The agent successfully assembled a comprehensive explanation of Cohere's business, positioning, pricing, and differentiation, but worked around a fundamentally inaccessible website. The main cohere.com domain is client-side rendered and returns no readable HTML content, forcing the agent to rely entirely on web searches and third-party sources (fueler.io, miracuves.com, metacto.com, analyticsinsight.net, sacra.com, etc.) rather than Cohere's own documentation. The agent did retrieve docs.cohere.com in step [10] but cited external sources for the final answer, indicating the primary site did not yield usable content.

• ›Steps [1–6] attempted to fetch cohere.com directly (homepage, /pricing, /about, /products, /command) but all returned client-side rendered HTML with no actual content visible in the response bodies—just boilerplate head tags and image preloads. This rendered the site effectively invisible to the agent.
• ›The agent pivoted to web search (steps [7–9], [11–12]) and found third-party pricing aggregators, blog posts, and comparison articles that contained the factual claims used in the final response (e.g., Command R+ at $2.50/$10 per 1M tokens, 17,000 active enterprise customers, Coral product details). These searches surfaced the information, not the primary domain.
• ›Step [10] fetched docs.cohere.com and returned structured HTML, but the agent did not cite this in the final sources list, suggesting the documentation either required interactive parsing or was not as useful as the third-party aggregators for the broad overview task.
• ›The agent explicitly called out in its answer that 'the website is client-side rendered' and that pricing/product details 'had to come from third-party aggregators rather than official sources'—a candid acknowledgment of the site's inaccessibility.
• ›All seven sources cited in the final answer are external (third-party blogs, comparison sites, news outlets) except for one docs.cohere.com link. The cohere.com homepage, pricing page, and about page—the most obvious places to find this information—contributed zero citations, indicating they were not machine-readable or content-bearing.