The agent successfully assembled a comprehensive explanation of PR Agent by combining fragments from the marketing site's metadata, the GitHub README, and prior knowledge about competitors. The marketing site itself proved largely inaccessible (HTML truncation, no docs routes, JavaScript rendering issues), forcing the agent to rely heavily on the GitHub repository and web search results to answer the task. The agent delivered a satisfying answer despite poor site navigability.

• ›Steps [1] and [4] returned only HTML metadata (title, description, keywords) from the marketing site—no body content was accessible due to truncation or JavaScript rendering. The site meta-description ('AI PR reviews for GitHub pull requests on your servers. MIT-licensed, no per-seat fee, deploy with Docker Compose...') was the only substantive marketing content the agent could retrieve from the domain itself.
• ›Step [7] (GitHub README) was content-bearing and provided the core technical details: architecture (web + worker processes), required infrastructure (Postgres, LLM API), and supported platforms. This became the authoritative source for 'what it does' and 'how it works' because the marketing site could not deliver it.
• ›Step [11] (web search for 'PR Agent GitHub app self-housed reviews pricing') surfaced competitor comparison articles (CodeRabbit, Greptile) that weren't from pr-agent-site.vercel.app itself, yet became critical for answering the 'how it's different' question. The marketing site had no visible comparison section.
• ›The agent cited the marketing site as a source in its final response, but only because it was the starting point and appeared in search results—not because the site's content was actually retrieved or useful. The GitHub repository README and web search results carried the substantive answer.
• ›Site agent-readiness is poor: truncated HTML responses, inaccessible /docs routes (404), no pricing page, no feature comparison table visible in the fetched content, and no embedded examples or screenshots. The agent had to work around these gaps by inferring from metadata and external sources.