How I Reverse-Engineered ChatGPT's Ranking Algorithm (And What It Means for Your Content)
You can verify everything I'm about to show you in less than 30 seconds.
Last week, while diving through ChatGPT's source code, I stumbled upon something fascinating: the actual configuration files that control how ChatGPT decides which websites to cite. Not theories. Not best practices. The actual production settings.
And here's the kicker, you can see it yourself right now. Open any ChatGPT conversation, click "View Source Code," hit Ctrl+F, and search for "rerank." There it is, plain as day.
The Discovery
What started as casual curiosity turned into a deep dive through thousands of lines of configuration data. Hidden in the React Router context stream, I found the complete retrieval and ranking pipeline that ChatGPT uses every time it searches the web.
The star of the show? A reranking model called ret-rr-skysight-v3.
reranker_model: "ret-rr-skysight-v3"
This isn't just a search algorithm, it's a sophisticated neural network that takes initial search results and completely reorders them based on signals we're only beginning to understand.
The Freshness Obsession
But here's what should make every content creator sit up straight:
use_freshness_scoring_profile: true
This single line of code confirms what many suspected but couldn't prove: ChatGPT is obsessed with fresh content.
That comprehensive guide you perfected in 2022? It's already losing ground to mediocre content published last week. This isn't a bug, it's a feature. The system literally has a dedicated scoring profile for prioritizing recent information.
I found additional evidence in the retrieval prompt itself, which explicitly instructs: "Use the web tool instead when the user asks about recent events / fresh information."
The Multi-Stage Filtering Pipeline
The configuration reveals ChatGPT doesn't just search and rank. It runs content through multiple sophisticated filters:
enable_query_intent: true
enable_source_filtering: true
enable_mimetype_filtering: true
vocabulary_search_enabled: true
use_coarse_grained_filters_for_vocabulary_search: false
Here's what's actually happening when ChatGPT searches:
Intent Detection: First, it figures out what you're really asking for—definition, comparison, tutorial, etc.
Vocabulary Analysis: It recognizes domain-specific terminology (and rewards sites that use it correctly)
Source Filtering: Different sources get different treatment
MIME Type Awareness: It knows if it's looking at HTML, PDF, or other formats
Fine-Grained Filtering: No broad strokes—everything gets precise evaluation
The "Slurm" Mystery Solved
One of the most intriguing findings was this setting:
use_light_weight_scoring_for_slurm_tenants: true
What the hell is "slurm"?
The answer was hiding in the connector configuration:
"slurm_dropbox"
"slurm_sharepoint"
"slurm_box"
"slurm_canva"
"slurm_notion"
"Slurm" refers to connected third-party services. When ChatGPT searches your personal Dropbox or Notion, it uses lightweight scoring. But when it searches the public web? Full neural reranking arsenal deployed.
This reveals something crucial: ChatGPT uses completely different strategies for public vs. private content.
The Disabled Mystery
Equally interesting is what's turned OFF:
use_relevance_lmp: false
I don't know what "LMP" stands for (Language Model Predictions? Latent Matching Process?), but ChatGPT has explicitly chosen not to use it. This suggests the system might be relying on more traditional information retrieval methods combined with the neural reranker, rather than some bleeding-edge experimental feature.
What This Actually Means for Your Content
After analyzing these configurations, here's what actually matters:
1. Freshness Isn't Optional
That freshness scoring isn't a gentle preference—it's a hard-coded priority. You need a content refresh strategy, period. Quarterly updates minimum.
2. Intent Clarity Is Everything
With intent detection active, your content needs to telegraph exactly what it is. Don't write a product comparison disguised as a blog post. Make your content's purpose crystal clear in structure and language.
3. Domain Expertise Pays Off
The vocabulary search system with fine-grained filters rewards proper use of technical terminology. This isn't about keyword stuffing—it's about demonstrating genuine expertise through precise language.
4. The Reranker Changed the Game
Getting into initial search results isn't enough. The ret-rr-skysight-v3 model will reshuffle everything. Your content needs to be so good it survives any reordering.
The Uncomfortable Truth
These configurations reveal an uncomfortable truth for SEO professionals and content creators: the game has fundamentally changed.
You're not optimizing for a search algorithm anymore. You're optimizing for a multi-stage AI pipeline that includes:
Intent detection
Vocabulary analysis
Freshness scoring
Source filtering
Neural reranking
Gaming this system would require fooling multiple independent evaluators, including a sophisticated neural network we don't fully understand.
The Bottom Line
The configuration doesn't lie. These settings are live in production, affecting every ChatGPT search right now.
The path forward isn't about tricks or hacks. It's about becoming genuinely authoritative, staying relentlessly current, and communicating with crystal clarity. The machines are getting better at recognizing quality. Maybe it's time we focused on actually delivering it.
You can verify all of this yourself. Open ChatGPT → View Source → Search "rerank". The full configuration analysis and code snippets are available on my GitHub.
This analysis is based on ChatGPT Plus configuration from August 2025. Settings may vary by user type and region.
P.S. - The most mind-blowing part? This configuration is visible to anyone who looks. It's been sitting there in plain sight, controlling billions of AI-generated responses, just waiting for someone to dig in and decode it.