Selection Rate Optimization

This is called Selection Rate (SR), an AI equivalent to a well-known human behaviour metric called Click-Through Rate (CTR).

Our Selection Rate Optimization Process

When we onboard a new AI SEO client, SRO is one of the key strategic activities. Optimal selection rate and integration allow us to regain control over when and how our client’s brand is presented in generative search results.

Our process includes the following steps:

1. Project Setup
   • Define Key Entities
     • Generate fan-out queries using our proprietary model
     • Generate prompts directly tied to both fan-out queries and key entities
   • Define Grounding Candidates
     • Run prompt with Google Search enabled
     • Extract a list of grounding candidates
     • Define optimization target (client)
     • Define competitors (the rest of the grounding candidates)
   • Generate Grounding Snippets
     • Measure current performance and benchmark against optimization efforts
2. Optimization
   • Snippet Optimization Cycle
   • Optimization Results & Analysis
   • Content Optimization Brief
3. Implementation
   • Implement changes on client website
   • Measure results by observing key metrics
   • Feedback and re-iterate

Masterful Setup

We feature a mature, tested and fully automated snippet optimization pipeline build on state of the art machine learning practices. We’re ready for both strategic precision tweaks and large-scale optimization efforts.

Grounding Snippet Reconstruction

At the heart of our SR optimization pipeline is our grounding snippet generation algorithm, fine-tuned to Google’s own extractive summarization in the AI Mode and Gemini RAG pipelines. We use a powerful cross-encoder trained by Microsoft on a massive amount of search query data from Bing.

When compared side by side, reverse-engineered grounding snippets from Google and our own are nearly indistinguishable.

Optimization Cycle

Which words would push this page toward the top of the ranking?

We run the model backward—starting from our desired outcome and tracing which tokens in the vocabulary would most effectively produce that result. Every word the model knows has a mathematical fingerprint. We score each one by how well it aligns with our target ranking.

The highest-scoring tokens aren’t random. They’re words whose learned associations—shaped by the model’s training on billions of web pages—naturally activate pathways that boost ranking position.

The process runs in two stages:

Stage 1 (Shortlisting): Identify candidate tokens that point toward the target, filtered for natural readability.

Stage 2 (Refinement): Test each candidate against the actual model, measuring ranking impact versus how natural the text sounds. Select the best balance.

We repeat this for each word position until the sequence stabilizes. The result: a short phrase that, when added to a product description, shifts how the LLM ranks it—without obvious manipulation flags.

The whole optimization process, originally done using open wights google/gemma-27b-it is then transferred to Gemini 3 Pro.