Chunking

When AI search generates an answer, it cites the most relevant chunk, not the whole page. Two versions of the same blog post can produce completely different AI quotes depending on how they're chunked. Anthropic and OpenAI engineering blogs report that well-tuned chunking improves RAG retrieval accuracy by 30–50% over baseline. This is where the GEO principle "write in chunks" comes from.

Fixed-size chunking: Splits by a fixed token count like 500 or 1,000. Simple, but breaks mid-sentence and loses context.

Recursive (sentence/paragraph): Splits paragraphs, then sentences, then words, preserving natural boundaries. The default in most RAG pipelines.

Semantic chunking: Uses embedding similarity to detect topic shifts and split there. Highest quality but computationally expensive.

Document-aware chunking: Uses Markdown or HTML ### headings as boundaries. Most effective for structured content like blog posts.

Overlap: Duplicates 10–20% of content across adjacent chunks so context doesn't get lost at the seam.

Sections must stand alone: Chunks typically correspond to ### sections. If a section depends on the previous one to make sense, it breaks when cited in isolation.

Include the subject and context inside each section: Write "Powerblog handles…" not "this tool handles…", each paragraph should be self-contained.

Right length: Very short sections lack enough information to be worth citing; very long sections dilute their embedding meaning. 200–500 words is the sweet spot.

Headings at topic shifts: If a single section mixes topics, chunkers split in awkward places. Add a clear ### heading whenever the topic changes.

FAQ blocks: Q&A pairs naturally form self-contained chunks, so breaking key questions into an FAQ section dramatically raises citation probability.