What Princeton proved about getting cited by AI

Why this study matters

Most advice on ranking inside AI answers is guesswork and vendor opinion. This study provides controlled benchmark evidence, although it should not be treated as a universal ranking rule. "GEO: Generative Engine Optimisation", by Aggarwal et al., was published at ACM KDD 2024 — a top-tier machine-learning conference — by researchers from Princeton University, IIT Delhi, Georgia Tech and the Allen Institute for AI. It's the first large-scale academic study to test what content changes actually improve citation rates inside generative engines.

The team built a benchmark called GEO-bench, ran 10,000 real user queries, and measured how nine different content tactics changed a source's visibility inside AI-generated answers. In their words, GEO "can boost visibility by up to 40% in generative engine responses."

The five tactics that worked

Within the benchmark, the stronger-performing interventions included:

• Cite sources — adding inline references to credible sources for your claims. Adding relevant citations helped in parts of the benchmark, especially for lower-ranked source positions. The size of the effect varied by domain and baseline position.
• Add statistics — replacing vague claims with specific numbers, percentages and dates. Statistics were one of the stronger interventions in the benchmark, but the effect varied by query and domain.
• Add quotations — including direct, attributable quotes from credible third parties. Engines extract quotes as evidence.
• Optimise fluency — clear, well-structured, readable writing that an engine can lift cleanly into an answer.
• Authoritative voice — confident, expert phrasing rather than hedged or promotional copy.

The biggest effects came from combining tactics — particularly statistics plus fluency — not from any single trick.

The honest caveat

This is evidence, not a rulebook. The study uses controlled benchmark interventions, but it does not establish durable organic discoverability in current commercial AI systems, and it notes that the effect of each tactic varies by domain, which is why domain-specific optimisation matters. It also can't tell you how fast these signals decay as engines change. The sensible response is to apply the proven tactics and then run your own measurement loop — many prompts, tracked over weeks — rather than trusting anyone who sells certainty about the algorithm.

What this means for your brand

Every tactic in the study is something we build into content at Cited: specific, cited, quotable, well-structured writing backed by a clear entity and real authority. It maps directly onto our four-phase method — Audit how engines cite you now, Structure answer-ready content, build Authority, then Monitor your share of answer. For the fundamentals, start with What is AEO? or What is GEO?

Source: Aggarwal, P. et al. "GEO: Generative Engine Optimisation," ACM KDD 2024. Findings summarised here in our own words; read the original paper.