Terminal Funnel: Agentes IA como Canal de Vendas
tomatefarcie123/terminal-funnel
Prove que agentes IA são canal de vendas legítimo com dados reais
Ver no GitHubAgent que demonstra como agentes de código Claude (e outras IAs) se tornaram novo canal de descoberta para ferramentas dev. Incluir case study com 163 dias de tráfego real, simulador de funil e análise honesta do que funciona (e não funciona) em marketing para IA.
Instalação
npx convez-ai tomatefarcie123/terminal-funnel1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Entender como agentes Claude Code descobrem e usam seu produto
- Testar estratégia de conteúdo otimizado para consumo por IA
- Analisar diferenças entre tráfego de humans vs agentes IA
Veredito

Case study raro com dados concretos (40k logs reais) que desafia suposições sobre marketing para IA. Útil para quem vende dev tools, mas honesto sobre as limitações (fetch ≠ citação ≠ conversão). O simulador de funil expõe bem as incertezas.
README
A tradução para o português deste README ainda está na fila.
The Terminal Is the Funnel
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.
This is a small, interactive, open-source web app that makes one argument and lets you stress-test it:
AI coding agents (Claude Code above all) are a new discovery & sales channel for developer tools — and a content site built for AI consumption wins that channel in ways "just improve your website" can't.
It's backed by 163 days of real CloudFront traffic (Jan 1 – Jun 12 2026) for one live developer tool — genymotion.com, an Android-emulator company — served through Rozz's AI site at rozz.genymotion.com.
▶ Live: https://rozz.site/terminal/
What the data actually shows
Parsed from 40,344 CloudFront log files. No estimates.
| Measured | Value |
|---|---|
| Live AI fetches (citation bots, during real user chats) | 15,221 |
| ChatGPT-User fetches | 15,134 |
| Reconstructed AI sessions (avg 2.1 pages, 38% multi-turn) | 7,260 |
| Claude-User fetches | 90 (80 with HTTP 200) |
| …of which from Claude Code (the terminal channel) | 74 (82%) |
| Distinct Claude Code-era sessions | 50 |
| Perplexity-User fetches | 7 |
Top of demand (what AI fetched most): the homepage, then the mobile-testing-security topic, then the pricing Q&A, then the Burp Suite security tutorial — i.e. buyer questions. The on-site chatbot, by contrast, is ~95% hobbyists (run TikTok on PC, set language to Chinese, play Minecraft). The hobbyists are on the website; the buyers are in the AI.
The hero example is a real session: on Mar 27 2026, 18:30:56→18:31:06 UTC, Claude Code fetched the pricing Q&A, then the gmsaas CLI runbook — pricing question to setup commands in ten seconds, inside the terminal.
What you can do in the app
- Run the agent yourself — pick a developer question and watch the same
WebFetchhit the marketing site (lossy, dead-end) vs the AI site (clean answer → runbook → command). - Read the 163 days — citation-channel growth, the crawl→citation lag (Claude's was shortest), the audience shift toward buyers, top content, and every Claude Code session request-by-request.
- The Channel Simulator — an explicit funnel where every multiplier is a knob you set. Defaults are anchored to observed Genymotion values (e.g. the citation rate that rose 14% → 83% after the AI site) and labelled measured vs modeled.
- Skeptic's corner — the five best objections (incl. "Google says don't build a separate AI site" and "fetched ≠ cited"), answered honestly, with the supporting and contradicting data.
What this is — and isn't
It measures what AI systems chose to fetch, and the shape of those sessions. It does not prove the AI cited or recommended the tool (fetched ≠ cited ≠ converted), the user's exact query, or revenue. The calculator is a model layered on measured retrievals, with every assumption exposed. See the in-app methodology section and data/README.md.
Honesty, on purpose (for a developer crowd)
The app deliberately features what didn't work: the AI-native llms.txt discovery file is barely used (crawlers still rely on sitemap.xml; independent tests put llms.txt usage near 0.1%), and the rewritten content pages mostly sit unread — Q&A pages drew far more citation fetches than cleaned pages (5,300+ vs ~3,100 in the data). Building a parallel AI site does not replace good website hygiene — Google's "one web, helpful content, structured data" guidance is correct and is table stakes.
Run it locally
No build step — vanilla HTML/CSS/ES-module JS with hand-rolled SVG charts.
git clone https://github.com/tomatefarcie123/terminal-funnel
cd terminal-funnel
python3 -m http.server 8137
# open http://localhost:8137
Repo layout
index.html # the page
css/style.css # terminal-noir design system
js/ # data loader, charts, terminal sim, calculator, content
data/ # the real (aggregated, non-PII) genymotion dataset + README
docs/ # strategy/requirements + research notes
Privacy & data ethics
Only bot / AI-agent traffic and aggregate counts are published. Raw IPs are never exported; sessions were reconstructed with a salted hash of /24 subnet + edge that is discarded and never published. CloudFront edge locations (coarse, city-level) are kept because they tell the "buyers around the world" story without identifying anyone.
Built as a hackathon experiment by Rozz — AI-search infrastructure for products that want to be found, read, and recommended by AI. MIT licensed.