What Perplexity says about itself
Start with what Perplexity states officially, because most of what circulates online about its ranking is inference from outside researchers, not a published algorithm. Perplexity's own help center describes the product plainly: it searches the web in response to a question and returns "accessible, conversational answers backed by verifiable sources," with content "sourced from the web in real-time" and citations drawn from "reputable news organizations, academic publications, and established content sources," according to Perplexity's help center. That is a description of intent and category — real-time retrieval, credible sourcing — not a scoring formula. Perplexity has not published the exact weights it assigns to relevance, authority, freshness or any other signal, and every specific percentage you will read about its ranking, including the ones in this article, comes from outside researchers testing the product from the outside, not from Perplexity's own documentation.
Retrieval comes before ranking
The architecture itself is not a mystery: Perplexity is built on retrieval-augmented generation. A query triggers a live web search rather than a pull from a static, pre-trained knowledge base, which is also why Perplexity answers can reflect information from the same day. Independent analysis of the product's behavior, including a technical breakdown from ZipTie, describes this as a multi-stage pipeline: an initial retrieval step pulls a batch of candidate pages for the query, a reranking step scores and narrows that batch, and only then does a language model draft an answer that cites specific passages from the surviving pages. The reranking stage is the part site owners actually care about, because it decides which of the retrieved pages make it into the final answer.
What independent researchers report drives the ranking
Because Perplexity does not publish its ranking weights, most of what is known about ranking factors comes from third parties reverse-engineering behavior across many test queries. Third-party analysis of AI-search visibility reports that content relevance to the query, domain authority, content freshness, source diversity and the presence of structured data all factor into which pages get cited, with the relative importance shifting by query type — informational questions leaning more on relevance and freshness, commercial questions leaning more on trust signals. Other researchers frame the same idea in more practical terms: pages that state a fact plainly, in a form a model can lift as a self-contained answer, are cited more often than pages that bury the same fact in marketing copy. None of these weightings are official; they are patterns observed by outside researchers running repeated test queries against the live product, and they should be read as directional evidence rather than a documented spec.
How many sources actually make it into an answer
The number of sources behind a single Perplexity answer varies by study and by query complexity, and the range reported is wide enough that it is worth citing more than one figure. Analysis from Uprise Digital reports an average of roughly 18 sources cited per answer in its sample. A separate 2024 academic study found just over 5 sources per response on a different query set. ZipTie's breakdown of the pipeline describes Perplexity pulling a wider pool of 5 to 10 candidate pages through hybrid search before a multi-stage filtering process narrows that down to a smaller number of pages that actually get cited in the visible answer. The honest summary: Perplexity typically evaluates more pages than it cites, the cited count is small relative to the web, and no single number applies to every query — simple factual questions tend to draw fewer citations than open-ended or comparative ones.
What separates cited pages from ignored ones
- The page answers the question in its own words, early. A model pulling a citable passage favors a page where the answer sits in the first few sentences of a section, not one buried after three paragraphs of preamble.
- The page is fetchable and parseable. If a page blocks crawlers, requires JavaScript to render its core content, or returns a different response to a bot than to a browser, it cannot be retrieved in the first place — a retrieval failure looks identical to a ranking loss from the outside.
- The domain shows up consistently on the topic. Independent research on source diversity, including an arXiv audit of generative search engines, found that AI answer engines draw from a set of domains that recur across related queries rather than sampling the entire web evenly each time — an unfamiliar domain has a harder time breaking in on a competitive topic than an established one.
- The claim is specific and checkable. Vague marketing language does not lift cleanly into a cited sentence; a concrete number, date or named fact does.
What this means if you run a business site
You cannot see Perplexity's ranking weights, and neither can anyone else outside the company — so the practical response is not to chase a formula but to remove the failure modes that keep a page out of the retrieval pool at all. That means a page has to be fetchable by Perplexity's crawler, has to state its key facts plainly near the top of the relevant section, and has to exist on a domain that publishes on the topic more than once. Agent Ready Search's scan checks the fetchability and structure half of that equation directly — whether your page can be retrieved and parsed in the first place — because that is the part an AI-search-readiness check can measure directly, on the page itself. The authority and consistency half is a longer-term editorial commitment, not a technical fix.