Knowledge cutoff

A knowledge cutoff is the fixed point in time after which a model's training data ends, so the model has no built-in knowledge of events, pages, or facts that first appeared after it¹. It is a property of how the model was trained, not of the product you are using: two systems built on the same base model share its cutoff, even if one can search the web and the other cannot.

The distinction that matters for AI search is between parametric knowledge and retrieved knowledge. Parametric knowledge is what the model absorbed during training, frozen at the cutoff. Retrieved knowledge is content the system fetches at answer time from outside the model. The cutoff bounds the first and says nothing about the second. A model whose training ended last year can still answer about this morning's news, but only by retrieving it, not by recalling it.

Status in 2026

Major models today ship with a documented cutoff. OpenAI, Anthropic, and others publish a per-model cutoff date²¹, and in practice it tends to lag the present by months. That lag is a primary reason the consumer engines added retrieval, though not the only one: retrieval also supplies verification, long-tail and private knowledge, and sources a user explicitly asks for. Anthropic's own documentation describes its web search tool as giving the model "direct access to real-time web content, allowing it to answer questions with up-to-date information beyond its knowledge cutoff," and states in the same place that "the response includes citations for sources drawn from search results"¹. Read those two sentences together and a large part of the GEO opportunity falls out of them: the cutoff creates a need to retrieve, retrieval is how the engine gets current information, and citation is how it credits what it retrieved. For current-information queries, the cutoff sits upstream of the citation surface.

The honest corollary is that retrieval is not automatic. For stable knowledge that sits well within its cutoff, an engine can answer from parametric memory alone, with no search and no sources; Anthropic's same web search documentation describes this case directly, noting that Claude answers without searching when a request draws on stable knowledge¹. When that happens there is nothing to cite, which is one reason some queries return no citations rather than citing anyone. Whether an engine retrieves at all depends on more than the cutoff: query type, the model's uncertainty, product policy, and whether the user asked for sources all feed the decision. The cutoff is one strong input to it, not the whole switch.

How to apply

The cutoff is a model property you do not control, but it tells you where citation opportunity concentrates. The work is to aim content at the queries most likely to trigger retrieval:

• Target fast-moving and post-cutoff topics first. A question the model cannot answer from training (this year's data, a just-shipped feature, an evolving standard) pushes the engine to retrieve, which is exactly where your content can be cited. Stable, long-settled topics are likelier to be answered from memory, but they still draw retrieval when they are specialized, contested, or the user asks for sources, so evergreen content is not excluded, just less reliably triggering.
• Account for per-engine retrieval behavior. Engines differ in how readily they retrieve: Perplexity lists sources on essentially every answer (so the cutoff matters least there), ChatGPT retrieves conditionally, and Claude answers stable-knowledge requests directly without searching¹. The same query can be a citable retrieval event on one engine and a no-citation parametric answer on another, so read a single not-cited per engine, not as a site-wide verdict.
• Keep recency signals honest and current. When an engine retrieves for a time-sensitive query it tends to favor content it can tell is recent and credible, so maintain accurate datePublished / dateModified rather than cosmetic date-bumping (see freshness signals).
• Optimize the retrieval layer, not the model weights. You cannot influence what the next training run absorbs or when the next cutoff lands, and a brand recalled from parametric memory provides no source link or verifiable provenance by itself. The citable surface is the retrieved answer, so the leverage is in being retrievable and quotable.

What to skip: do not treat a model's cutoff date as an optimization lever, and do not chase parametric inclusion. Neither is something you control or can convert into a citation, though the cutoff date is still useful diagnostic context for spotting answers that may be stale. Spend the effort on the retrieval layer that does produce citations.

How it relates to other concepts

• The cutoff is a core reason retrieval-augmented generation exists: RAG (Lewis et al. 2020) pairs a retriever with the model so an answer can draw on current or external content rather than only frozen parametric knowledge³. The cutoff is one problem retrieval addresses; RAG is a common architecture for it.
• It defines part of the job of the retrieval pipeline: fetch, rank, assemble, and attribute exist in part to supply the model with what its training left out.
• It is a major upstream reason for generative engine optimization: because engines retrieve to answer beyond the cutoff, much of the citation surface GEO targets exists in the first place.
• It is documented engine-side in Claude citations: Anthropic ties web search, the knowledge cutoff, and citations together in one tool description, the clearest vendor statement of the cutoff-to-citation chain.
• It raises the stakes for freshness signals: post-cutoff content competes on recency, and most assistants tend to prefer fresher content for time-sensitive queries.
• It is a standing reason for hallucination grounding: answering a post-cutoff question from stale parametric memory is a common way to produce confident but wrong output, and grounding the answer in retrieved sources is the mitigation.