Claude Opus 5.5 made this entire video in code. 2 prompts. 13 scenes. 2,148 frames. Two voices. No video AI model, no After Effects, just code. And the story is the real trick: Give Claude one MCP server, and it becomes a video crew 👇

The 8 AI engines in our study behave very differently. @grok cited an average of 65.4 sources per answer. @Gemini cited 2.0. @perplexity_ai and Grok cited sources in 100% of answers. Gemini did in 28%. If you only check one assistant, you are measuring that one assistant.

The sources AI cites change with what you sell. For business software, 22.6% of answers with sources cited a review site like G2. For everyday products: 1.6%. Everyday products lean on publications (51.9% vs 30.3%) and YouTube (14.3% vs 6.4%). Same engines, same method, different places to show up. Work out which side you are on before you pick a channel.

What AI engines cite when you ask a "best X for Y" question. Out of 928 answers that cited sources: 37.6% cited a tech or business publication. 15.5% cited a review or comparison site. 10.1% cited Reddit. 9.1% cited YouTube. 92.8% cited at least one blog or list-style page. The most cited domains: - @Forbes (107 answers). - @techradar (97). - toolradar (96). - @tomsguide (95). - @Reddit (94). The page that ranks for "best X for Y" is the page the assistant reads.

I asked 8 AI engines 150 buying questions and logged every source they cited. 1,200 answers. 928 cited at least one source. Reddit showed up in 10.1% of those. The average hides the story. Google AI Overview cited Reddit in 44.3% of its answers with sources. Gemini, 52.4%. ChatGPT, Claude, and Muse Spark: 0%. Those two Google engines produced 89% of every Reddit citation in the study. So whether Reddit is worth your time for AI visibility depends on which assistant your buyers use. Full report below

When a buyer asks @ChatGPT which tool to pick, are you in the answer? @appearsn checks 7 AI engines, shows the sources behind each answer, and helps you show up in them. Reddit first, with every reply approved by you.

Nanobot just dropped a banger for everyone who wants a private AI agent without the bloat! @openclaw is great, but Nanobot is now fully available and 98% smaller while delivering the same core functionality. Most AI agents are heavy, complex, and hard to audit. Nanobot just solved this by staying under small but still acting as a full-featured personal assistant. Here is everything you need to know to build a secure, auditable agent in under 2 minutes: Why Nanobot is a Game Changer Massively Lightweight: 98% smaller than OpenClaw, easy to inspect the entire codebase. Model Agnostic: Use OpenAI, Anthropic, or run 100% local models via Ollama for total privacy. Stateful Memory: It builds a local graph of your history. If you work on a project today, it remembers the context next week. Instant UI: Connects directly to Telegram, WhatsApp, or Discord so you can run terminal commands from your phone. The Power of "Small" by the Numbers 24,000+: GitHub stars and counting. 98%: Reduction in size compared to OpenClaw. 5 Mins: Total setup time from pip install to Telegram bot. 20+: Max tool iterations for complex multi-step tasks. How to Build Your "Research Agent" Install: Run pip install nanobot-ai to get the core engine. Onboard: Use nanobot onboard to auto-create your workspace and "Soul" files. Connect: Add your API keys (or local endpoint) to ~/.nanobot/config.json. Interface: Get a token from BotFather on Telegram to control your machine via chat. Level Up: Add the Brave Search MCP tool to give your agent real-time internet access. The takeaway? AI doesn't need to be a resource hog to be powerful. Nanobot is the perfect "sandbox-first" agent for devs who value privacy and speed. I have shared a link to the full tutorial and GitHub repo in the next post!

This might be controversial, but… AI-generated code could kill your startup. Everyone’s obsessed with how fast AI can ship features. But the faster you move, the less you actually understand your own product. @ArvidKahl breaks it all in the current podcast: