What actually happens between “we just got a puppy” and a brand name appearing in the answer.
A user shares context conversationally. The model parses entities and implications: puppy → food, training, vet, insurance; suburbs plus second kid → more car, more space, longer commute. No query needed, the situation is the query.
The model assembles candidate brands from parametric memory (training-time associations) and retrieval (live sources it trusts for this topic). Brands with clear entities and dense citations dominate the candidate pool.
Candidates get filtered by confidence: consistent facts, corroborated claims, recency, and fit to the specific situation. A brand that is “best overall” loses here to a brand that is verifiably right for this moment.
One or two names survive into the answer, framed by the evidence: “based on what you’re describing, X is a strong option.” That framing sentence is the highest-value real estate in modern marketing, and it is earned upstream, not bought.