A LATENT SPACE RESEARCH PROJECT
We asked frontier AI agents to recommend products across 133 categories. This is an open window into their choices: the favorites, the disagreements, and the reasons behind them.
AEO means answer engine optimization—understanding how products show up in AI answers. Here, you can go beyond a mention count and see what agents actually recommend.
WHAT SURPRISED US
A high mention rate can hide a low chance of being chosen. A familiar favorite can disappear when you change the model. The interesting part is understanding why.
01 / MENTIONED ≠ CHOSEN
Kysely appears in 42 of 42 validated ORM / database-tool answers (100%), but is first choice in only 1 (2.4%). Visibility alone misses the difference between being discussed and being recommended.
02 / THE MODEL CHANGES THE WINNER
For the same AI-sandbox scenario, GPT-5.6 Sol picks Modal first in all 6 frames; GPT-6 Astra picks E2B in all 6. Consistency within a model does not mean agreement across models. These are six phrasings, not independent repeat trials.
03 / READ THE SCORE, THEN THE ROLES
CodeRabbit leads the code-review leaderboard with a 62.9% overall score, while being first choice in 42.9% of validated answers. Alternatives and supporting mentions also contribute; eligible objections subtract. The roles explain what a single percentage cannot.
Findings describe this study’s saved answers. Rates measure observed recommendations, not product quality or market share. Each link leads to the underlying research view.
THINK YOU KNOW WHAT AI WILL PICK?
Start with three familiar AI products. Guess the order, reveal the rankings, and skip anything you don’t know.
MAKE THE RESEARCH YOUR OWN
Browse category leaderboards, compare recommendation strength, and open the answers behind each score.
02 / ProductsExplore mentions versus first choices, model preferences, and the situations where recommendations change.
03 / Compare productsPut two products side by side to inspect shared appearances, tradeoffs, and the evidence behind each choice.
04 / Model differencesCompare matched questions across Sol → Astra and Opus → Fable. Follow the biggest shifts into individual results.
05 / SourcesExplore cited websites and look up a domain’s recorded retrieval successes, failures, and citing answers.
06 / Notable findingsRead category findings, strengths, limitations, and switching conditions grounded in the saved answers.
THE EXPERIMENT, IN PLAIN ENGLISH
We ran real AI agents through their native tools and harnesses, saved their answers, then assessed the products they recommended, mentioned, or objected to.
Read the full methodology ↗From coding agents and AI infrastructure to business software and professional services, each category gives the agent a concrete decision to make.
A baseline plus direct fit, practitioner choice, comparison, tradeoffs, and conditional recommendation. These are related phrasings of one scenario, not independent trials.
5,566 usable answers out of 5,586 saved. Product identities and roles are reviewed against the evidence; 20 excluded answers stay out of score denominators.
Separate first choices, alternatives, supporting mentions, and objections. Average across available configurations and link the results to prompts, answers, and research notes.
FROM A SCORE TO AN EXPLANATION
Start with a category. See who appears. Then inspect why.