Only 20% of brand mentions overlap in Claude vs Claude Code

Profound gave Claude and Claude Code the same prompts with web search switched on. Claude searched in more than 93% of its responses and Claude Code in only 13%, Profound said in an analysis published on August 31. The two products run on the same models, yet on average they named just 1 in 5 of the same brands for the same prompt.

Profound helps brands show up in AI answers, and the study draws on two datasets. The first holds 24,135 responses to 1,724 prompts from 11 randomly sampled categories, including a coding subset, run from July 13 to July 23 with web search enabled on every platform. The second covers the top 1,000 pages each product’s agent visited between July 18 and August 18 on domains Profound tracks internally. Profound scraped those pages with Firecrawl and had gpt-4.1-mini, a small OpenAI model, sort them by page type. It estimates that Claude Code now has 2–4 million weekly active users.

Same models, different harness

Profound’s explanation is that Claude Code has the same intelligence as Claude but different instructions, tools and interfaces. Together these make up the harness, the layer of prompts, tools and control loop wrapped around a model. A recent paper on arXiv names the practice of shaping that layer “harness engineering”, describing it as “broadening the focus from a model’s context to the full set of mechanisms configured around it.” Profound draws a similar line from its own data: “Foundation models may not be good proxies for the harnesses that they are part of.”

How often does each one search?

Profound calls the search gap the clearest difference: 13% of responses for Claude Code against 93% for Claude, with web search enabled on both. Claude Code still mentioned a similar number of brands per response, 6.6 against Claude’s 5.2.

Each platform was more consistent with itself than with the other. Two separate Claude responses to the same prompt shared about 1 in 2 of the brands they mentioned, and two Claude Code responses shared 2 in 5. Across the two products the figure fell to 1 in 5. The same pattern showed up in wording: repeated responses from one platform were more semantically similar to each other than Claude and Claude Code responses to the same prompt.

Profound then looked at which brands came up in the coding subset of 2,800 responses across 200 prompts. Among the top 15 brands, Claude leaned toward code editors and IDEs. Claude Code more often surfaced code-quality and development-workflow tools.

The answers also look different on the page. Claude Code’s responses were shorter and more structured:

• Mean response length: 459 words for Claude, 322 for Claude Code
• Responses with a list: 56% for Claude, 94% for Claude Code
• Responses with a table: 11% for Claude, 54% for Claude Code

Docs versus robots.txt

Claude Code rarely searches and rarely cites sources, but its agent still produced substantial traffic across the domains Profound tracks, and it went to different places. Nearly three-quarters of the observed Claude Code visits landed on documentation, informational and pricing pages. Only 5% of Claude’s agent visits went to those pages. Claude’s agent instead sent 60% of its visits to robots.txt files, sitemaps and home pages, against 4% for Claude Code’s.

Both agents visited brand-owned content more than earned media or social pages. Profound reads the split as two different jobs. Claude’s agent tries to work out what a site contains, while Claude Code’s goes after specific information on page types it already knows. For Claude Code, Profound concludes, broad site visibility may matter less than making specific technical and pricing details easy to find.

What Profound tells marketers

Profound’s main advice is to treat Claude and Claude Code as separate answer engines, because what works for one may not reliably work for the other. As more harnesses attract meaningful user traffic, it says, marketers may need to measure each one on its own. For Claude Code, that means keeping documentation, informational and pricing pages current and machine-readable, so the agent can pull the exact details it needs to evaluate and use a product.

Concretely, Profound says pages should state clear, extractable facts about uptime, latency, environment support and similar details. Its example of a good claim is “Supports Python 3.10–3.13, Node.js 20+, and Go 1.22+”, which an answer engine can retrieve more easily than “works with your existing stack.” Headings should be written as questions, such as “How do I authenticate API requests?”, and the answer should come before any background or explanation.