Anatomy of AI Answer Engines: Inverted Index vs Neural RAG Vectors
Traditional search engines (Google PageRank, Bing Lucene) operate on an inverted keyword index. They scan an index of tokenized terms, compute TF-IDF and BM25 relevance scores, and rank pages using link equity and anchor text authority.
Frontier AI Answer Engines (ChatGPT Search, Perplexity Sonar, Google Gemini Grounding, Claude Web Search) operate on a fundamentally different pipeline: Multi-Stage Retrieval-Augmented Generation (RAG).
RFC 9309 lookup; sub-120ms HTTP/3 retrieval.
Boilerplate, scripts, & ads stripped to extract clean text.
Content sliced into 512-token dense embeddings.
Cosine similarity against user prompt & query intents.
LLM generates answer with anchored inline citations.