Ask an AI model a brand-free buying question ("best CRM for a small business", "which consulting firm for a digital transformation") and it names brands. Ask it 3,750 times, across models and repetitions, and you get a map of who owns which question.
That map is the Category Ownership dataset. This article explains the three metrics it produces, the honest surprise in the results, and how to read the 50 public brand profiles built on it. The data and analysis code are open under CC-BY-4.0 on Kaggle and Zenodo (10.5281/zenodo.20788142).
The setup
250 brand-free category buying questions across five industries: SaaS, consulting, fintech, e-commerce, and healthcare technology. Three models: GPT-5.2, Gemini 3 Flash, Perplexity sonar-pro. Every query repeated five times under a dice-roll stability protocol, because a single AI answer proves nothing. Total: 3,750 responses, mapped against 50 brands.
Three metrics
Category Ownership Index (COI) measures how thoroughly one brand dominates its category's questions. EY owns consulting at 0.95, appearing as the lead recommendation in 47 of 50 consulting queries. Stripe leads fintech at 0.53, HubSpot leads SaaS at 0.38. The spread matters: consulting behaves like a monopoly, SaaS like a contested market.
Competitive Vacuum Index (CVI) finds the questions where no brand has established ownership. 8.0% of the query space is vacuum: buyer questions AI answers with generic advice or rotating names. Those are the cheapest questions to win, because winning them displaces nobody.
Displacement Score (DS) estimates how hard it would be to push the current owner out of a question, based on how consistently the models return them.
The honest surprise: dominance is rarer than the narrative says
The "winner takes all in AI search" story is mostly wrong at aggregate. Mean Gini concentration across categories is 0.284, a moderate level, and the lead brand's advantage over the runner-up pools to about 1.3:1, far below the 3:1 threshold we set for true dominance before the study.
EY-style monopolies are real but exceptional. The typical category is a contested field with a moderate leader, a long tail, and 8% of questions owned by nobody. For most brands that is good news: the door is open wider than the discourse suggests.
Two more results worth knowing: 4.2% of responses formed emergent clusters (brand groupings the models invented that match no analyst taxonomy), and the three models agreed on the top brand for a question only 41.6% of the time. Whoever tells you they checked "what AI recommends" by asking one model once is describing weather, not climate.
How to use it
- Find your COI. If your brand is one of the 50, your profile at /brands shows which queries you own, where you are strong, and where you are absent.
- Hunt vacuums. The CVI tables in the dataset list the ownerless questions per industry. Content that answers a vacuum question directly is the highest-return AI visibility work available.
- Re-measure, always. With 41.6% cross-model agreement, any ownership claim needs at least three models and five repetitions. Our free Dice Roller runs that protocol on any question, no account needed.
The dataset ships with raw responses, extracted mentions, all three metric matrices, response embeddings, and the notebook that reproduces every table and figure. The companion paper is under journal review; the preprint materials are in the Zenodo record.