How Perplexity Selects and Ranks Its Sources

Two streams fill the candidate set Perplexity ranks. The first is a proprietary index, pulled in-house and rebuilt on the Vespa serving engine across billions of web pages; Vespa chief executive Jon Bratseth documented that deployment on 15 April 2025. The second is third-party search. Perplexity chief executive Aravind Srinivas told Fast Company on 21 June 2024: "We don't just rely on our own web crawlers, we rely on third-party web crawlers as well."

Retrieval names the lookup that runs before a single sentence is written. PerplexityBot fetches the candidate pages. The model then drafts the answer and attaches a numbered marker to each grounded sentence, so every claim in the output points at a source a reader opens. Ranking inside that candidate set rewards one stated verifiable fact over an assertion of expertise carrying none.

Citation density is the count of distinct sources attached to one generated answer. Qwairy measured it across 118,101 AI-generated answers carrying 669,065 citations during the third quarter of 2025, and Perplexity placed first.

The arithmetic decides the rest. An answer holding 21 slots admits 21 domains; an answer holding 8 admits 8. A domain outside the leading three results still lands inside a 21-slot list.

Counts move with methodology. Qwairy counted every citation attached to an answer, including the extended source list behind the visible summary.

Citations per answer, five providers measured

Perplexity cites more sources per answer than ChatGPT or Google AI Overviews. Across the five providers Qwairy reported — Perplexity, Google AI Overviews, Gemini, ChatGPT and Microsoft Copilot — Perplexity ranked first at 21.87 citations per question, ChatGPT fourth at 7.92, a ratio of 2.76 to 1.

PerplexityBot is Perplexity's crawler. Perplexity's documentation, checked 21 September 2026, states two facts: its purpose is surfacing and linking websites in Perplexity search results, and it is not used to crawl content for AI foundation models. Blocking it drops the domain from Perplexity's index and buys no training protection.

A second token behaves differently. Perplexity-User fetches a page when a user's question requires it, and Perplexity documents that this fetcher ignores robots.txt rules because a person requested it. Two tokens, two behaviours: PerplexityBot indexes, Perplexity-User fetches live.

Three access conditions

1. Permission

   robots.txt admits the PerplexityBot user agent, and the firewall admits the IP ranges at perplexity.com/perplexitybot.json.

2. Rendering

   the answer text sits in the first HTML response. Vercel and MERJ logged 24.4 million PerplexityBot fetches in one month, recording no JavaScript execution, 17 December 2024.

3. Status

   the URL returns 200, with no login wall, consent gate or interstitial.

An attributable claim is a statement a reader traces to a named quantity and a named basis without leaving the sentence. Perplexity ranks candidates inside an already-relevant set, so attributability breaks the tie. The numbered marker has to point at text that settles the sentence preceding it.

Three properties of an attributable claim

1. Quantification

   the claim names a figure, a date or a named entity, never a vague quantity.

2. Provenance

   the claim names its measurement basis or its source inside the same sentence.

3. Independence

   the claim resolves in one or two sentences, with no pronoun reaching back into the paragraph above it.

Two sentences show the gap. "We are a leading AI visibility consultancy" carries no figure and no basis. "Perplexity averaged 21.87 citations per question across 118,101 answers measured by Qwairy in Q3 2025" carries both, and it survives extraction whole.

Attribution is measured rather than assumed. The citation-rate method states how a share is computed across a fixed prompt set, and the weekly monitoring run records which assistant named which source on which date.