The agent assembled a comprehensive explanation of Anthropic by combining sparse website fetches with web search results and external sources, since the anthropic.com site itself is JavaScript-heavy and lacks centralized, machine-readable product and pricing information. The site's structure—splitting corporate messaging across anthropic.com and claude.com, returning 404s for key pages like /about and /papers, and embedding content in JavaScript bundles—forced heavy reliance on external sources and prior knowledge (36% prior knowledge, 55% from previous artifact fragments) to answer the task.

• ›Steps [1], [2], [3], [4], [6], [9], [10] fetched anthropic.com pages but returned HTML stubs with JavaScript placeholders rather than readable text content about what the company does, pricing, or how it differs from competitors. The actual product positioning ('Claude') and pricing details live at claude.com, not the corporate domain.
• ›The agent relied primarily on web search results [12, 13, 14] and the Claude platform docs [15] to answer the task. Search result snippets and external third-party comparison sites (tech-insider.org, MindStudio, SpenHound) provided the substantive answers about pricing tiers, enterprise market share, Constitutional AI, and competitive positioning—none of which were clearly present in the fetched HTML.
• ›Key pages returned 404 errors: /about [5] and /papers [7] do not exist, blocking direct access to company background and research outputs. This forced the agent to infer company information from Wikipedia and news articles instead of first-party documentation.
• ›The site's navigation structure is fragmented: API documentation is at platform.claude.com (separate subdomain), pricing is at claude.com/pricing (not anthropic.com/pricing despite the fetch), and research presence is opaque. An agent cannot reliably discover the full product story from anthropic.com alone.
• ›The final answer cites 12 external sources and only 3 anthropic.com URLs; the agent explicitly flagged in its response that the website is 'heavily JavaScript-based, making it difficult to extract clear text content' and that 'pricing could be clearer.' This is candid evidence that the site did not self-serve the answer.