The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it)
Latent Space Frontier built an AEO tracker analyzing 6 prompt variations across 7 frontier AI models covering 161 categories, revealing how AI agents recommend products and tools Strong self-bias observed: models tend to recommend competitors from the same ecosystem (Claude Code recommended by Claude, Codex by Sol/Astra, Cursor by Grok, Devin by SWE-1.7) 28 out of 161 categories showed universal dominance across all surveyed frontier models, while many others remain competitive battlegrounds New
Analysis
TL;DR
- Latent Space Frontier built an AEO tracker analyzing 6 prompt variations across 7 frontier AI models covering 161 categories, revealing how AI agents recommend products and tools
- Strong self-bias observed: models tend to recommend competitors from the same ecosystem (Claude Code recommended by Claude, Codex by Sol/Astra, Cursor by Grok, Devin by SWE-1.7)
- 28 out of 161 categories showed universal dominance across all surveyed frontier models, while many others remain competitive battlegrounds
- Newer model generations (Astra, Fable) show increased confidence and efficiency — searching fewer sources and changing recommendations less when questions are paraphrased
- AEO practices like markdown content-negotiation were validated as real and impactful; failures in these practices actively discourage models from reading content
Why It Matters
This research provides the first large-scale empirical look at how AI agents influence product discovery and recommendation — a critical factor as AI agents become primary interfaces for users. For AI practitioners and companies, understanding AEO dynamics is now essential for visibility in agent-driven search, and the observed model biases reveal both opportunities and risks in how products get recommended.
Technical Details
- Methodology: Extended AmplifyingAI's framework to run 6 prompt variations across 7 frontier models (Claude Opus/Fable, Gemini Sol/Astra, Grok, Muse, SWE-1.7) covering 161 categories spanning coding agents, AI podcasts, sandboxes, managed databases, ASR models, angel investors, corporate spend, and payroll software
- Scoring system: A proprietary AEO score weighting first choices, alternative choices, and mentions, with negative weights applied to mild and strong anti-recommendations
- Data transparency: Every prompt and answer pair is publicly inspectable; contamination checks were performed and found negative
- Source extraction: Top cited sources influencing agent recommendations were extracted, along with analysis of top recommendation failures
- Limitations: Gemini/Antigravity, GLM/Zcode, and DeepSeek/DeepCode were excluded due to rate limits and errors in the first run
Industry Insight
- AEO is emerging as a legitimate and high-value optimization discipline — as models become more confident and less random in their recommendations, the ROI of proper AEO practices (like markdown content-negotiation) will only increase
- Companies should monitor "soft biases" where models favor same-ecosystem products and invest in AEO strategies that can break through these recommendation loops, especially in the 133 non-universal categories that remain competitive
- The shift toward fewer source lookups and higher confidence in newer model generations means first-mover advantage in AEO will be critical — once a product establishes dominance in a category, it becomes increasingly difficult for competitors to displace it
Disclaimer: The above content is generated by AI and is for reference only.