Hallucination
A confident, fluent but factually wrong or fabricated statement produced by a language model, including invented citations, prices…
GEO AITraining cutoff
Knowledge cutoff
The date after which a language model has no training data, so anything more recent is unknown to it unless supplied through retrieval.
In full
Cutoffs explain why an assistant may describe a product's old pricing or a defunct feature: the fact was true in training data and nothing retrieved corrected it. Grounded systems mitigate this, but only when retrieval fires and finds current information. For brands the implication is that keeping authoritative, easily retrievable current facts on the open web is the only practical way to overwrite stale parametric knowledge.
Example
An assistant without search access lists a company's pre-rebrand name because the change happened after its cutoff.
Related terms
Grounding
Constraining a language model's output to information retrieved from a specified source — a search index, a document set or a database —…
Brand mention monitoring
Tracking where a brand is named across the web and inside AI answers, including whether the description is accurate and the sentiment…
Retrieval-augmented generation
An architecture where a system retrieves relevant documents or passages first and then has a language model generate an answer conditioned…
Context window
The maximum amount of text a language model can consider at once, which bounds how much retrieved content can be supplied when generating…