Claude and Claude Code Read Different Pages. And Name Different Brands.

The AI-visibility firm Profound ran 1,724 prompts through two products built by the same company on the same underlying models, then compared 24,135 answers. Published August 24, the analysis set Claude, the assistant people chat with, against Claude Code, the version that runs as an agent. The prompts ran between July 13 and 23.

They behaved almost nothing alike.

Claude searched the web in more than 93% of its responses. Claude Code searched in 13%. Where each one named companies, Profound found they “only mention 1 in 5 of the same brands.” Ask Claude the same question twice and the overlap is 1 in 2.

Two products. One vendor. Four fifths of the brand recommendations don’t match.

They aren’t reading the same internet

The page-visit data is the part that should change somebody’s content plan. Profound tracked the top 1,000 pages each one visited between July 18 and August 18.

Claude Code sent 73% of its visits to documentation and pricing pages. Claude sent 5% there. Running the other direction, Claude spent 60% of its visits on robots.txt files, sitemaps and homepages, the things a crawler reads to orient itself. Claude Code spent 4%.

The shape of the answers differs too. Claude Code averaged 322 words and used a list in 94% of responses. Claude averaged 459 words and used lists in 56%.

Profound’s recommendation is unusually plain for a vendor post:

“Treat Claude and Claude Code as separate Answer Engines. What works for one may not reliably work for the other.”

And the specific version, for anyone selling to technical buyers:

“If your customers use Claude Code, focus on keeping documentation, informational, and pricing pages current and machine-readable.”

This is not a small slice of traffic

Documentation has quietly become a primary destination for machines rather than people.

Mintlify hosts documentation for software companies. It analyzed 790 million requests across its customers’ sites over 30 days. AI coding agents accounted for 45.3% of them, “which is nearly tied with traditional browser traffic at 45.8%.” Claude Code alone generated 199.4 million of them. As Mintlify put it: “Claude Code, on its own, generated 199.4 million requests. That’s more than Chrome on Windows (119.4M).”

Together, Claude Code and Cursor made up 95.6% of identified agent traffic.

Meanwhile the audience keeps growing. SE Ranking studied 101,574 websites across 250 countries. It found Claude’s referral traffic up 386% between January and April, Claude Code’s weekly active users doubled since January 1, and its business subscriptions up fourfold. Yulia Deda, the SEO and content marketing expert who ran that research, summarized the shift in one line:

“Claude is becoming less like a chatbot and more like a workspace.”

A workspace reads different things than a chat window does.

The part every one of these companies has an interest in

Now the caveats, because they carry real weight and none of these posts lead with them.

Profound sells answer-engine visibility monitoring. Mintlify sells documentation hosting. SE Ranking sells SEO software. Every measurement in this article was produced by a company whose product the measurement argues you need. That doesn’t make the numbers wrong. It does mean nobody in this dataset is disinterested, and it’s the first thing to say out loud before a budget moves.

The sampling deserves the same treatment. Mintlify’s 45.3% is 45.3% of traffic to documentation sites, exactly where you’d expect coding agents to concentrate. Read as a claim about the whole web, it’s badly wrong. Read as a claim about your docs, it’s the most useful number here.

Mintlify at least discloses its own limit, which is more than most: “OpenAI’s Codex doesn’t include one, which means its traffic can’t be distinguished from generic HTTP clients,” so “the actual share of AI agent traffic is likely higher than what we’re reporting here.”

Profound’s post carries no methodology caveats at all. No sampling discussion, no representativeness note, no significance thresholds, on a 10-day prompt window. Register that absence next to a finding as strong as four-fifths brand divergence.

What this changes

For most of the last year, “optimizing for AI search” has been discussed as one job. Our own coverage has largely treated it that way. That includes the Reddit citation collapse we reported on August 20, where two trackers measuring the same thing produced different answers, and the 200-character listing we described as the unit of AI search visibility.

The Profound data points at a second problem underneath that one. Even when trackers agree, they may be measuring a machine that isn’t the one reaching your buyer.

Three things follow:

1. Find out which one your buyers actually use. A company selling developer tools and a company selling to CMOs are optimizing for different readers who happen to share a brand name.
2. Pricing and documentation pages are now brand surfaces. If 73% of an agent’s visits land there, the page your marketing team never touches is doing your positioning.
3. Ask any visibility vendor which harnesses they measure. A harness is the software wrapper that turns a model into a working agent. Profound’s own conclusion is that the list keeps growing: “as more harnesses attract meaningful user traffic, marketers may need to measure them as distinct Answer Engines.”

The uncomfortable version is simpler. If you’ve been buying one AI-visibility report and reading it as the answer, you’ve been looking at one of at least two machines, and quite possibly not the one your customers open.