Parametric Knowledge

Parametric knowledge is the information a language model learned during training and stores in its parameters (its weights), which it can use without looking anything up. When an assistant explains photosynthesis, names the capital of Peru or describes what a well-known software company sells without running a search, it is drawing on parametric knowledge. Information fetched at answer time through search or documents is the contrast, often called retrieved or non-parametric knowledge.

The terms come from research. In 2019, Fabio Petroni and colleagues at Facebook AI Research and University College London showed that BERT, without any fine-tuning, recalled relational facts from fill-in-the-blank prompts about as well as traditional NLP methods with some access to oracle knowledge, and asked whether language models could serve as knowledge bases themselves (Petroni et al.). The 2020 paper that introduced retrieval-augmented generation framed its design as combining 'parametric memory' (the model) with 'non-parametric memory' (a searchable index), precisely because parametric knowledge is hard to update and cannot show its sources.

Parametric knowledge has four properties that matter in practice:

• It is frozen at a knowledge cutoff. Anthropic, for example, lists June 2026 as the reliable knowledge cutoff for its current Claude models (Anthropic); anything published later can only reach those models through retrieval.
• It reflects how often and how consistently something appeared in the training data. Widely and consistently described facts are learned well; rare or contradictory ones are learned poorly or not at all.
• It carries no citations. A model answering from memory cannot link to where it learned something, so these answers never send referral traffic.
• It can be wrong with confidence, which is one source of AI hallucination.

Modern assistants mix both kinds of knowledge and decide per question which to use. ChatGPT searches on a minority of prompts: Profound reported in February 2026 that about 18% of the ChatGPT conversations it analyzed from late 2025 triggered a web search, and Nectiv measured searches on 31% of prompts in October 2025. Everything else is answered from parametric knowledge. Even grounded answers lean on it, because the model's prior understanding shapes which queries it writes and how it reads the sources it finds.

For GEO, the distinction drives two different strategies. Retrieved visibility, with citations and links, depends on crawlability, rankings for the model's search queries and strong passages, and it can change within days. Parametric visibility depends on what the web said about your brand, consistently and in many places, before the model's cutoff, and it changes only when new models are trained. A brand founded after a model's cutoff has no parametric presence in that model at all. Measure the share of your prompts that are answered without retrieval using grounding rate, and track how models describe you when they do not search, which is what share of model captures.

The two feed each other. Reference sources such as Wikipedia, which Profound found cited in about one in six ChatGPT conversations that included citations, appear both in training data and in live retrieval. Pages that are cited often today are likely to be part of what future models learn from.