What actually moves an AI answer

The honest state of this field: there is one well-known peer-reviewed study that tested content changes against generative engine visibility, and almost everything else you will read is inference from observation. So it is worth reading that study carefully rather than quoting its headline.

“GEO: Generative Engine Optimization” was published at KDD 2024 by a team from IIT Delhi and Princeton. They built a 10,000-query benchmark, applied nine content treatments to source pages, and measured how much of the generated answer each source accounted for (Aggarwal et al., arXiv:2311.09735).

What worked

Three treatments came out ahead: adding quotations from credible sources, adding relevant statistics, and citing sources. Those “achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric”.

What failed is as useful. Keyword stuffing, the reflex import from classic SEO, “[doesn’t] perform well”. On their Perplexity replication it performed 10% worse than doing nothing at all.

The effect also varied by domain: statistics helped most in law, government and debate topics; quotations in history and society topics; citations in factual and legal ones. There is no single treatment that works everywhere, which is the paper’s own stated conclusion.

The finding nobody quotes

Visibility gains depended heavily on where the page already ranked. Citing sources produced “a substantial 115.1% increase in visibility for websites ranked fifth in SERP, while on average, the visibility of the top-ranked website decreased by 30.3%”.

If you are already the leading source, these techniques can cost you. If you are fifth, they can roughly double your presence. The authors frame this as generative engines democratising visibility for smaller publishers. Either way, your starting position changes which advice applies to you.

The caveats that matter

The primary test engine was the authors’ own harness: the top five Google results fed to gpt-3.5-turbo in 2023. That is not a 2026 commercial assistant. “Visibility” is their constructed metric, not traffic or revenue. And the paper says plainly that “methods may need to adapt over time as GEs evolve, mirroring the evolution of SEO”, and that it never tested what these changes do to your search rankings.

So treat “up to 40%” as a best-case figure on one metric, in one harness, three years ago. The durable takeaway is narrower and still valuable: text that carries evidence gets used more than text that asserts. Quote sources, include figures, attribute claims.

Where the line is

There is a second literature here, on making models cite you by manipulating them rather than informing them. A Berkeley team demonstrated adversarial text that reliably promotes low-ranked products, with attacks that “transfer effectively to state-of-the-art conversational search engines” (Pfrommer et al., arXiv:2406.03589). It works, it is prompt injection, and it is the AI-era equivalent of a link farm: detectable in retrospect, and worth nothing when it is.

Google’s own guidance meanwhile tells site owners to ignore several popular tactics, including rewriting content for AI systems and chunking it into small pieces, and to avoid chasing inauthentic mentions (Google).

The overlap between “what the research supports” and “what platforms endorse” is small but real: write things worth quoting, structure them so they survive retrieval, and measure with a sample big enough to detect a change. Whether the change shows up at all depends on whether your problem was retrieval or the model’s memory of your brand.