Here's the answer: assistants assemble recommendations from three pipes. Training knowledge, what the model absorbed about your category over years of web consensus. Retrieval and live search, what it looks up right now. And partner data, the structured feeds like reviews and shopping integrations. That means you influence answers on two clocks: fresh evidence can move mentions in weeks, while the trained consensus moves in quarters.
Now the number that surprises everyone: AI answers are not Google rankings read aloud. Ahrefs found only about 12 percent of AI-cited URLs rank in Google's top 10 for the same prompt. Ranking well and being cited are correlated cousins, not the same sport, which is why we already do SEO is the most expensive sentence in the answer economy.
What actually drives citations, per the big studies: mentions. Brands in the top quartile for web mentions get roughly 10 times more AI visibility. Muck Rack's analysis of 25 million links puts earned media at 84 percent of AI citations in its pool, while Profound's 11.8 billion citation study finds about 57 percent going to brand-owned sites. The studies disagree because they measure different pools, and the operator's conclusion sits in the middle: you need both. Owned evidence answers the factual questions. Earned evidence decides the trust-heavy, best-of answers where purchases happen.
One bonus while it lasts: AI-referred visitors convert at extraordinary rates, because the answer already did the selling.
The full machine room tour is Volume 2 of our Primacy Playbook.
The machine has no mystery left.