Generative Engine Optimization: A New Era for Online Information

Large language models are changing the way we find information. Instead of traditional search results, users now get synthesized answers. This shift has sparked a new trend: Generative Engine Optimization, or GEO. Let's break this down: GEO targets the evidence pool and generation process of these engines.

The Influence Question

As GEO gains traction, two main risks emerge. First, the concentrated influence from low contestability and system sensitivity. With fewer ways to challenge the information synthesized by LLMs, who holds the real power? Second, undisclosed commercial influences may be embedded in these engines, impacting the reasoning behind the answers.

Strip away the marketing and you get a clear view: a system where the lines between information and commercial interests blur. The reality is, this could significantly impact the information we consume daily. Are users aware of the potential biases lurking behind these polished answers?

Academic vs. Industry Practices

A formal GEO pipeline can highlight where optimization takes place. Comparing academic and industry practices uncovers a third risk: blind spots. The visibility and evaluation asymmetries between offline setups and deployed systems present challenges. Academic methods might not align perfectly with real-world deployments, leading to gaps in how these systems are optimized and evaluated.

Here's what the benchmarks actually show: offline evaluations often don't capture the full complexity of live systems. This discrepancy can lead to oversight in system performance and influence.

Governance and Measurement

What can be done? The call for answer-level governance and measurement is growing louder. Stronger contestability, precise disclosure, and black-box auditing of material influence are necessary. But will the industry step up to the plate?

Deployment-aligned metrics are also key for exposure persistence. Users need transparency, and systems should be designed to withstand scrutiny. The architecture matters more than the parameter count ensuring fairness and transparency in information generation.

The numbers tell a different story when you consider the potential impact on user trust and information integrity. In a world where synthesized answers become the norm, ensuring these systems are free from undue influence isn't just a technical challenge, but an ethical one.