A developer asked Claude Code how much Genymotion costs. Ten seconds later, the developer asked Claude Code showed the developer how to set it up. Both answers came from the AI site. The pricing page and command line instructions were structured for exactly this: clean, extractable content that Claude Code can fetch and act on in the same session.
The logs show what happened during a session on March 27 hitting the AI site for Genymotion. 18:30:56 UTC — Claude Code fetches /qna/what-pricing-plans-are-available-for-genymotion.html. 18:31:06 UTC — Claude Code fetches /runbooks/gmsaas-cli-runbook.html. No website visit happened. No marketing fluff happened. A developer evaluated a product and started implementing the product inside the same terminal window. The publisher hosts Claude-User on the AI site. There have been 14 requests from 10 unique sessions in six days. 12 of the requests came from Claude Code.
Key findings
- Claude-User: 14 requests (Mar 25–30), first-ever direct Claude citations from the AI site
- 12 of 14 requests came from Claude Code (user-agent:
claude-code/2.1.83-84) - 2 of 14 requests came from claude.ai web search (user-agent:
Claude-User/1.0) - The pricing Q&A page was fetched 6 times from 6 different sessions
- The gmsaas CLI runbook was fetched 3 times
- Time from ClaudeBot mass crawl to first Claude-User retrieval: 5 days
The data
All LLM bots (Mar 24–31, 2026)
| Bot | Requests | Category | Change vs. prior week | |---|---:|---|---| | ChatGPT-User | 1,008 | Citation | Stable (~1,200 plateau) | | ByteSpider | 337 | Training | Up from 91 | | OpenAI SearchBot | 123 | Search index | Stable | | OpenAI GPTBot | 64 | Training | Up from 37 | | ClaudeBot | 33 | Training | Down from 958 (post-crawl monitoring) | | PerplexityBot | 24 | Search index | Stable (post-indexing) | | Claude-User | 14 | Citation | First appearance | | Meta AI | 9 | Training | Stable |
Claude-User content breakdown
| Content type | Requests | % | |---|---:|---:| | Q&A Page | 6 | 42.9% | | Homepage (index) | 3 | 21.4% | | Runbooks | 3 | 21.4% | | GEO Page | 1 | 7.1% | | Robots.txt | 1 | 7.1% |
What Claude Code is
Claude Code is Anthropic’s terminal-based AI coding assistant for reading, editing, debugging, and automating work inside codebases. Claude Code is one of the fastest-rising developer tools in 2026. Estimates put Claude Code at over 20 million daily installs. Some reports say Claude Code exceeded 60% developer adoption within nine months.
The user-agent string in the logs confirms the tool and version. The user-agent string is Claude-User (claude-code/2.1.83) and Claude-User (claude-code/2.1.84). Different developers run different versions of Claude Code. Each version shows up independently in the logs.
Claude Code doesn’t operate like a web chatbot. A ChatGPT-User session happens in a browser tab. The developer reads the answer. Then the developer opens a new tab to visit the docs. Then the developer goes back to the editor. Claude Code collapses that workflow into one step. The answer, the documentation, and the implementation happen inside the terminal.
Using the terminal as a sales channel is, to the publisher’s knowledge, genuinely new and incredibly powerful.
The sessions
Every Claude-User session tells a story.
Session D: 10 seconds (March 27)
| Time (UTC) | Page | User agent |
|---:|---|---|
| 18:30:56 | /qna/what-pricing-plans-are-available-for-genymotion.html | claude-code/2.1.84 |
| 18:31:06 | /runbooks/gmsaas-cli-runbook.html | claude-code/2.1.84 |
Two things made this possible.
First, the content is structured for AI retrieval. The pricing answer lives on a dedicated Q&A page with Schema.org QAPage markup. The markup includes one question and one complete answer. The markup is extractable in a single fetch. But the page never reaches Claude Code directly. Claude Code’s WebFetch tool passes the HTML through Haiku. Haiku is a small fast model that summarizes the content. Haiku enforces a 125-character limit on direct quotes. Haiku can’t reproduce the pricing table verbatim. Haiku paraphrases and condenses pricing content. Haiku sometimes invents helpful framing that wasn’t in the original. Claude Code actually sees a lossy summary, not the page itself. Without the structured Q&A format, that summary would be even more lossy. The summary would contain fragments from a marketing site. The marketing fragments would have no clear answer to extract. With the structured Q&A format, Haiku has a clean signal. Haiku gets one question and one complete answer. Haiku consistently preserves the pricing tiers even through summarization.
Second, the AI site was deliberately built with Claude Code in mind. The gmsaas CLI runbook isn’t a page that existed on Genymotion’s main website. The gmsaas CLI runbook is a step-by-step command reference. Rozz generated the step-by-step command reference specifically for developer tooling sessions. Rozz designed the runbook for sessions where a developer is already in a terminal. The Q&A pages link to relevant runbooks. An AI agent navigating from a pricing question can discover an implementation path. Claude Code went from “how much does it cost?” to “here are the setup commands” in 10 seconds. The content was designed to support that exact transition.
Session C: 12 minutes (March 26)
| Time (UTC) | Page | User agent |
|---:|---|---|
| 17:13:15 | /qna/what-pricing-plans-are-available-for-genymotion.html | claude-code/2.1.84 |
| 17:18:38 | /runbooks/gmsaas-cli-runbook.html | claude-code/2.1.84 |
| 17:25:28 | /index.html | claude-code/2.1.84 |
Session F: claude.ai web search (March 30)
| Time (UTC) | Page | User agent |
|---:|---|---|
| 07:59:59 | /robots.txt | Claude-User/1.0 |
| 08:00:00 | /runbooks/gmsaas-cli-runbook.html | Claude-User/1.0 |
The only non-Claude Code session is the claude.ai web search session. The user agent is Mozilla/5.0 ... compatible; Claude-User/1.0. The claude.ai web search feature uses that user agent. Someone used Claude in a browser. The person asked about Genymotion’s CLI. Claude’s web search fetched the runbook.
Two Claude channels are now open. Claude Code is the developer terminal. claude.ai is the web chat.
What this means for developer tools
The website is no longer the interface. The AI agent is the interface.
For three months, ChatGPT-User has been the citation channel. Someone opens ChatGPT. The person asks about Android emulators. The answer cites Genymotion. This citation behavior is discovery and evaluation. The developer still needs to leave ChatGPT. The developer still needs to visit the website. The developer still needs to read the docs. Then the developer still needs to go back to the editor. Then the developer starts building.
Claude Code removes every step between “I want this” and “I’m building with this.” The developer is already in the codebase. When Claude Code fetches the pricing page, the answer appears in the terminal. When Claude Code fetches the CLI runbook, the setup commands are right there. The next step isn’t “open a browser.” The next step is writing the integration.
Content is no longer read by users on a webpage. Content is consumed by AI tools inside workflows. Workflow output is working code. Workflow output is not a bookmark to visit later.
This is what the AI site was built for. The pricing Q&A page answers the purchase question in structured, extractable format. The gmsaas CLI runbook provides step-by-step commands. The topic taxonomy links pages. Linked pages let an AI agent navigate from evaluation to implementation in a single session. Rozz built infrastructure for Genymotion. The infrastructure includes dedicated Q&A pages, executable runbooks, Schema.org markup, and a topic structure that AI agents can traverse. Claude Code consumed the infrastructure. ChatGPT-User consumed the infrastructure. PerplexityBot consumed the infrastructure.
The Claude pipeline is complete
Three stages exist. Each stage is observable in the logs.
| Stage | Bot | What happens | |---:|---|---| | 1. Crawl + index | ClaudeBot | Content collected, knowledge base updated | | 2. Monitoring | ClaudeBot | Robots.txt + sitemap checks, freshness assessment | | 3. Live retrieval | Claude-User | Real-time fetches during user conversations |
Stage 3 is what the publisher observes now. The full timeline is as follows.
| Date | Event | |---|---| | Mar 2 | ClaudeBot sweeps topic pages (“reading the map”) | | Mar 10–17 | ClaudeBot regresses to monitoring only (0 Q&A pages) | | Mar 20 | ClaudeBot mass crawl: 577 requests triggered by per-topic sitemapindex | | Mar 25 | Claude-User first appearance: pricing Q&A fetched from Claude Code |
Five days pass from mass crawl to live retrieval.
Claude-SearchBot: answered
In article 8, the publisher asked why Claude-SearchBot had never visited the AI site. The answer is now clear. Anthropic doesn’t need a separate search bot. ClaudeBot handles both training and search indexing. Claude-User retrieves content without Claude-SearchBot ever visiting.
| Role | OpenAI | Anthropic | |---|---|---| | Training + indexing | GPTBot + OAI-SearchBot | ClaudeBot (one bot, both functions) | | Live retrieval | ChatGPT-User | Claude-User |
Perplexity: 25% and climbing
Perplexity’s citation rate rose from 17% to 25%. PerplexityBot’s deep indexing event (March 10–15) produces results two weeks later. The crawl-to-citation pattern has now been validated on three platforms.
| Platform | Deep indexing | First citations | Lag | Current citation rate | |---|---|---|---|---| | ChatGPT | Jan 7 (547 req) | Late January | ~3 weeks | 83% | | Perplexity | Mar 10–15 (511 req) | ~Mar 24 | ~2 weeks | 25% (up from 17%) | | Claude | Mar 20 (577 req) | Mar 25 | 5 days | Early (14 requests) |
Where we are
Three months ago, one platform cited Genymotion: ChatGPT at 14%.
Today:
ChatGPT: 83% citation rate. ~1,000 requests/week. Stable. This is the proven channel.
Perplexity: 25% citation rate. Up from 17%. Growing.
Claude: live retrieval confirmed. 14 requests in 6 days. 12 requests came from Claude Code. Developers are evaluating and implementing inside coding sessions.
Same infrastructure exists. Same content exists. Same Schema.org markup exists. Three platforms exist. Three timelines exist. Same outcome exists.
The next question is whether Claude-User follows the same growth curve as ChatGPT-User. ChatGPT went from 42 requests in January to 1,200/week by March. Claude-User is at 14 in the first week. The developer audience using Claude Code may be smaller than ChatGPT’s general audience. The sessions are higher intent. A developer asking about pricing inside a terminal is closer to a purchase decision than someone asking ChatGPT in a browser.
We’re iterating. Stay tuned.
Get This for Your Site
This is what the Genymotion AI site was built for. Structured Q&A pages answer purchase questions. Executable runbooks provide step-by-step commands. Clean, extractable content is linked by topic taxonomy so AI agents can navigate from evaluation to implementation in one session.
Claude Code consumed that content. ChatGPT-User has been consuming that content for months. PerplexityBot indexed all of it. Rozz builds this automatically.
$997/month | ChatGPT at 83%. Perplexity at 25%. Claude Code sessions starting.
→ rozz@rozz.site
Data source: CloudFront access logs for rozz.genymotion.com, March 24–31, 2026. Bot classification is based on User-Agent strings. Claude-User session reconstruction is from timestamped request logs and user-agent version matching.
Serial tech entrepreneur with 10+ years experience building AI systems including Aristotle (conversational AI analytics) and products for eBay and Cartier. Previously founded Squid Solutions. Built AI products like Aristotle, the conversational big data analytics chatbot, and an AR jewelry try-on device for Cartier.
March 31, 2026 | Data period: Mar 24–31, 2026
rozz @ rozz.site | © 2026 ROZZ. All rights reserved.