On September 7, 2026, artificial intelligence newsletter and podcast Latent Space, led by Swyx and Alessio, released an empirical study investigating AI Engine Optimization. Titled the Frontier AEO Tracker, the research systematically examines how frontier language models evaluate and recommend software products. As developers and enterprise workflows increasingly delegate technical discovery to autonomous agents, understanding how these systems select tooling has become critical for the software industry. The publication offers one of the first data-driven benchmarks on the mechanisms governing product visibility in modern generative pipelines.
To carry out the analysis, the researchers constructed an automated testing pipeline powered by GPT-6 Astra. The benchmarking environment systematically queried seven leading models, including GPT Sol, GPT Astra, Claude Opus, Claude Fable, Grok, and SWE-1.7. Each model was evaluated using six distinct prompt variations across 161 predefined software categories. This setup allowed Latent Space to observe recommendation stability across diverse prompt formulations and assess how model architectures differ when answering identical inquiries.
A primary finding of the tracker is the presence of a measurable, systematic self-bias across the leading frontier models. Systems consistently favor tools originating from their own corporate or developer ecosystems. For instance, Anthropic's Claude models heavily prioritize recommending Claude Code, whereas OpenAI models consistently favor Codex. Similarly, xAI's Grok predominantly recommends Cursor, highlighting how proprietary alignment or specific training distributions influence recommendation outputs.
Beyond platform bias, the study highlights severe consolidation in several key software categories. In 28 of the 161 examined categories, an uncontested first-place winner has emerged across all tested models, regardless of prompt variations. For emerging startups and alternative tooling, this dynamic presents a formidable barrier to organic discovery. Once an incumbent software tool becomes entrenched across frontier training sets, autonomous agents repeatedly surface it, effectively shutting out competing alternatives.
The report also uncovers the primary technical factors dictating whether autonomous agents cite and recommend a given product. According to the findings from Swyx and Alessio, product inclusion does not rely on traditional web metadata. Instead, the most effective levers are a semantic documentation structure and the implementation of Markdown content negotiation on product landing pages. Web destinations that seamlessly serve raw Markdown to automated agents experience significantly higher recommendation frequencies.
The findings of the Frontier AEO Tracker indicate a structural shift in how software discovery functions in an era dominated by autonomous agents. Traditional search engine optimization focused on human web crawlers is rapidly giving way to optimization tailored specifically for LLM context windows. Software providers must now architect their technical documentation for direct agent consumption if they wish to remain discoverable. As autonomous pipelines increasingly decide which tools get installed, failing to adapt to AEO dynamics risks complete invisibility across modern development stacks.