Why it is not a search engine

A search engine holds documents and hands one back. A chat model holds neither. The retrieval literature states the difference plainly: large pre-trained language models "store factual knowledge in their parameters",¹ the parameters a training run produces rather than any stored document. The same paper reports the cost: such a model's ability to access and precisely manipulate that knowledge is limited, and two problems stay unsolved for a model working from parameters alone, providing provenance for its decisions and updating its world knowledge.¹ Those are the two complaints a reader arrives with. There is no citation because there is no stored document to cite, and a fact cannot be corrected the way a page can be edited.

The literature names the two kinds of memory this page needs. A parametric memory is the trained model itself. A non-parametric memory is an external index that a retriever searches.¹ Retrieval is the second kind attached to the first, and the paper measures the effect: the generated language becomes more specific, more varied and more factual than a parametric-only baseline.¹

Product documentation describes the same shape from the other side. Uploading files for a model to search is framed as augmenting the model's inherent knowledge by giving it access to those collections, and only content actually retrieved comes back with an attribution naming the file.² Web search is a tool a developer adds to a request, after which the model decides when to use it.³ Anthropic documents the criteria: it searches when a request depends on information that is current, changing or outside its training data, and answers directly for stable knowledge such as established facts and coding concepts.³ OpenAI documents its equivalent as a tool the model may or may not invoke depending on the input.⁴ Anthropic also documents that search can be switched off for a whole organisation, which is what an optional capability looks like.³

The boundary. None of this tells you whether the app in front of you has search turned on. In the interfaces where the behaviour is documented it is an add-on that somebody enables, and within a single answer the share that came from parameters against the share that came from retrieved text is not something a vendor publishes. Interpretability work supplies the reason a lookup cannot happen at all: what is stored is spread across many neurons, each neuron serving many concepts, so there is no record sitting there to fetch.⁵ The diagram below answers the same question two ways, one a search engine finding a stored page and the other a chat model with no retrieval turned on, to show why only one of the two paths can hand back a link.

graph LR
  A[Same question] --> B[Search engine]
  A --> C[Chat model, no retrieval]
  B --> D[Finds a stored page]
  C --> E[Generates from parameters]
  D --> F[Hands it back, with a link]
  E --> G[Writes wording, no link]

Figure: The same question resolved two ways; only the search engine's path returns a page to point at.