A developer asks their AI coding agent a question — mid-session, never opening a browser. The agent fetches a page, reads it, and acts. That fetch is a sales call your analytics will never see. We measured it on 163 days of real traffic for one developer tool, and tested what actually wins it: improving your website or a separate AI site.
reconstructed from a real session in the logs — Mar 27, 2026 · 18:30:56→18:31:06 UTC
For one real Android-emulator company, the on-site chatbot logged what visitors actually want. The AI site logged what AI assistants fetch to answer buyers. They tell completely different stories.
667 conversations
<5% showed commercial intent. The rest:
Hobbyists. Buyer rate stuck at 7–8% for months — and it didn't move.
buyer-intent
up from 37% — buyer/pro reads overtook hobbyist reads. What AI fetched:
Buyers. Sessions that look like procurement — invisible to the website.
Two different datasets, kept separate. The website figures
(667 conversations, intent mix) come from Genymotion's on-site chatbot logs, published in
Rozz's Insights series. The AI-site figures throughout the rest of this
page come from CloudFront access logs for rozz.genymotion.com. We never
mix them.
The hobbyists are on your website.
The buyers are in the AI.
The scenarios you're about to run aren't hypothetical — they're these people's actual questions.
Lives in Claude Code / Cursor. Hits a real problem while coding — "run 20 emulators in CI without a Mac farm." Asks her agent in place; never opens a browser if it can answer.
Wants: a command to run, not a brochure. Evaluate → decide → implement without leaving the terminal.
Tech lead doing procurement research in ChatGPT / Perplexity, not Google. Multi-round sessions that end on the pricing question. The vendor never sees it happen.
Wants: a confident shortlist. If your content isn't retrievable, a competitor's is.
Head of Growth / DevRel at the tool. Markets to enterprises; the website audience is hobbyists; the real buyers evaluate in AI, where analytics shows nothing.
Wants: to see the invisible AI demand — and influence what the assistant recommends.
This is what an AI coding agent's WebFetch actually does: pull a URL, convert to
markdown, summarize through a small fast model, enforce a tight quote limit. Pick a developer's question
and watch the same agent hit the marketing site vs the AI site.
How this is built: the AI-site content is real (pulled from the live Genymotion AI site); the marketing-site extraction reflects what a markdown-then-summarize pass recovers from a page that is ~75% navigation/CTA chrome with no machine-readable Q&A markup. We dramatize extractability and format, not specific prices.
Not our framing — Anthropic's. The official Claude Code docs say WebFetch
"converts the response to Markdown… and runs the prompt against the content using a small, fast model.
For most fetches, Claude receives that model's answer, not the raw page… lossy by design,"
and it sends an Accept header that prefers Markdown over HTML. Clean,
answer-shaped content survives that pipeline; marketing chrome doesn't.
Four reasons the channel is real — and one reason to stay honest about it. Sources in the research notes.
Every figure below is parsed from 40,344 CloudFront log files for the Genymotion AI site, Jan–Jun 2026. No estimates. Methodology & limits →
These are counts of what AI assistants fetched on the Genymotion AI site — not what they said, whether they cited or recommended Genymotion, or whether anyone bought. Fetched ≠ cited ≠ converted. ~70% of AI traffic carries no referrer, so attribution is genuinely hard. We measure retrieval demand; the Channel Simulator models the rest, with every assumption exposed.
Daily fetches by real-time citation bots — the AI reading your content during a user's conversation.
A platform mass-crawls (training/index), then starts fetching live for users. The gap is the lead time on a new channel.
Citation-bot content reads, classified by audience, by month.
Top content fetched by citation bots across the window.
Every Claude-User session on the AI site — the developer evaluations that happened inside a coding tool. — came from Claude Code.
Google is unambiguous: don't build a separate site to game AI — one web, helpful content, structured data. That's correct, and it's table stakes. The AI site isn't a ranking trick (canonicals point home) — it's a different surface for a different reader. Here's the honest comparison.
This table is about the assistant channel — ChatGPT, Claude, Perplexity — which retrieve live and which our logs show fetching the AI site 15,000+ times. Google's AI Overviews / AI Mode are a different system that ranks from Google's own index, and there the answer is the opposite: improve your one site, no separate property, no special markup. Independent studies (Princeton, Ahrefs, Semrush) agree with Google that content and third-party mentions beat parallel sites for that channel. We're not arguing with Google — we're measuring a channel Google doesn't run.
And it isn't free. A separate AI site is real work to stand up and keep in sync: harvesting Q&As from real conversations, rewriting docs to extraction-fluency rules, authoring runbooks, maintaining a topic taxonomy, and keeping canonicals pointed home. Against 50 Claude Code sessions in six months for one tool, that's a bet on the trajectory — not a slam dunk today. Improve your main site first (it's cheaper and Google-endorsed); add the AI surface when the main site plateaus or can't host the buyer long-tail.
Rozz scores any domain's AI-readiness on seven dimensions. Note where the weight sits — extractability + markdown conversion = 38%. Schema.org is only 10%. The game is "can a small model lift your answer cleanly," not "do you have JSON-LD."
The one thing the logs truly measure is retrieval demand. Everything downstream is a model you control. Drag the knobs — defaults are anchored to Genymotion-observed values where noted (the rest are yours), and the whole funnel is explicitly a model, not a measurement.
This is a model, not a measurement. We can count what AI fetched; we cannot see what it said, whether it recommended you over a competitor, or whether anyone bought. Every multiplier above is yours to set or zero out.
If you're a developer, you're already drafting objections. Good. Here are the five best ones, with honest answers.
No purchase required to start — most of this you can do this afternoon with your own logs.
ChatGPT-User, Claude-User, PerplexityBot, OAI-SearchBot, GPTBot, ClaudeBot. You'll likely find demand you never knew you had.Source: CloudFront access logs for the Genymotion AI site, Jan 1 – Jun 12 2026. Raw IPs are never published; sessions are reconstructed with a salted hash of /24 subnet + edge, then discarded. Full method & the raw aggregated JSON are in the public repo. External claims about Google's stance, GEO practice, and Claude Code adoption are cited in the research notes.