GEOhandbook

Articles ·

How Perplexity actually chooses its sources — and why your rank-12 page can still get cited

3 min read

The one-line version: Perplexity runs a live web search on every query, breaks that query into sub-queries, fetches roughly ten candidate pages in full, and picks citations with a three-layer machine-learning reranker that scores topical authority and semantic depth. Ranking on Google is neither necessary nor sufficient. A page at Google position 12 can still be cited if its passages are more directly answerable than the ones above it (Fokal).

Why Perplexity is different — the citation-volume gap

Perplexity cites drastically more sources per answer than ChatGPT. On multi-constraint queries, Machine Relations Research measured an average of 17.7 citations per Perplexity answer vs. 3.4 for ChatGPT. An independent analysis of 118,000 AI responses put Perplexity at 21.87 citations per answer across all query types. For a typical query Perplexity may cite 15–20 pages while ChatGPT cites 3–4.

Perplexity’s architecture is described as breadth-oriented: it decomposes complex queries into narrower sub-queries, pulls from many sources at lower individual absorption depth, and spreads citation weight widely. ChatGPT is the opposite — fewer sources, more language and structure taken from each (Machine Relations). For a publisher, that means Perplexity gives you many more chances at a citation, but each citation carries less answer weight than a ChatGPT one.

The retrieval pipeline, step by step

According to Fokal’s teardown of Perplexity’s documented behavior:

  1. Query decomposition. A conversational question like “What CRM is best for small B2B sales teams?” becomes several narrow keyword searches rather than one broad query. Your page needs to answer focused sub-questions, not just provide a sprawling overview.
  2. Blended index retrieval. Perplexity queries its own crawler (PerplexityBot) plus multiple third-party search APIs. Unlike ChatGPT Search, which draws exclusively from Bing’s index, Perplexity uses a blended index. Bing indexation helps but isn’t the only gate — the page needs to be crawlable by PerplexityBot and indexed by at least one major search index Perplexity queries.
  3. Full-page reading. Candidate pages are read in full by a URL-fetching tool, not just their search-result snippets. The default max_tokens_per_page parameter is 4,096 tokens per page (Fokal).
  4. Three-layer machine-learning reranker. Machine Relations describes the pipeline as: initial retrieval → quality filters → discard entire result sets that don’t clear thresholds. The L3 reranker evaluates topical authority and semantic depth.
  5. Manual authoritative-domain boosts. Perplexity maintains manual lists of authoritative domains that receive algorithmic boosts — Machine Relations names Amazon, GitHub, LinkedIn and Coursera among them.

What this means practically

  • Break your content into answerable sub-passages. A single wall of prose that “covers” the topic is worse than four self-contained 60–100-word chunks that each cleanly answer a sub-question. This is the specific reason Perplexity’s architecture rewards structure.
  • Recency matters more here than on ChatGPT. Perplexity’s index is continuously refreshed. ChatGPT relies on a static training snapshot plus SearchGPT for freshness. Updating a canonical page bumps its Perplexity ranking almost immediately.
  • Get crawlable to PerplexityBot. Don’t rely on Bing indexation alone. Add User-agent: PerplexityBot allow rules to robots.txt, verify the bot can render pages that use client-side JS, and watch server logs for its user agent.
  • Aim for the 4,096-token window. That’s roughly 3,000 words of clean prose. Pages significantly longer than that get truncated — critical passages near the end may never reach the reranker. Put the answerable content in the first two-thirds.

Where Perplexity beats Google — and where it doesn’t

Perplexity ranks passages, not pages. A well-structured page ranked lower on Google can beat a poorly structured page ranked higher, because Perplexity is picking specific paragraphs and code blocks rather than choosing the strongest URL overall (Fokal). This makes it the AI engine where careful information architecture — clear H2/H3 hierarchy, semantic HTML, one-idea-per-paragraph — has the highest leverage.

But breadth-first retrieval has a known weakness: retrievers may systematically favor documents with particular structural properties regardless of informational value (Machine Relations). That’s why lists, tables and FAQ blocks over-index on Perplexity citations — not because they’re inherently better content, but because they’re inherently more retrievable.

Actionable checklist

  1. Add a 60–100 word direct answer as the first paragraph of every page.
  2. Break long content into <h2> sections with sub-questions as <h3>.
  3. Use FAQPage JSON-LD for at least one FAQ block per pillar page.
  4. Explicitly allow PerplexityBot in robots.txt.
  5. Keep page word count under ~3,000 for maximum reranker coverage.
  6. Update canonical pages quarterly — recency is a Perplexity ranking factor in a way it isn’t on ChatGPT.

Related reading: the GEO guide covers the full eight-pillar framework, and the checklist turns this into a copy-and-run implementation list.