What changed, precisely
A search engine returned ten positions and let the buyer choose. A generative engine returns one paragraph naming two or three companies. The competitive question stopped being where you rank and became whether you are in the sentence at all.
That single change has an uncomfortable consequence: there is no page two to fight for, and no gradual slide. A company is either named in the paragraph or it is not, and when it is not, its absence is total.
How an engine assembles an answer
With live search enabled, the sequence is roughly consistent across the four major engines: the prompt is expanded into several retrieval queries, a set of documents is fetched and ranked for relevance, a subset is passed into the context window, and the answer is generated from that subset with citations attached to the passages used.
Two properties of that pipeline matter commercially. First, the retrieval step, not the model's memory, decides most brand mentions on current-state questions. Second, the context window is small relative to the web, so being the eleventh-best source is functionally the same as being absent.
Insight
Citation sets are concentrated and unstable at the same time. Profound measured citations per answer falling 10.4% and unique cited domains falling in 56 of 58 industries after a single model update in June 2026. Concentration is the norm, which is why source-level work outperforms broad content programmes — and volatility is the norm too, which is why depending on one source is a position you can lose in a week.
What carries over from SEO, and what does not
The pattern is that anything which makes a document easy to fetch and interpret still matters, while anything built to game a ranking formula does not survive the move to retrieval plus generation.
The four levers that move a mention
Measuring it without fooling yourself
Engines are non-deterministic, so a single flattering answer proves nothing. What holds up is a fixed question set, fielded the same way on each pass, scored on mention and citation behaviour rather than on the wording of the reply.
That is the whole basis of the Signal Score: twenty buying-intent questions, four engines, eight signals, read against the competitors in the same category. Re-run at day 45, the comparison means something. Run once, in isolation, it is still just an anecdote with a number attached.