Skip to content
VectorCite
Free audit

Glossary entry training-data

Training data

The corpus a model was trained on, distinct from the live web it retrieves citations from at query time.

Filed under
Core concepts

Definition

Training data is the static corpus used to pre-train or fine-tune a language model. Common Crawl, books, code, web pages, all frozen at a cutoff date. Engines reference training data for general knowledge but cite the live web (RAG) for current facts, comparisons, and product info. The GPTBot, ClaudeBot, Google-Extended user-agents are training crawlers; the *SearchBot variants are retrieval crawlers.

Example

Claude's training data has a cutoff (e.g. April 2024). For 'best CRM 2026' Claude invokes web tool. The cited sources are live retrievals, not training data.

How to optimize

Be in both. Allow training bots (GPTBot, ClaudeBot, Google-Extended) AND retrieval bots (OAI-SearchBot, Claude-Web, PerplexityBot). The training presence helps with brand recognition; the retrieval presence wins citations.

Check a page against a real AI answer

Claude’s live answer to your buyer’s question, its sources, your readiness score and the fixes. Free, no signup, about half a minute.

Run a free audit