People also ask
In practice
Every answer is a blend, and the blend’s proportions differ per model, per query and per whether the engine has web search enabled. Training-data presence: the brand appears in the corpus the model learned from — established through crawl inclusion (Common Crawl and licensed scrapes), coverage and mentions accumulated over years; its virtues are memory (the model recommends you unprompted, search off) and persistence — its vices are staleness and opacity (the model may describe last year’s product with this year’s confidence).
Retrieval: the engine fetches current pages when composing — which makes freshness, indexation and crawler access the levers, and makes the answer layer responsive to the placement and content work within update cycles. The programme treats them as separate lines with separate tests: the search-off prompt run reveals memory (“secured” status — worth knowing, not manipulable on demand); the search-on run reveals retrieval — and the two disagree often enough that both must be measured.
The strategic implications: a brand absent from training data competes entirely on retrieval — winnable, since most engines’ live layer indexes the same web the classic index does; a brand present in training data enjoys a floor of unprompted recognition that placements cannot buy — and must still win retrieval, because memory recommends what it remembers, not what you shipped last sprint. The reporting frame: per model, per intent — presence from memory, presence from retrieval, and the overlap — because the two layers move on different clocks and neither substitutes for the other.
See also: AI search visibility monitoring, Share of model, LLM optimisation.
Related service: Brand mentions.
