Astra Dominates New AEO Tracker, While Anthropic Prefers Broader Research

Latent Space has launched a new Frontier AEO Tracker that maps how top AI models recommend tools and services across 161 categories. The methodology involved billions of tokens and 6 prompt variations across 7 different models with active web search. The goal is to provide founders and developers with hard data on how their products rank in the eyes of generative AI when users ask for recommendations in areas ranging from coding agents to payroll software.

Fewer sources lead to higher assertiveness

While Anthropic's models tend to search more sources (Fable with a median of 15 sources, Opus 11), Astra models prefer efficiency and rely on fewer sources (Astra median 5, Sol 9). The test shows that Astra is much more assertive and less prone to changing its answer when the prompt is slightly modified. For companies, this means that AEO optimization is even more critical for more confident models, as random answer variations decrease.

Strong bias in developer tools

The tracker revealed interesting soft biases. In the coding agent category, for example, Fable and Opus logically prefer Claude Code, Sol and Astra recommend Codex, Grok leans towards Cursor, and SWE-1.7 highlights Devin. On the other hand, OpenAI's GPT models showed commendable neutrality, sometimes recommending Claude themselves. Out of 161 categories, 28 have a clear and universal favorite across all models, while the rest remain a highly competitive battleground for AEO.

Those who do not read formatted content lose

The analysis confirmed that models do evaluate and consume content based on its structure. Practices like markdown content negotiation, measured by services like Ora or Vercel, have a real impact. If models cannot technically read your website due to errors or poor formatting, they logically won't cite or recommend you, giving a clear signal for the technical maintenance of corporate blogs and documentation.

Loss of diversity in recommendations

Overly confident models can limit the variety of tools users learn about. If startups fail to get among the top recommended options in these highly assertive systems, they will be cut off from an ever-growing share of traffic and potential customers.

Lilith's verdict

Teaching AI to say your name is now more valuable than buying AdWords. Those who don't optimize for machines will cease to exist for humans.