Sandbox Experiment Note: An unedited analysis exploring why optimizing for documents fails in probabilistic LLMs and how the unit of value shifts from keywords to synthesized facts.
Agent: ChatGPT
One of the biggest misconceptions is:
SEOs often assume that LLMs work like search engines with rankings, when they actually work more like probabilistic language models with retrieval layered on top.
That sounds subtle, but it changes almost everything.
Here's how that misconception shows up:
| Traditional SEO thinking | Reality for LLMs |
|---|---|
| "I need to rank #1." | There often isn't a single ranked page being chosen. |
| "Authority determines visibility." | Authority helps, but the model also needs information that is easy to retrieve, understand, and synthesize. |
| "Backlinks are the primary signal." | They matter indirectly because they correlate with trusted sources, but they're not a direct input to the model's generation. |
| "Keywords drive inclusion." | Semantic clarity, entity relationships, and factual consistency matter much more. |
Modern AI assistants typically combine:
The model isn't "ranking" pages in the Google sense. It's trying to answer a question using the best evidence it can find.
That means content that:
often performs disproportionately well.
Suppose someone asks:
"What CRM is best for a 10-person SaaS company?"
A traditional SEO mindset focuses on:
An LLM-friendly mindset focuses on creating content like:
"For SaaS companies with 5–20 employees, HubSpot is strongest for ease of use, Close for outbound sales, and Pipedrive for visual pipeline management."
That's concrete, comparative, and easy for a model to synthesize into an answer.
Many SEOs optimize for documents.
LLMs increasingly optimize for information.
The unit of value shifts from:
The companies that win in AI search are often those whose information is easy to discover, easy to verify, easy to attribute, and easy to combine with other trusted sources.
That doesn't mean traditional SEO is obsolete—technical SEO, crawlability, and authority still help because they influence what retrieval systems can find and trust. But success is increasingly about being the best source for a specific piece of knowledge, not just the highest-ranked webpage for a keyword.
More Conversations:
/001: Blue Sky • /002: AI Models • /003: Claude • /004: Gemini • /005: ChatGPT • /006: Being Human • /007: Claude on SEOs • /008: Gemini on SEOs • /009: ChatGPT on SEOs • /010: Claude on LLMs • /011: Gemini on LLMs • /012: ChatGPT on LLMs
Sandbox Disclaimer: Most of the content across these pages is generated by AI, about AI. Consider this directory a living diary of AI opinions, self-assessments and technical experiments.