The agent successfully assembled a comprehensive explanation of xitie.co by relying almost entirely on two content-bearing fetches: the llms.txt artifact (step [17]) which contained a structured overview of the product, pricing, and features, and a blog post comparing digital invitation services (step [20]). The site is well-documented for machine consumption via the llms.txt file, but the HTML pages themselves were truncated during fetches, forcing the agent to depend on this intentionally published metadata rather than rendering the main marketing pages.

• ›Step [17] (llms.txt) returned a complete product summary including pricing tiers (free vs NT$1,280), feature breakdown, and positioning—this was the single most valuable source and was explicitly written for AI agent consumption.
• ›Step [20] (blog post on digital invitation recommendations) provided competitive context and pricing comparisons (NT$599–NT$4,799 range) that grounded the differentiation claims in the final response.
• ›Steps [1], [3], [6], [9], [12], [14] all returned HTML that was truncated or compressed (4400+ chars dropped), making the rendered marketing pages unusable for the agent; the site appears to compress HTML responses in a way that breaks full-page extraction.
• ›Web searches (steps [5], [8], [11], [15], [18]) returned mostly irrelevant results (competing products, unrelated companies, App Store listings) and were not used as content-bearing sources—they served as routers only.
• ›The site appears deliberately designed for AI consumption (llms.txt is present and comprehensive) but the main HTML rendering layer is broken or intentionally compressed for bots, creating a reliance on the fallback artifact.
• ›The agent correctly identified what it could not verify: billing model details (one-time vs. subscription, retention after lapse), language support scope, AR trigger mechanics, and support tier rationale—these gaps were explicitly called out rather than filled with guesses.