Grounding
Connecting an AI's response to real, verifiable sources rather than memory alone.
Grounding refers to tethering an AI's outputs to specific, verifiable real-world information — so responses are traceable to a source rather than purely generated from training weights.
Why it matters: A grounded response can say "According to the Q2 2026 earnings report, revenue was $4.2B [source]." An ungrounded one says "Revenue was approximately $4B" — possibly accurate, possibly hallucinated.
Methods for grounding:
- RAG (Retrieval-Augmented Generation) — the most common: retrieve relevant documents and insert them as context
- Web search — real-time retrieval from the internet (Perplexity, Bing AI, Google AI Overviews)
- Structured data injection — paste tables, JSON, or database results directly into the prompt
- Citations — instruct the model to cite specific passages; models like Claude can quote source text
Grounded vs. grounded-aware: Grounding is not just about sourcing — it's about the model actually using the provided source rather than ignoring it and relying on training memory. Verify that responses reference injected content, not just assert matching facts.
In plain terms
The difference between a journalist who files a story from memory vs. one who grounds every claim in a document, interview, or official record — both may say the same thing, but only one can be fact-checked.
Related concepts
Retrieval-Augmented Generation
Giving the AI access to your documents before it answers.
Hallucination
When an AI generates confident-sounding facts that are simply wrong.
Knowledge Cutoff
The date after which an AI model has no training data — and doesn't know what happened.
Vector Database
A database that stores meaning, not just text — finding similar content by concept rather than keyword.