GPT-6 Luna:
AI recommendation benchmark.

Organic asks a buyer question without supplying the target domain and checks whether the business appears in the returned recommendation list. Mentioned asks the same question with the target domain supplied. Both use the model’s existing knowledge, without browsing.

For each completed question pair, subtract the Organic score from the Mentioned score. Average the pairs within each domain, then average across distinct domains. A +39 gap means 39 score points, not a 39% improvement.

A completed Organic result contributes 0 if the target business is absent from the returned list. Unfinished or failed pairs are excluded from both averages. Absence from this list does not establish that the model has never heard of the business.

These tables show each model’s available sample, which may include different businesses and questions. They are descriptive comparisons, not a controlled leaderboard. A fair head-to-head test needs the same domains, questions and test settings.

The benchmark uses completed baseline question pairs and counts each distinct domain once per model. Custom questions and internal/demo projects are excluded. Local fake analyses and results without completed training run records are excluded. When multiple projects track the same domain, only its most recently updated eligible project is included for that model.

Scores are published only when enough distinct businesses have eligible analyses. Only aggregate scores are public; sample counts, customer domains, buyer questions and raw answers are not shown.