Skip to main content
ZICQ

Wiki AI Concepts

AI Search (Answer Engine)

AI Concepts
Aliases: Answer Engine AI Search Engine Grounded Search ·2026-11-18

AI Search (often called an Answer Engine) combines live web retrieval with large language model synthesis to deliver direct, source-cited answers rather than a raw list of ten blue links.

Today this pattern spans standalone consumer engines (such as Perplexity and You.com), integrated search overviews in traditional engines (Google AI Overviews, Bing Copilot), and agent-oriented search APIs (such as Tavily and Exa) purpose-built for LLM tool calls.

Dimension Traditional Keyword Search AI Search (Answer Engine)
Output Format Ranked links + short snippets Direct structured synthesis with inline source citations
Multi-Hop Synthesis User opens multiple tabs and compares manually Engine decomposes sub-queries, reconciles conflicts, summarizes diffs
Latency & Cost Sub-100ms, extremely cheap per query Seconds (search + fetch + LLM inference), significantly higher compute cost
Primary Failure Mode SEO spam and clickbait farms Misattributed citations, summarization hallucinations, stale caches

Core Reference Pipeline

User query
  └── Query rewriting & sub-question decomposition
       └── Live retrieval (web index + semantic search APIs)
            └── Clean page extraction, chunking, and reranking
                 └── [RAG](/wiki/rag) grounded generation → streamed answer with inline citations

For multi-minute Deep Research queries, an iterative agent runs multiple search-read-refine cycles—adjusting follow-up queries based on gaps discovered mid-run before assembling a cited report.

Key Engineering Challenges

  • Citation faithfulness (Grounding): Every claim, statistic, and version number should trace cleanly to a fetched excerpt from the linked citation rather than unverified parametric memory;
  • Freshness vs. AI-generated web spam: Ranking pipelines must elevate authoritative primary sources and official documentation while filtering low-signal AI-generated content farms;
  • Agent-native search APIs: Unlike legacy SERP scrapers optimized for browser snippets, modern search APIs (such as Tavily and Exa) return clean extracted Markdown and semantic filters tuned to preserve context budget during tool calling.

Further Reading