Reciprocal Rank Fusion (RRF) Explained

Reciprocal Rank Fusion (RRF) is the algorithm ChatGPT uses to decide which sources to cite

RRF is a result merging algorithm that combines multiple ranked lists into a single unified ranking. ChatGPT generates multiple search queries for each user question, retrieves Bing results for each query, and uses RRF to merge all results into one ranked list. Pages that consistently appear across multiple query variations accumulate the highest scores and get cited.

The RRF formula

The formula is straightforward:

RRF_score(d) = ∑ 1 / (k + rank_i(d))

Where:

• d is the document (web page)
• k is a constant, typically 60
• rank_i(d) is the rank position of document d in the i-th query result list
• The sum is across all query variations where the document appears

For a page ranked #1 in one query: score = 1/(60+1) = 0.0164. If the same page also ranks #3 in a second query: total score = 0.0164 + 1/(60+3) = 0.0164 + 0.0159 = 0.0323. The more queries a page appears in, the higher its cumulative RRF score.

Why RRF rewards consistency over dominance

Consider two competing pages:

• Page A ranks #1 for 2 query variations: RRF score = 2 × (1/61) = 0.0328
• Page B ranks #5 for 10 query variations: RRF score = 10 × (1/65) = 0.1538

Page B wins by a factor of nearly 5x — even though it never ranks #1. This is the core insight: broad, consistent visibility across many related queries is mathematically superior to dominating a few queries.

How ChatGPT generates query variations

When you ask ChatGPT something, it generates what the industry calls "fanout queries". For a question like "best cybersecurity consultant in Australia", ChatGPT might generate:

• "top cybersecurity consulting firms Australia"
• "Australian cybersecurity companies reviews"
• "cyber security consultant Sydney Melbourne"
• "VAPT services Australia"
• "Essential Eight compliance consultant Australia"
• "cybersecurity audit Australian business"

Each of these queries returns a separate ranked list from Bing. RRF merges all lists together. A consultancy that ranks in the top 10 across all six queries will dominate the final merged ranking — even if it never ranks #1 for any individual query.

What this means for content strategy

Build topic clusters, not single pages

A single "ultimate guide" page can only rank for a limited number of query variations. But a topic cluster — 10-20 interlinked articles covering every facet of a subject — can rank across dozens of query variations. Each article captures different fanout queries, and the domain accumulates aggregate RRF scores.

Target long-tail variations

Every long-tail query your content ranks for adds to your cumulative RRF score. Research every way someone might ask about your expertise and create content that addresses those specific variations. FAQ sections, comparison articles, checklists, and how-to guides each capture different query types.

Internal linking amplifies RRF

Strong internal links between related articles signal to Bing that your content forms a cohesive knowledge cluster. This improves individual page rankings across Bing, which directly improves your RRF scores in ChatGPT's fusion algorithm.

ChatGPT's confirmed RRF parameters

Developers who inspected ChatGPT's network requests found three key parameters:

• rrf_alpha: 1 — the scaling factor is neutral. All query fanouts carry equal weight
• rrf_input_threshold: 0 — no minimum threshold. Results from every query variation are included, even pages that appear in only one search
• ranking_model: null — no custom ML model overrides the RRF algorithm. It's pure mathematical fusion

This is significant because it means the system is predictable. There's no "black box" ML model reranking results unpredictably — it's straightforward math that rewards consistent Bing visibility.

RRF vs other fusion methods

RRF was chosen over alternatives like CombSUM, CombMNZ, and Borda Count because of its key advantages:

• No score normalisation needed — RRF works on rank positions, not raw scores, making it robust across different query types
• Resistant to outliers — a single very high or very low score doesn't distort the merged ranking
• Simple to implement and audit — the algorithm is deterministic and reproducible