The agent successfully assembled a comprehensive explanation of Mirai Labs by combining web search results with targeted fetches of pages that search discovery surfaced. The site itself is poorly navigable — direct URL guesses and conventional paths (/pricing, /docs, /about, /product/*) returned 404s, and homepage content was truncated by client-side rendering. The agent worked around this by relying on 50% web search to find accessible landing pages (/about-us, /inference-runtime, /conversion-optimization-toolkit, blog post), then fetched those. The final answer is substantially complete (what they do, who it's for, pricing, differentiation, competitors) but built more from search snippets and external sources than from the site's own structure.

• ›Search results (steps 8, 9, 15, 16) were the primary discovery mechanism — they surfaced the existence of /about-us, /inference-runtime, /conversion-optimization-toolkit, and blog content that direct URL guesses missed.
• ›Direct navigation attempts (/pricing, /docs, /about, /product/apple-inference-sdk) all returned 404s, indicating either site restructuring or incomplete routing. The agent had to abandon conventional site exploration.
• ›The homepage and most fetched pages used Next.js with client-side rendering and returned heavily truncated HTML (4400 char truncation noted repeatedly), making it impossible to extract full content from the page bodies. Actual product details, feature lists, and pricing tables were not accessible in the fetched responses.
• ›Pricing information came entirely from external sources (FutureTools, SoftwareWorld, Tracxn reviews cited in search results), not from a native /pricing page or pricing table on the site itself.
• ›The agent cited 7 sources in the final response, but only 4 are from trymirai.com URLs (the others are GitHub, G2, FutureTools, Slashdot). This reflects that 50% of fetch sourcing came from web search and the site's own navigation was unreliable.
• ›The final answer correctly identifies Apple-specific focus, per-active-device pricing ($0.10/month), and key products, but includes caveats about what could not be verified (model catalog, Android support, technical documentation) due to inaccessible pages.