GEO: Generative Engine Optimization

GEO asks what SEO asks, aimed at language models: when a developer asks an LLM what tool to use, does your product come up, how prominently, and is what the model says about it true. In the Sourcegraph case study that shaped it, the corpus is 322 base prompts across five categories tracing the developer purchase journey (category search, comparison, alternative search, use case, problem framing), expanding to 577 with persona, phrasing, and agent-prompt variants, each mapped to one of 23 expected-outcome scenarios and checked against six known misrepresentations of the product.

Scoring is layered cheap-to-expensive. Presence and structural prominence (first-mention offset, list rank, share of voice against a locked competitor set) are deterministic string-and-offset work, reproducible and far cheaper than judges, which makes them the iteration layer. The semantic layer, an LLM judge scoring sentiment, accuracy, completeness, and competitive framing, is opt-in and paid, run only once the deterministic pass has narrowed where a closer read is worth paying for.

The baseline run asked Claude about Sourcegraph and its competitors and put numbers on the question: across 526 responses, Sourcegraph surfaced at a 61.5% mention rate with a 26.9% share of voice against a locked competitor set where rivals like ripgrep and ack drew their own mentions. The same run counted how often the model’s description tripped one of the six known misrepresentations, so the record is not just whether the tool comes up but whether what the model says about it holds.

The statistics get the same care as the prompts. A power-analysis pass decomposes within- and between-prompt variance into a cluster-adjusted design effect, which sizes how many repetitions a given effect size needs to detect. A regression harness then treats model updates as deployments: it grades a metric pass, warn, or fail only when the change clears both an effect-size threshold and a two-proportion z-test, so a model release that erodes your share of voice surfaces as a failing check instead of a feeling.

The harness itself is product-agnostic. The Sourcegraph numbers above come from a private run; the open-source version (GEO_public) ships against a demo product with a seed corpus and a prompt-generator, and pointing it at a new product is a config file, not a rewrite.