ML//Knowledge Cutoff

- The date at which an LLM's training data was frozen. Everything published after this date is invisible to the model unless retrieved in real time.

The date at which an LLM's training data was frozen. Everything published after this date is invisible to the model unless retrieved in real time.

Not a clean cut — training data is a snapshot of the web at crawl time, processed and filtered over months. The "cutoff" is the latest date in the corpus, not a single collection event.

Implications for web devs: even with perfect AEO, your content only enters a model's knowledge if it was crawled before the cutoff and survived filtering. Getting into training data is a months-long pipeline, not an instant event.

Real-time retrieval bypasses the cutoff: Perplexity, ChatGPT with search, and Google AI Overviews fetch live content. This is faster but less sticky — the model does not "know" you, it just found you.

Training data vs retrieval knowledge: training data shapes the model's priors (it believes things). Retrieved content is contextual (it reads things). Being in training data means the model recommends you unprompted. Being in retrieval means it cites you when asked.

Current cutoffs (as of early 2026): GPT-4 ~Apr 2024, Claude 3.5 ~Apr 2024, Gemini ~varies. Each new model generation pushes the cutoff forward.