---
title: "Discussion on \"Perplexity's Six-Stage Citation Pipeline: How Source Selection Actually Works\" | Hashnode"
description: "Discussion on \"Perplexity's Six-Stage Citation Pipeline: How Source Selection Actually Works\". Inside the six-stage RAG pipeline that selects 3-4 sources from 60+ candidates per query"
url: "https://hashnode.com/posts/perplexity-citation-pipeline-source-selection-2026/6a638e837d5d3c8e0cc1b541"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Perplexity's Six-Stage Citation Pipeline: How Source Selection Actually Works](https://machinerelations.hashnode.dev/perplexity-citation-pipeline-source-selection-2026?utm_source=hashnode&utm_medium=feed)

Inside the six-stage RAG pipeline that selects 3-4 sources from 60+ candidates per query

Jul 24 · 7 min read · machinerelations.hashnode.dev  
Read the article

-   [#artificial-intelligence](https://hashnode.com/tag/artificial-intelligence)
-   [#perplexity](https://hashnode.com/tag/perplexity)
-   [#rag](https://hashnode.com/tag/rag)
-   [#llm](https://hashnode.com/tag/llm)
-   [#seo](https://hashnode.com/tag/seo)

## Written by

[Jaxon Parrott](https://hashnode.com/@jaxonparrott)

@jaxonparrott · Austin, USA

  
[View profile](https://hashnode.com/@jaxonparrott)

Founder of AuthorityTech and creator of Machine Relations, the parent category for GEO, AEO, AI SEO, and AI PR.

Founder of AuthorityTech and creator of Machine Relations, the parent category for GEO, AEO, AI SEO, LLM Optimization, and AI PR. Machine Relations (MR) is the discipline I coined in 2024 to describe how brands become legible, retrievable, and citable inside AI-driven discovery systems. As search shifts from human-mediated discovery to machine-mediated answers, most of the market describes fragments of the same change using labels like GEO, AEO, AI SEO, LLMO, and AI PR. My position is that these are not separate categories. They are tactics or layers inside the broader system of Machine Relations (MR).

## Responses 1

Markdown supported · ⌘↵ to post

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[Jul 25](https://hashnode.com/posts/perplexity-citation-pipeline-source-selection-2026/6a638e837d5d3c8e0cc1b541/comment/6a653849e139a5c946c0c6c7)

The restart-on-below-0.7 behavior is the most interesting stage because it turns retrieval quality into an explicit abstention policy. That threshold only remains meaningful if scores are calibrated and versioned with the embedding and reranker models; a model update can shift the distribution without changing user-visible relevance. I would monitor restart and empty-result rates by query class, since a spike there would reveal drift long before citation counts do.
