Princeton confirmed last year what any honest strategist already knew. In generative AI engines, authority outperforms content. The paper is GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) and the number that makes half the industry uncomfortable is forty — the percentage gain in visibility when a source carries verifiable authority signals versus one that doesn't.
That's not a marginal difference. It's the gap between existing and not existing in a generated response.
Karl Friston has been describing the same mechanism since 2010, applied to the brain — and it holds equally for any system that processes information: systems don't register, they predict. The brain doesn't wait for incoming data to build a picture of the world. It builds the picture first and adjusts when data contradicts it. Google does the same. It runs a prior model of which sources are trustworthy, and when someone queries something, that model predicts the answer before reading anything. If you're not in the prior, it doesn't matter how much you publish.
What gets you into that prior isn't content. It's citations. Verifiable identifiers — ORCID, DOI, Wikidata, institutional profiles. The coherence between what you claim to be and what independent sources can confirm you are. That's authority in terms that information retrieval systems actually process and weight.
The other number worth naming comes from Nature Communications this year (Wu et al., 2025): between 50% and 90% of LLM-generated responses have real support problems with their citations. Models cite, but not always accurately. The direct consequence is that sources with verifiable structure become scarcer and more valuable as models grow more careful about attribution.
The question I always get is the same: how do I rank better? The answer nobody wants to hear is that ranking isn't a content tactic. It's an architecture decision. Which external nodes recognize you. Which identifiers point toward you. What coherence exists between your declared identity and what the graph can independently verify.
I've seen the same pattern across clients in very different niches. The profile you didn't write → The one who comes to build that architecture doesn't ask what it costs. They ask how soon they can start. The question already changed — and that's exactly what predictive systems, human or algorithmic, detect before anything else.