How Perplexity ranks and cites pages (2026)

Perplexity is the one major answer engine that has published how its search pipeline works. Its September 2025 engineering post describes the index, the unit it retrieves and the ranking stages, so some advice here is documented by Perplexity rather than inferred. Numbers from outside studies name their sample, their date and how strong the evidence is. Engines change monthly, so treat every number as a dated snapshot.

How Perplexity's search index works

Perplexity searches its own web index of more than 200 billion URLs, built on the Vespa search engine. Perplexity's post on its search API (Sep 2025) says machine learning decides which pages to index and how often to refresh them, and that the fastest storage tier favors “authoritative domains” and “undercovered topics.” Pages are parsed with rulesets tuned for language models. All of this is confirmed by Perplexity.

Perplexity's index is its own, so ranking in Google or Bing is not the whole story. Perplexity relied on the Bing API in early versions, and it has not denied using third-party feeds today. PerplexityBot, its crawler, builds the index.

Perplexity also rewrites questions into searches, though less than some engines. Peec's analysis of 5 million fan-out queries found about 1.4 searches per Perplexity prompt, compared with 2.1 for ChatGPT and 6.8 for Grok (plausible, one vendor).

How Perplexity ranks pages: spans, prefilter and cross-encoder rerank

Perplexity ranks pieces of pages, not whole pages, in four stages. Its engineering post says each document is split into “self-contained spans” that are “individually retrieved and ranked,” the most atomic units possible.

Stage	What Perplexity says it does	What it means for a page
1. Span splitting	Breaks each document into self-contained sub-document spans	Each section is judged on its own, without the rest of the page
2. Hybrid retrieval	Combines lexical (keyword) and embedding (meaning) search	Use the words people search with, and state each idea plainly
3. Prefilter	Removes “non-responsive or stale content”	Sections that answer nothing, or look out of date, drop out early
4. Fast scorers, then cross-encoder rerank	Cheap scorers narrow the set, then a cross-encoder compares each span with the query	The final pick depends on how directly the span answers the query

The mechanism is confirmed by Perplexity. Any uplift from writing for it is plausible, not measured. Perplexity does not document how it orders citations, so citation order should not be read as rank.

Why self-contained sections matter on Perplexity

Self-contained sections matter on Perplexity because its retriever ranks each span separately, so a section that needs the rest of the page to make sense competes poorly. Span-level writing is a design target Perplexity documents itself, not a theory from the GEO industry. A span that opens with “It also supports this” gives the reranker nothing to match.

Writing for span-level retrieval looks like this:

Open each section with a sentence that answers its heading directly, the answer-first pattern.
Name the subject again instead of starting with “it” or “this”.
Keep each number next to its unit, date and source.
Define a term where it is used, not only in an earlier section.
Keep one idea per section, under a heading that states it.

Formatting is a smaller lever than content. CITECHOICE, a causal replay study (2026), found that structured rendering added about 0.50 citations per answer for a page that was already retrieved, but did not make retrieval itself more likely. The wider evidence on what does and doesn't move citations is in the guide to what moves AI citations.

PerplexityBot vs Perplexity-User: crawler access

Perplexity uses two crawlers. PerplexityBot builds the search index and respects robots.txt, while Perplexity-User fetches pages live when a user's question needs them and generally does not follow robots.txt. Both points are confirmed by Perplexity's documentation, as reported.

To be in Perplexity's index, PerplexityBot must be allowed in robots.txt and must not be challenged by a firewall or CDN bot rule.
A robots.txt rule generally won't stop Perplexity-User. A server or firewall rule would, at the cost of those live fetches.
AI crawlers generally don't run JavaScript, according to crawler data from a hosting platform and repeated fetch tests, so the main text belongs in the server HTML.

Robots.txt shows intent, not behavior. Server logs show whether PerplexityBot actually reaches the pages that matter. optmz's site audit checks robots.txt rules for AI crawlers and whether content is in the server HTML, and its Real AI data feature verifies AI crawler fetches from server logs. The AI crawler user-agent list has the exact tokens.

How freshness affects Perplexity citations

Freshness affects Perplexity directly, because its ranking pipeline includes a prefilter that removes stale content before the final rerank. Perplexity's Search API also returns a date and a last-updated field for each result, which shows the system tracks page dates. Both are confirmed by Perplexity's documentation.

Outside studies agree on the direction. Ahrefs' freshness study (17 million URLs) found AI assistants cite pages 25.7% newer than Google's organic results. A lab study of LLM reranking (Fang et al., SIGIR-AP 2025) found that a recent date alone could move a passage up by as many as 95 ranks. Freshness is a strong factor for Perplexity, though no study isolates how Perplexity itself weighs dates.

Perplexity's citations were also the steadiest of the engines Profound tracked (Jul 2025, about 80,000 prompts per platform): 40.5% of cited domains changed after one month, against 54.1% for ChatGPT and 59.3% for AI Overviews. Keep pages current by updating their facts, show a visible updated date that matches, and never change a date without changing the content.

Does Google rank predict Perplexity citations? About 29% overlap

Google rank predicts Perplexity citations only partly. Ahrefs (Jul 2025, 15,000 prompts) found about 29% of Perplexity's citations overlapped Google's top 10, higher than ChatGPT's 8% in the same study but far from a match. CiteLens (Jul 2026) reported 89%, and the gap between the two studies is unresolved.

Perplexity runs its own index, so the safer reading is that Google rank is a partial signal. Being indexed by PerplexityBot, ranking somewhere for the searches Perplexity runs, and having a span that answers the query directly all count for more than a top-3 Google position for the head term.

Reddit, YouTube and the sources Perplexity cites most

Reddit and YouTube are the most-cited domains on Perplexity in the large studies, but their shares swing widely from month to month. Profound's analysis of 680 million citations (to June 2025) put Reddit first at 6.6% of Perplexity citations and YouTube second at 2.0%. Tinuiti and Profound later measured Reddit at 24% of Perplexity citations in January 2026. Perplexity also over-indexes on academic sources and documentation. All of this is plausible: several datasets, measured differently, unstable over time.

The swings are too large for fixed rules. On ChatGPT, Ahrefs saw Reddit's share fall from 19.1% to 2.3% between March and September 2026, and Profound's 680-million-citation shares predate September 2025, since when source shares in such datasets have swung by as much as 10 times. A rule like “post on Reddit to get cited” has no stable evidence behind it. Check which domains are actually cited for the questions a client cares about, week by week.

Perplexity API vs the Perplexity app

The Perplexity API and the Perplexity app share a search system but use different settings, so API results are a close relative of the app, not a copy. Perplexity's FAQ says the API uses the same search system “with differences in configuration” and that the model may differ. Perplexity staff on its forum have said the app's Pro Search is multi-step, while the sonar-pro API model was single-step. That evidence is strong.

Users report about 2–10 citations per API answer, compared with 10–40 or more in the app.
Surfer and cloro (Oct 2026, 1,000 prompts, one run each) measured 19.7% URL overlap between Perplexity's API and app answers (plausible).
Perplexity retired its Sonar chat-completions API on September 27, 2026 in favor of an Agent API, according to a third-party integration notice. Perplexity staff point to the pro-search setting for answers closer to the app.
The Search API returns a title, URL, snippet, date and last-updated time for each result, and accepts country and domain filters.

Perplexity API data is best treated as a labelled proxy for what app users see. The glossary entry on API vs app explains the difference. Results agree better at the domain level than at the URL level, so compare the two by site.

What doesn't help on Perplexity: schema, llms.txt and bolted-on stats

JSON-LD schema, llms.txt files and inserted statistics or quotations have no good evidence of helping Perplexity citations. Perplexity's ranking stages work on the visible text of each span, and nothing in its engineering post gives markup a role in retrieval or reranking.

JSON-LD: live fetch tests by searchVIU and Pilot Digital found that ChatGPT, Claude and Perplexity strip script-based JSON-LD when they read a page. Schema still matters for Google rich results.
llms.txt: Ahrefs found that 97% of the files on 137,000 sites were never fetched, and no study shows a citation effect.
Added statistics, quotes or “according to” lines: a 2026 re-run of the original GEO tactics on 10 modern engine families found effects indistinguishable from zero, or negative.
Keyword stuffing: measured as harmful in the original GEO benchmark, and still null to negative in 2026.

Real specifics are a different matter from bolted-on ones. A controlled study of 252,000 paired trials (“What Gets Cited”, 2026, read at abstract level) found that explicit prices and recent timestamps gave consistent gains, while formatting-only edits did little. Numbers that answer the question belong in the span; numbers added for show do not help.

Checklist: how to get cited by Perplexity in 2026

A page gets cited by Perplexity when PerplexityBot can index it, its sections answer questions on their own, and its facts are current. Work through these steps in order.

Allow PerplexityBot in robots.txt and in firewall and CDN bot rules.
Put the main text in the server HTML.
Check server logs for PerplexityBot fetches of the pages that matter.
Give every section a heading that states its topic and a first sentence that answers it.
Make each section readable alone: repeat the subject, define terms, keep numbers with their units and dates.
Cut filler that answers nothing, which the prefilter is designed to drop.
Update time-sensitive pages when their facts change, with a visible date that matches.
Treat Google rank as a partial signal and Reddit or YouTube shares as unstable.
Measure with repeated samples, and keep API and app results apart.

Sources