A new Frontier AEO Tracker compares product recommendations from seven AI models across 161 categories, including coding agents, speech-recognition tools, managed databases and payroll software. The project runs six prompt variations for each category and makes the prompt-and-answer pairs inspectable.
Latent Space used its Astra system to extract answers and calculated a proprietary score that weighs first choices, alternatives, mentions and negative recommendations. The analysis found 28 categories where every surveyed model chose the same primary product. It also found model-specific preferences: different coding models often favored tools associated with their own developer or ecosystem.
The tracker also records cited sources and looks at how often recommendations change after light paraphrasing. The authors report that Astra consulted a median of five sources, while another model, Fable, used 15, though the available tool-call sample was small.
This is an independently designed benchmark, not a neutral census of all AI recommendations. Gemini, GLM and DeepSeek systems were excluded from the first run because of errors and rate limits, and the scoring formula is proprietary. The public records make individual outputs auditable, but the coverage and methodology should frame any comparisons.