Claude Searches Web Differently From Claude Code, Data Shows

Profound, a company that tracks brand visibility in AI answers, found that Anthropic's Claude and Claude Code produce distinct responses to the same prompts, with Claude using web search in 93% of responses versus 13% for Claude Code, and brand mentions overlapping by only about 20% on average. The data, based on 1,724 prompts tested between July 13 and July 23, also showed Claude Code's responses were shorter and more structured, averaging 322 words versus Claude's 459, and that Claude Code's agent visits were mostly documentation and pricing pages, while Claude's agent focused on robots.txt, sitemaps, and homepages.

Claude and Claude Code provide different responses to the same prompts, based on new data https://www.tryprofound.com/blog/claude-and-claude-code-are-distinct-answer-engines from Profound, a company that monitors how brands appear in AI answers. In their tests, Claude Code used web search in 13% of responses, while Claude did so in 93%. On average, the brands mentioned for the same prompt overlapped by about 20%. Profound sells AI visibility tracking and promotes its platform in the report. How The Responses Differ Between July 13 and July 23, the company tested 1,724 prompts on both products, generating 24,135 responses with web search enabled for both. Interestingly, repeat runs from the same product were more consistent than the two products were with each other. For example, two responses from Claude to the same prompt shared about half of their brands, while two responses from Claude Code shared about 40%. Despite searching less often, Claude Code mentioned nearly the same number of brands per response, 6.6 compared to Claude’s 5.2. For coding prompts, Claude leaned toward code editors and IDEs, whereas Claude Code more often surfaced code-quality and workflow tools. Additionally, Claude Code’s responses tended to be shorter and more structured, averaging 322 words versus Claude’s 459. Over half of Claude Code’s responses included a table, compared to 11% for Claude. What Each Agent Reads The second dataset shows the top 1,000 pages visited by each product’s agent between July 18 and August 18, across many domains the company tracks internally. Almost three-quarters of Claude Code’s observed visits went to documentation, informational, and pricing pages, while only 5% of Claude’s visits did the same. In contrast, 60% of Claude’s visits were to robots.txt files, sitemaps, and homepages, compared to just 4% for Claude Code. The company interprets this as Claude’s agent exploring what a site contains, whereas Claude Code’s agent goes after specific details on known page types. For such pages, the report’s advice is to state clear facts like “Supports Python 3.10–3.13, Node.js 20+, and Go 1.22+,” and lead with the answer under question-shaped headings. The company mentions that an AI model labeled the page types, but the report doesn’t specify any human review or explain how they distinguished between the two agents’ traffic. Why This Matters The data suggests measuring Claude Code separately from Claude when your customers are using it. In one dataset, agents displayed different page-visit patterns, and in another, brand mentions overlapped by about 20% on average. The report’s authors suggest clearly listing technical and pricing details on pages visited by Claude Code’s agent, although they don’t explicitly test whether these changes boost brand mentions or visibility. Looking Ahead AI products with shared models might require their own measurement as they attract meaningful user traffic. Claude Code is one example of that type of product. The discussion around page visits is based on one month of agent traffic across the domains Profound monitors internally. Featured Image: Cast Of Thousands/Shutterstock