Reciprocal Rank Fusion & AI Citations
ChatGPT does not use a single ranking to decide which sources to cite. It runs multiple queries, gets multiple ranked lists, and fuses them together using Reciprocal Rank Fusion (RRF). This has direct, measurable consequences for content strategy -- and it gives topic clusters a mathematical advantage over standalone pages.
The RRF Formula
$$\text{RRF}(d) = \sum_{r \in R} \frac{1}{k + r(d)}$$
Where:
- d = a document (webpage)
- R = the set of all ranked lists
- r(d) = the rank of document d in ranked list r
- k = a constant (set to 60 in ChatGPT's implementation)
In plain terms: for each ranked list a document appears in, add 1 / (60 + its rank position). The document's final score is the sum across all lists.
Why k = 60?
The constant k controls how much rank position matters. A high k value (like 60) means the difference between rank 1 and rank 5 is small, but appearing in multiple lists matters a lot. ChatGPT chose k = 60, which heavily rewards breadth of relevance over dominance in a single ranking.
The ChatGPT DevTools Discovery
Metehan found these parameters by inspecting ChatGPT's network traffic in browser DevTools during search-enabled conversations:
| Parameter | Value | What It Does |
|---|---|---|
rrf_alpha | 60 | The k constant in the RRF formula |
rrf_input_threshold | Varies by query | Minimum number of input ranked lists before RRF fusion activates |
These are not theoretical. They are live configuration values observed in ChatGPT's search requests.
How RRF Works in Practice
When a user asks ChatGPT a complex question, the system:
- Decomposes the query into multiple sub-queries
- Retrieves a ranked list of results for each sub-query
- Fuses all ranked lists using RRF with k=60
- Selects citations from the top of the fused ranking
The Math Advantage of Topic Clusters
This is where RRF changes everything for content strategy. A domain that ranks across multiple sub-queries accumulates RRF score faster than a domain that ranks #1 for only one.
Example 1: The 1.8x Advantage
Query: "Best project management tools for remote teams"
Sub-queries generated:
- "project management tools comparison"
- "remote team collaboration software"
- "best tools for distributed teams"
Single-page domain (ranks #1 for one sub-query only):
- RRF score = 1/(60+1) = 0.0164
Topic cluster domain (ranks #3, #5, #4 across all three):
- RRF score = 1/(60+3) + 1/(60+5) + 1/(60+4) = 0.0159 + 0.0154 + 0.0156 = 0.0469
The topic cluster domain scores 2.9x higher despite never ranking #1.
Example 2: The 10x Advantage
Scale this to a query that decomposes into 10 sub-queries. A domain with a deep content cluster covering all 10 facets of a topic will appear in most or all ranked lists. A competitor with one strong page appears in 1-2 lists at best.
| Scenario | Lists Appearing In | Avg Rank | RRF Score | Multiplier |
|---|---|---|---|---|
| One strong page | 1 | #1 | 0.0164 | 1x |
| Topic cluster (moderate) | 5 | #4 | 0.0781 | 4.8x |
| Topic cluster (deep) | 10 | #5 | 0.1538 | 9.4x |
Example 3: The 60x Advantage (Extreme Case)
For broad informational queries that decompose into 20+ sub-queries, a domain with comprehensive topical coverage can accumulate RRF scores 60x higher than a competitor with a single authoritative page. The math is unforgiving.
This Changes Strategy
Traditional SEO rewards ranking #1 for a single keyword. RRF rewards covering a topic comprehensively. A domain that ranks #5 across 10 related queries will outperform a domain that ranks #1 for one query. This is why topic clusters are not optional in AEO -- they are a mathematical requirement.
Multi-Source Integration
ChatGPT's RRF implementation does not just fuse results from web search sub-queries. It integrates multiple source types, each with their own ranked list:
| Source Type | Description | Weight in Fusion |
|---|---|---|
webpage | Standard web search results | Primary |
webpage_extended | Extended snippets from long-form content | Primary |
grouped_webpages | Multiple pages from the same domain, grouped | Boosted (same-domain clustering) |
image_inline | Image results relevant to the query | Secondary |
The grouped_webpages source type is particularly important. When ChatGPT detects multiple relevant pages from the same domain, it groups them -- and this grouping can boost the domain's aggregate RRF score even further.
Strategic Recommendations
For Content Teams
Build topic clusters, not standalone pages. Every page in your cluster is another entry point into RRF ranked lists. Five pages covering different facets of a topic will almost always outperform one comprehensive page.
Cover sub-topics explicitly. When planning content, ask: "What sub-queries could this topic decompose into?" Write a page for each one.
Internal link clusters tightly. ChatGPT's
grouped_webpagessource type rewards domains with clearly related content. Strong internal linking signals topical relationships.Target long-tail variations. Each long-tail page you publish is another potential appearance in a ranked list. RRF rewards breadth.
For Technical SEOs
Monitor query decomposition patterns. Use ChatGPT's search mode to see how it breaks down complex queries. Map these decompositions to your content coverage.
Track multi-query visibility. Traditional rank tracking measures one keyword per page. For AEO, track how many sub-queries your domain appears in for a given topic.
Optimize for
grouped_webpages. Clear site architecture, consistent URL structures, and strong internal linking help ChatGPT group your pages together.
For Strategists
Prioritize topic depth over topic breadth. It is better to thoroughly cover 5 topics (with 10+ pages each) than to thinly cover 50 topics (1-2 pages each). RRF rewards depth within a topic.
Use the RRF formula to model expected citation probability. Before publishing, estimate how many sub-queries your content cluster will cover and calculate the expected RRF advantage.
The Core Principle
RRF means that appearing everywhere on a topic matters more than ranking first for one query. Build deep, not wide.