Claude and Claude Code Are Distinct Answer Engines

What happened

Researchers at Profound published an observational study showing that Claude and Claude Code, despite sharing the same underlying Anthropic model, behave like two entirely separate answer engines. Analyzing 24,135 responses gathered from prompts run between July 13 and July 23, 2026, the study found that the two products search at dramatically different rates, surface different brands in response to the same prompts, and send their agents to categorically different types of webpages.

This piece explains what the Profound study measured, how the gap between Claude and Claude Code emerges mechanically, and why the findings matter to anyone trying to make their product visible inside AI-driven answers. It is aimed at developers, content strategists, and marketers who build or optimize for tools that sit on top of Anthropic's models.

Profound analyzed two datasets. The first contained 24,135 Claude and Claude Code responses across 1,724 prompts spanning 11 randomly sampled categories, including a 2,800-response coding subset drawn from 200 prompts. The second dataset tracked the top 1,000 webpages visited by each product's agent over the 30-day window from July 18 to August 18, 2026, using pages scraped via Firecrawl and then categorized by page type using gpt-4.1-mini. Web search was enabled for all platforms throughout the prompt-response collection period.

How it works

The headline finding was stark. When web search was available, Claude searched in over 93 percent of its responses. Claude Code searched in only 13 percent of responses. Despite this difference in retrieval behavior, Claude Code still mentioned an average of 6.6 brands per response, compared to Claude's 5.2, suggesting it draws on training knowledge rather than live retrieval to generate recommendations.

Claude Code is Anthropic's coding harness, estimated to have two to four million weekly active users. Although it runs on the same model intelligence as Claude, it operates with different system instructions, different tools, and a different interface optimized for software development workflows. Those configuration differences shape downstream behavior in measurable ways: mean response length in Claude is 459 words versus 322 words in Claude Code; Claude Code includes a list in 94 percent of responses compared to 56 percent for Claude; and Claude Code includes a table in 54 percent of responses compared to 11 percent for Claude.

Why it matters

The brand-mention divergence follows from the same configuration gap. Claude and Claude Code mention only one in five of the same brands on average when given identical prompts. By contrast, two distinct Claude responses to the same prompt overlap in brand mentions roughly one in two times, and two distinct Claude Code responses overlap about two in five times. Each product is far more self-consistent than it is consistent with the other. In coding prompts specifically, Claude leaned toward code editors and IDEs in its top-15 brand list, while Claude Code more frequently surfaced code-quality and development-workflow tools.

The study introduces a concept worth sitting with: the harness is not a neutral wrapper around a foundation model. It actively shapes which brands get mentioned, which pages get read, and what kind of evidence gets cited. If marketers or developers treat Claude as a reliable proxy for Claude Code visibility, they will likely misjudge both. Profound's framing is that harnesses with meaningful user traffic need to be treated as distinct answer engines in their own right, with separate optimization strategies.

For Claude Code specifically, its agent concentrates almost all of its observed page visits on a narrow slice of the web. Nearly three-quarters of Claude Code's agent traffic across the tracked domains went to documentation, informational, and pricing pages, versus just 5 percent for Claude's agent. Claude's agent spent 60 percent of its visits on robots.txt files, sitemaps, and home pages, the crawl-discovery layer of a site. Claude Code's agent skipped that discovery layer almost entirely and went straight to structured reference material.

Who is affected

The most directly affected group is any company that sells developer tools, APIs, or software infrastructure and wants Claude Code to recommend its product. Since Claude Code's agent specifically targets documentation, informational, and pricing pages, brands that have thin, vague, or structurally opaque documentation are at a disadvantage. Statements like "works with your existing stack" are harder for an answer engine to extract and relay than precise claims such as "Supports Python 3.10 through 3.13, Node.js 20+, and Go 1.22+." The framing of headings also matters: question-shaped headings like "How do I authenticate API requests?" followed immediately by the answer are more extractable than narrative prose.

Marketers working outside developer tooling are affected in a different way. Claude Code's two to four million weekly active users represent a meaningful audience even for non-coding contexts, but that audience is querying through a harness that rarely cites sources and relies more heavily on training-time knowledge. Brands that have not built presence in training data and have not verified their accuracy in structured pages may simply not appear, regardless of how well-optimized their content is for conventional web search or even for Claude itself.

What to watch next

The Profound study covers a specific 30-day window and a specific sample of internally tracked domains, so replication across broader domain sets would strengthen confidence in the ratios it reports. The 13 percent search rate for Claude Code and the three-quarters figure for documentation-and-pricing visits are useful anchors, but both could shift as Anthropic updates Claude Code's system prompt or default tooling. Anyone relying on these numbers for optimization decisions should treat them as a starting point for ongoing measurement rather than fixed constants.

A broader question the study raises but does not resolve is how many other harnesses built on Anthropic's, OpenAI's, or Google's models exhibit similar divergence from their base products. Anthropic's own products are a relatively transparent case because Profound could compare them directly. Third-party harnesses, enterprise deployments with custom system prompts, and embedded AI features in software tools likely show equivalent divergence without leaving comparable data trails. Builders optimizing for AI visibility may eventually need to audit each harness their customers use, not just the foundation models underneath.