Training data versus retrieval

Parametric memory is what a model produces with no tools: facts compressed into weights during training. Non-parametric memory is a set of documents fetched at query time and placed in the prompt, the arrangement behind retrieval-augmented generation. They fail differently. A model reciting from weights cannot show a source, and a retrieved page can be quoted accurately by a model that never saw it in training. Production engines carry both and decide per query whether to search^[3].

Vendors publish the date. Anthropic lists a reliable knowledge cutoff of June 2026 for Claude Fable 5.1, May 2026 for Claude Opus 5, January 2026 for Claude Sonnet 5 and February 2025 for Claude Haiku 4.5^[1], and treats that figure as distinct from the training data cutoff, publishing the pair for each model in its Transparency Hub rather than in the comparison table^[4]. Microsoft documents GPT-4.1 as trained to May 31, 2024^[5], GPT-5 to September 30, 2024^[2], GPT-5.5 to December 2025^[6] and GPT-5.6 to June 2026^[7], a gap that has closed from roughly eleven months to about one. The published date is an upper bound rather than a description of what a model knows, since effective cutoffs drift from reported ones when old pages recur in new crawls and deduplication misses near-duplicates^[8].

Retrieval is a tool call, and on most platforms the model decides. OpenAI documents that the model chooses whether to search based on the prompt, with tool_choice set to required when the caller wants a search every time^[3]. Google describes the same behaviour for grounding^[9]. Perplexity ships the opposite default and grounds every answer^[10]. A question about a settled, well-covered subject may therefore never reach a crawler, while a question about this week's prices always does. What the engine does after it decides to search is covered in query fan-out.

Popularity decides which store answers. On PopQA, a 14,000-question probe, models answer about well-known entities from weights alone and fail on the long tail, and scale does not close the gap^[11]; retrieval-augmented models beat models orders of magnitude larger there, while unassisted models hold their own on popular entities^[12]. A 2025 study of vision-language models points the other way: finetuned models rely more on memorised training data than retrieval-augmented ones do, and answer the WebQA test set at 72% against 52%^[13]. The two results are measured on different modalities and different benchmarks. A heavily discussed brand can be named without any page being fetched, and an obscure one can be cited without being remembered. Citation quality is a third question: an audit of four generative search engines found 51.5% of sentences fully supported by their citations^[14] and 74.5% of citations supporting the sentence they sat on^[15].

The two paths take different work. Training-time presence comes from being written about across the corpus a crawler sweeps, which is why brand mentions on third-party sites count even where they carry no link. Retrieval-time presence comes from being fetchable and quotable during the request, which is what AI crawlers and content freshness govern. Only the second is testable within a quarter: the GEO-bench study measured visibility gains of up to 40% from rewriting source pages^[16], while a training run is fixed and the next one is months away.