The GEO score: Quantifying the perfect content playbook

1. Ensuring Same Idea Space: Cohesion to Theme

We used sentence embeddings (OpenAI's text-embedding-3-small) to convert each article title into a high-dimensional vector representation. To ensure all titles live in the same conceptual neighborhood:

• Create a theme anchor by computing the centroid of all title embeddings
• Calculate cosine similarity between each title embedding and the theme anchor
• Compute the mean similarity across all titles

This gives us a value in [0,1] where higher values indicate the titles are thematically consistent: they're all in the same conceptual neighborhood rather than scattered across different domains.

2. Preventing Over-clustering: Top-k Neighbor Analysis

To ensure titles aren't just variations of the same concept, we analyze top-k nearest-neighbor similarity:

• For each title, find its k=3 most similar neighbors
• Calculate Top3_sim as the mean cosine similarity to these 3 neighbors

Using top-3 neighbors rather than single nearest neighbor reduces noise and provides a more stable redundancy measure. Low Top3_sim indicates healthy diversity: titles are related but not duplicative.

3. Composite Score

We use a harmonic mean to force balance between our metrics:

Where:

• C (Cohesion): Measures how well titles align with the theme (0 to 1)
• N (Non-redundancy): 1 - Top3_sim, measures diversity (0 to 1)

Both metrics are normalized to [0,1], making the score directly comparable across test runs. The harmonic mean ensures both metrics must be reasonably high: a high score means titles are cohesively on-theme while avoiding redundancy.

Site Creation Process

For each test site, we followed a standardized methodology:

1. Created small Astro blogs - Fast, lightweight static sites optimized for content delivery
2. Deployed statically on AWS - Used S3 + Amplify.
3. Registered with Google Search Console - To allow chatgpt to consume from the google SERP.
4. Waited for indexing.
5. Used the impaca.ai platform to monitor GEO visibility for all sites.

Group strategy

We created 5 groups of 6 sites each (30 total), where each group uses a unique gibberish word that only appears in that group's titles. This prevents groups from competing with each other or with existing, established sites on AI mentions.