metehan.ai Research & SOPs

Appearance

AI Search Engines Reference ​

Fresh

This content reflects Metehan Alp's research findings on AI search engine internals as of 2025-2026.

A reference guide covering how each major AI search platform ranks, retrieves, and cites content.

ChatGPT (OpenAI) ​

Ranking System ​

  • RRF (Reciprocal Rank Fusion) -- Combines multiple ranked lists from sub-query decomposition
  • Skysight Reranker -- Neural reranker that rescores results after initial retrieval
  • Freshness Scoring -- Time-decay weighting that favors recently published or updated content

Attribution ​

  • PSL-based attribution -- Uses the Public Suffix List to determine source domains for citation
  • Citations link to the source domain, not individual pages in many cases
  • Direct URL citations appear in ChatGPT's search mode (OAI-SearchBot powered)

Key Parameters ​

  • rrf_alpha -- Controls the weight balance in RRF fusion
  • ranking_model -- Selects which reranker model processes results
  • use_freshness_scoring_profile -- Enables time-decay weighting
  • enable_query_intent -- Activates intent classification for query routing

Content Signals ​

  • Authoritative domains preferred (high domain authority, established presence)
  • Long-form content with structured headings performs well
  • FAQ-structured content gets cited at higher rates
  • Original research and data-backed claims receive preferential citation

Perplexity ​

Ranking System ​

  • XGBoost L3 Reranker -- Machine learning reranker (L3 = third-level reranking pass)
  • Authoritative Domain Lists -- Curated lists of trusted domains per topic category
  • Topic Multipliers -- Boost factors applied to domains with established topic authority

Attribution ​

  • Inline citations with numbered references
  • Each claim in the response is attributed to a specific source
  • Multiple sources per response (typically 4-8)
  • Source preview cards with title, URL, and excerpt

Key Parameters ​

  • l3_reranker_drop_threshold -- Score below which results are dropped from consideration
  • subscribed_topic_multiplier -- Boost factor for domains the user follows
  • time_decay_rate -- How quickly older content loses ranking weight
  • embedding_similarity_threshold -- Minimum semantic similarity for inclusion

Content Signals ​

  • Factual density matters (claims per paragraph)
  • Structured data (tables, lists) gets preferentially extracted
  • Author credentials influence trust scoring
  • Domain authority weight is high (established publishers dominate)

Google AI Mode ​

Ranking System ​

  • 4-stage pipeline:
    1. Query understanding and decomposition
    2. Retrieval across Google's index
    3. Reranking with neural models
    4. Response generation with inline citations
  • Jetstream -- Google's serving infrastructure for real-time AI response generation

7 Ranking Signals ​

  1. Semantic relevance -- Content meaning matches query intent
  2. Source authority -- Domain authority and E-E-A-T signals
  3. Content freshness -- Recency of publication or last update
  4. Structured data -- Presence and quality of schema markup
  5. Content depth -- Comprehensive topic coverage
  6. User engagement -- Historical click-through and dwell time signals
  7. Entity alignment -- Named entity overlap between query and content

Content Signals ​

  • 500-token chunk processing -- Content is processed in approximately 500-token windows
  • Heading hierarchy matters heavily (clear H1 > H2 > H3 structure)
  • Answer-first content structure preferred
  • Lists and tables get preferential extraction

Google AI Overviews ​

Ranking System ​

  • Semantic relevance scoring against the query
  • Structured data weighting (schema markup)
  • Content comprehensiveness scoring
  • Source diversity (multiple sources preferred over single-source)

Content Signals ​

  • Long-tail, informational queries trigger AI Overviews most frequently
  • Content with clear question-answer structure performs best
  • Table and list formatting increases extraction probability
  • Pages already ranking in the top 10-20 are preferentially selected
  • Cosine similarity between content and query is a primary signal

Key Characteristics ​

  • Sources typically come from pages already ranking on page 1-2 of Google
  • 3-5 sources cited per AI Overview (average)
  • Visual elements (images, diagrams) may be included from source pages

Google Discover ​

Ranking System ​

  • NAIADES -- Google's Discover recommendation engine
  • 13 cluster types -- Content is categorized into 13 topic clusters for recommendation matching
  • Tombstoning -- Process that removes content from Discover feed after engagement drops

Content Signals ​

  • High-quality images (1200px+ width) are required for inclusion
  • Topic relevance to user interests (based on browsing history and search patterns)
  • Content freshness is heavily weighted (new content preferred)
  • E-E-A-T signals influence inclusion probability
  • Engagement velocity (how quickly new content accumulates interactions) affects promotion

Key Characteristics ​

  • Not query-driven (recommendation-based, not search-based)
  • Mobile-first (primarily surfaces on mobile devices)
  • Visual-first (image quality directly impacts CTR)
  • Ephemeral traffic (Discover traffic spikes and fades, unlike search traffic)