Configuration Parameters
Fresh
This content reflects Metehan Alp's research into AI platform internals as of 2025-2026.
A reference of all discovered internal configuration parameters across AI search platforms. These parameters are derived from research, reverse engineering, and published documentation -- they are not official documentation from the platforms themselves.
Research-Derived Parameters
These parameters are discovered through research and experimentation. Platform operators may change, rename, or remove them at any time without notice. Use as directional intelligence, not as guaranteed specifications.
ChatGPT Parameters
| Parameter | Type | Description | Impact |
|---|---|---|---|
rrf_alpha | float | Controls weight balance in RRF score fusion. Higher values increase the influence of top-ranked results relative to lower-ranked ones. | Determines how steeply rank position affects citation probability |
rrf_input_threshold | float | Minimum score a result must achieve in initial retrieval to be included in the RRF fusion step. | Acts as a quality gate -- content below this threshold is never considered for citation |
ranking_model | string | Selects which neural reranker processes results after initial retrieval (e.g., Skysight). | Different models may weight different content signals |
use_freshness_scoring_profile | boolean | Enables time-decay weighting that favors recently published or updated content. | When enabled, older content receives a scoring penalty proportional to age |
enable_query_intent | boolean | Activates intent classification that routes queries to different retrieval pipelines (informational, navigational, transactional). | Affects which content types are retrieved for different query categories |
vocabulary_search_enabled | boolean | Enables vocabulary-based search matching in addition to semantic search. | When enabled, exact keyword matches contribute to retrieval scoring alongside semantic similarity |
Perplexity Parameters
| Parameter | Type | Description | Impact |
|---|---|---|---|
l3_reranker_drop_threshold | float | Score below which results are dropped after the L3 (third-level) reranking pass. | Controls how aggressively low-relevance results are filtered. Lower threshold = more sources considered. |
subscribed_topic_multiplier | float | Boost factor applied to domains associated with topics the user has shown interest in. | Personalization signal -- domains in user's interest graph get a citation boost |
time_decay_rate | float | Rate at which older content loses ranking weight. Higher values = faster decay. | Determines how much freshness matters relative to authority and relevance |
embedding_similarity_threshold | float | Minimum cosine similarity between query embedding and content embedding for inclusion. | Content below this semantic similarity threshold is excluded from consideration entirely |
Google AI Mode Parameters
| Parameter | Type | Description | Impact |
|---|---|---|---|
| Chunk size | ~500 tokens | Content is processed in approximately 500-token windows during retrieval and generation. | Content sections should be structured to align with this processing window. Sections longer than 500 tokens may be partially processed. |
| Engagement tier thresholds | 100K queries | Content must be associated with topics receiving 100K+ queries to enter the highest engagement tier. | High-volume topics receive more AI Mode coverage. Niche topics may not trigger AI Mode responses. |
How to Use These Parameters
For Content Strategy
Understanding these parameters helps you make better optimization decisions:
- rrf_alpha and rrf_input_threshold tell you that ranking position matters, but the relationship is not linear. Small position improvements near the input threshold have outsized impact.
- use_freshness_scoring_profile confirms that content freshness is an active signal, not just a tiebreaker. Update content regularly.
- l3_reranker_drop_threshold means Perplexity aggressively filters low-relevance results. Your content must be highly relevant to survive 3 levels of reranking.
- embedding_similarity_threshold means semantic alignment (not just keyword matching) determines whether your content is even considered.
- 500-token chunk size means your content structure should create self-contained 375-word sections that make sense independently.
For Competitive Analysis
When a competitor consistently gets cited and you do not, these parameters suggest diagnostic questions:
- Are they ranking for more sub-queries than you? (RRF breadth)
- Is their content newer? (freshness scoring)
- Is their content more semantically aligned with the query? (embedding similarity)
- Are they on Perplexity's authoritative domain list? (topic multiplier)
- Does their content structure align with 500-token chunks? (chunk size)
Parameter Discovery Methods
These parameters were identified through:
- API documentation analysis -- Some parameters appear in public API docs or developer documentation
- Network request inspection -- Observing API calls between frontend and backend during search interactions
- A/B testing -- Publishing content variations and measuring citation differences to infer active parameters
- Open-source code analysis -- Examining open-source components of AI search systems
- Research paper references -- Academic papers from platform engineers that reference internal parameters
Further Reading
- AI Search Engines Reference -- How each platform uses these parameters in practice
- Platform Comparison -- Side-by-side feature comparison