Definition

Retrieval-augmented generation (RAG)

RAG is the architecture behind most cited AI answers: the system retrieves relevant documents at query time and conditions the generated answer on them, rather than relying only on model weights.

This is why current content can be cited by a model whose training cut-off predates it, and why AI visibility is a live measurement problem rather than a training-data one.

It also explains citation volatility: retrieval varies between runs, so the source set varies with it.

Related terms

Grounding Citation Non-determinism

Nostimates measures retrieval-augmented generation across twelve AI answer surfaces. Request access →

All terms

AEO (Answer Engine Optimisation)GEO (Generative Engine Optimisation)AI Overviews AI Mode Citation share n-sampling Query fan-out Grounding Share of voice (AI)Extraction success rate Surface Answer engine Citation Brand mention Mention rate Confidence interval Statistical significance Citation volatility Non-determinism Reproducibility Identity-free collection Personalisation Locale Residential proxy Bot detection Collection cost Credit Evidence retention Audit trail Parser version Canary prompt Data freshness Cadence Prompt set Head term Long-tail query Competitive set llms.txt robots.txt AI crawler Training crawler Hallucination Answer box Featured snippet Zero-click search Entity Entity salience Knowledge graph Structured data JSON-LD schema.org FAQ schema Answer-first content Chunk Embedding Semantic search Prompt injection Market coverage Sentiment analysis Position tracking White-label data