Casmos: Optimizing for LLM Citations Instead of Rankings

yyyokel.com

1 point·wompapumpum··2 comments

Been down a rabbit hole on how Perplexity, ChatGPT, and AI Overviews actually decide what to cite. Turns out it's nothing like traditional SEO anymore.

The short version: LLMs don't care about your domain authority or backlink profile. They care about whether they can extract a clean answer from your content. Structured stuff (Q&A format, comparison tables, FAQ markup) gets cited way more than 3000-word walls of text. And different models cite completely different sources for the same query—like 35-40% disjoint.

I put together a prompt system for Claude that walks through the whole thing: researching how LLMs behave in your niche, figuring out what's exploitable, building content that's easy for AI to parse and cite, seeding it across platforms these models actually crawl, then deciding what's working and what to kill.

It's five prompts that I run in sequence. Pretty tacticalless "best practices" and more "here's how these systems actually work right now." based on my findings.

Would love to hear if anyone else is seeing meaningful traffic from AI search yet, or if it's still mostly noise...

From a backend perspective, this matches what I've seen building RAG pipelines. The issue is usually that standard chunking algorithms often sever semantic context in long-form text, even with overlap. Structured data or explicit Q&A formats survive retrieval much better because the embedding actually captures a complete thought rather than a fragment.

Reply on news.ycombinator.com