On Perplexity, 64% of retrieved pages are never cited. Retrieval and citation measure different things.

On Perplexity, 64 percent of the pages it retrieves for a query are never cited in the final answer, according to a February 2026 benchmark study of more than a million AI citations from Peec AI. That number surprises people who treat retrieval and citation as the same event. They aren't. A page can clear the first filter, get pulled into an AI engine's candidate set, and still lose at the next stage, where the model decides which of those candidates actually earn a place in what the user reads. Three separate 2026 studies now map that funnel in detail, and the gaps between its stages are wider than most GEO advice admits.

What retrievability actually measures

Retrievability is the likelihood that a piece of content surfaces at all during an AI system's initial document search, before any citation or inclusion decision happens. A page can be well-written and still have low retrievability, because this stage runs on relevance and ranking signals, not on prose quality. Getting found comes first, and it's a different problem from getting chosen, a distinction most introductory GEO guidance skips past.

Most AI search systems don't retrieve whole pages anyway. They run passage retrieval, pulling specific chunks of text rather than entire documents, so a page can be retrieved for one paragraph while the rest of it never enters the system's consideration for that query. An 800,000-page citation study covered on this blog found that retrieval rank and topical relevance predicted citations far better than how many query variations a page happened to cover, a result that cuts against a lot of advice built around chasing query coverage instead of ranking well for the handful of queries that matter.

Why most retrieved pages never get cited

Retrieval only builds a candidate pool. What happens next is a separate filter, and most candidates don't survive it. On Perplexity, Peec AI's February 2026 benchmark found 64 percent of retrieved URLs get dropped before the answer is finalized.

Being retrieved by an AI search engine is necessary, but not sufficient, for being cited in its answer.

A 2026 measurement framework built by independent researchers Zhang Kai, He Xinyue, and Yao Jingang, analyzing 602 prompts across ChatGPT, Google's AI Overviews, and Perplexity, found the same drop-off from a different angle. Search triggers on 98.6 percent of ChatGPT prompts, 99.7 percent of Google prompts, and 100 percent of Perplexity prompts, so nearly every query produces a candidate pool. What shrinks is how much of that pool makes the cut: ChatGPT surfaces an average of 6.9 citations per answer, Google's AI Overviews 12.1, and Perplexity 16.4, each drawn from a far larger set of candidates. The distance between how many sources a system actually consults and how many it shows the reader is what this site's glossary defines as the attribution gap, and it's decided well before anyone gets to judging writing quality.

Getting cited doesn't guarantee the page shapes the answer

Even a source that clears retrieval and gets cited doesn't necessarily shape what the AI actually says. Zhang, He, and Yao's framework measures a third stage they call citation absorption: how much language, evidence, or structure a cited page actually contributes to the final answer, separate from whether it was linked at all.

Their data, built on the public geo-citation-lab dataset, found pages containing numbers or statistics averaged 61.6 percent higher absorption scores than pages without them, a larger gap than any other content difference they measured. Encyclopedia-style pages scored the highest average influence of any content type in the dataset; general news pages ranked among the lowest. That lines up with what this blog found digging into the original 2024 GEO paper, where quotations and statistics were the two rewrites that moved citation-worthy visibility the most, for the practical reason that a specific number or a named claim gives a model something concrete to lift without inventing or paraphrasing a fact it could get wrong. The glossary term for a passage's ability to be lifted cleanly on its own is extractability, and content that already reads like a quotable fact tends to score well on it by default.

The same page can clear this funnel on one engine and fail on the next

Clearing retrieval, selection, and absorption on one AI engine says almost nothing about performance on another. A September 2026 study by independent researcher Benjamin Tannenbaum compared ChatGPT, Microsoft Copilot, Google, and Perplexity on identical prompts and found that 84.9 percent of engine pairs shared zero cited URLs for the same query.

Tannenbaum's dataset covered 15 fixed commercial prompts and 589 citation observations across 528 unique URLs. Across any two engines, the overlap in exactly which pages got cited for the same query averaged a Jaccard similarity of 0.0079, against roughly 0.127 expected if citations were handed out at random. Only 3.6 percent of observed URLs showed up on more than one engine, and no single engine captured more than 42.6 percent of the full four-engine citation set. A separate 126-million-prompt study this blog covered previously found the same instability one stage earlier in the funnel: the gap between a brand getting mentioned by an AI engine and that engine actually citing it as a source ranged from 30 percent on Gemini to 64 percent on Google's AI Overviews. Retrieval, selection, and absorption apparently don't transfer across engines any more reliably than mentions and citations do within a single one.

What this means for measuring and improving AI visibility

Tracking one citation number tells a brand almost nothing about where it's actually failing. A brand that never gets retrieved has a different problem than one that gets retrieved constantly but rarely selected, which is a different problem again from one that gets cited but contributes nothing to what the answer actually says.

Diagnosing which stage is failing changes what to fix. Low retrievability points toward relevance and ranking signals, the same levers behind ordinary search visibility. Getting retrieved but rarely selected points toward how the content competes against other candidates for the same query, which is what the citation rate metric is built to track. Getting selected but barely absorbed points toward the writing itself: whether a paragraph carries a concrete number or claim a model can lift on its own, the same discipline behind the answer-first writing framework this blog has covered at the paragraph level before.

None of the three studies behind these numbers claim the funnel behaves identically on every engine or holds steady as models change. The 2024 paper that first measured statistics and quotations is already two model generations removed from the systems people use today, and Tannenbaum's own same-engine comparison, run a day apart, found 67 percent of one engine's cited URLs turned over in that single day. Every figure in this piece is a snapshot of how 2026's systems behaved when someone measured them, not a fixed conversion rate.