Assistants and answers
How an assistant chooses which pages to open before it writes — a selection problem, not a ranking one, and it rewards different things.
Also called Source selection · Извлечение источников
A search engine orders every candidate and shows you ten. An assistant picks a handful, reads them, and writes an answer that has to be defensible from what it read. The second job favours a page that states a claim, attributes it, and gives the figure with its period and base — not because a model prefers rigour, but because a page that says something specific is a page the answer can be grounded in.
That is also why extractability is a structural property rather than a writing style. A claim buried in a paragraph of hedging survives no summarisation; a claim in a sentence, with its source beside it, does. The site-level version of the same point: a machine that can fetch your text without executing a page of JavaScript, and can read a table as a table, has more of you to work with.
Where this is genuinely uncertain, it is worth saying so. Nobody outside the operators knows the weighting, the operators change it without announcement, and every confident public account of “how to rank in AI search” is inference from small samples. What can be established from your own side is narrower and more useful: whether the pages were fetched at all, and whether what they contain can be quoted without being rewritten.
The tactics do not carry over. Keyword density, link volume and title optimisation address an ordering problem that is not the one being solved here, and the effort is better spent on whether a specific claim on the page can be lifted with its source attached. The test is mechanical: read one of your pages and try to write two sentences that answer a buyer's question, citing it. If you cannot, neither can the machine.
Knowing the definition is not the same as being able to check the figure. These are the procedures that do the second thing.
- Testing a promise about AI answers
- “We will get you cited by ChatGPT and the other assistants.” · 20 minutes, 5 questions.
The definitions are the easy part. Whether the figure on your dashboard was computed this way is a different question, and usually the more expensive one.