What is Inference (in LLMs)?
Inference is the moment a trained model runs: it takes the prompt plus any retrieved context and generates the answer, token by token. Training happens rarely and bakes in knowledge; inference happens on every question and assembles the actual answer a buyer reads.
Why the distinction matters for GEO
You influence training slowly (coverage that enters future datasets) and retrieval quickly (crawlable pages that get pulled in at answer time). Both meet at inference. A brand invisible in training data can still appear in answers through retrieval, and a brand with strong priors can be contradicted by what retrieval finds today.
Inference-time variability
Generation is probabilistic: the same question can produce different answers run to run. That is why credible AI visibility measurement uses repeated runs and trends rather than a single screenshot of one answer.