AI Overview Scoring Simulator

Made by Tyrone Sims, Green Owl Systems

When someone searches for your service, Google's AI answer at the top quotes a few paragraphs from a few websites. Put your page up against your competitors and see whose paragraphs get quoted, and what to change so it picks yours.

Ben Franklin Plumbing vs Roto-Rooter, emergency water heater repair in Lakeland, FL
Your service plus your city works best.
Search the phrase above on Google and copy the links of the top results. Up to 6 pages.
One per line. Google's AI also answers the related questions behind a search. Each line gets its own head-to-head.
Advanced settings
How it works

Google's AI doesn't read your whole page. It breaks every page into paragraphs and runs them through 3 rounds, each one stricter than the last. Round 1 asks "is this on topic?" Round 2 asks "does it cover every part of the search: the service, the problem and the city?" Round 3 reads each paragraph next to the search like a person would and picks the best answers. This tool runs the same 3 rounds with open AI models. The tech details:

  1. Stage 1, broad net. Each page is fetched, stripped of menus, footers and scripts, and cut into passages of about 80 words with their section heading attached. Every passage from every page goes into one pool and gets a dense BGE-M3 embedding. Cosine similarity to the query keeps the top 50.
  2. Stage 2, token-level scoring. Late-interaction MaxSim, the ColBERT method. Each word of the query finds its best-matching phrase in the passage, using BGE-M3 vectors, and the passage scores the average of those best matches. A passage has to cover every part of the query (the service, the problem, the city) to score well here. The top 15 survive.
  3. Stage 3, the final boss. The bge-reranker-base cross-encoder reads the query and each passage together and outputs a relevance probability from 0 to 1. The top passages are the ones an AI Overview would quote. A page's score is its best passage times 1000.

A simulator, not Google. It models how passage rankers work with open models, so use it to compare passages, not to predict exact placements.