The three conditions
Every tactic worth doing maps onto one of these. Anything that maps onto none of them is decoration.
- Retrieval: the assistant has to be able to fetch and parse a page about you, which is ordinary technical accessibility work
- Consensus: several sources the model trusts have to describe you the same way, which is why third-party accuracy matters more than your own adjectives
- Match: the description has to fit the question, so a page about what you do for one kind of buyer wins that buyer's question and no other
- Outside your control: training data, model updates, which sources a given assistant favours this month
- Also outside your control: whether a given phrasing triggers live search at all, which varies by model and by question