LLM

A large language model is a system trained on vast amounts of text to predict what words should come next. Everything an assistant appears to do — explaining, comparing, recommending — is built on that single mechanism.

Two consequences follow, and both matter commercially. First, the model has no database of facts to look up; it has patterns. When it names your brand from training alone, it is reproducing an association it formed from how often and how consistently you appeared across the text it learned from. Second, the output is probabilistic — ask the same question twice and you can get two different answers, which is why visibility has to be measured as a rate rather than checked once.

Modern assistants soften the first limitation by retrieving live pages at question time and reasoning over those instead. That is the part you can influence directly, and it is where most AEO work actually operates.

Understanding that a model predicts rather than looks up explains why brand consistency matters so much. A company described the same way across many sources builds a strong pattern the model can reproduce confidently. One described five different ways produces a hedge — or a competitor's name instead.

Asked which tools lead a category, a model with no retrieval names the three brands that appeared most consistently in its training text — which may be two years out of date. Given live search, the same model returns a current list built from pages it just read. Same system, entirely different answer.