine isn't a black box you can only guess at. There's a specific mechanism behind it, and understanding that mechanism tells you exactly what to do differently in how you write. This chapter explains how generative engines decide what to pull into an answer and, just as importantly, how that process differs from one engine to the next.
Most AI answer engines are built on a technique called retrieval-augmented generation, or RAG. Instead of answering purely from what a model memorized during training, the system first retrieves a set of relevant documents or passages from an index (or a live web search), then generates its answer grounded in what it retrieved, typically attaching citations back to those sources.
That two-step process matters because it means getting cited is really two separate problems: first, does your content get retrieved at all, and second, once retrieved, does the model choose to quote or link it rather than a competing source. Both steps reward different things than classic search ranking does.
Retrieval systems aren't scanning for keyword density or backlink counts. Research on how these systems select sources points to a few consistent properties:
That last point is easy to underestimate. According to analysis from Yext and Profound on how ChatGPT, Perplexity, Gemini, and Claude decide what to cite, roughly 85% of brand mentions in AI answers come from third-party pages, not the brand's own website. In other words, what other credible sites say about you carries more weight in an AI's answer than what you say about yourself.
One of the most useful, and most overlooked, facts about GEO is that different engines have meaningfully different citation behavior. That same Yext and Profound research found that ChatGPT leans heavily on its training data and high-authority reference sources, with Wikipedia alone accounting for roughly 48% of its top citations, and only searches the live web when it judges that necessary. Perplexity, by contrast, searches the live web on essentially every query and tends to cite around 16 sources per answer.
The overlap between what the two engines actually cite is small: only about 11% of cited domains show up in both ChatGPT's and Perplexity's citations, according to the same analysis. That's a meaningful number for anyone tempted to optimize for a single engine.
Note: The practical implication of that 11% overlap figure is that there is no single trick that gets you cited everywhere. Winning one engine's citation pattern doesn't transfer automatically to the next one.
Because engines pull from different mixes of sources, and weigh first-party versus third-party content differently, the only strategy that reliably works across all of them is being consistently and credibly present: accurate information on your own site, plus a genuine trail of third-party mentions, reviews, and references that corroborate it. Trying to reverse-engineer one engine's current citation quirks is a short-lived tactic; broad, real credibility is not.
The foundational research behind the term GEO, a study led by researchers at Princeton and published at ACM KDD 2024, tested a range of content-optimization techniques directly against live generative engines and measured the effect on visibility. Its headline finding was that the right optimizations lifted a page's visibility in AI-generated answers by up to 40%, and the single highest-impact techniques it identified were adding relevant statistics, citing credible sources, and including direct quotations, in short, making content verifiable rather than merely well-written.
That finding lines up with the retrieval behavior described above: a system built to ground its answers in retrieved evidence naturally prefers passages that already look like evidence.
Tip: When you're evaluating whether a page is likely to get cited, ask a blunt question: if a fact-checker read only this paragraph, could they verify the claim without leaving the page? If not, that's usually the gap to close first.
HarperFlow writes every article with this retrieval behavior in mind: statements are backed by named sources where a claim needs one, passages are structured so a single section can stand on its own, and the tone stays factual rather than promotional, because that is exactly what a retrieval-and-citation system is built to prefer.