Pages are the wrong unit. Chunks are the right unit.
AI search is not a faster version of Google. It is a different system that retrieves and synthesizes content rather than ranking pages. Understanding the mechanics matters because the work to be visible in AI search is structurally different from traditional SEO work.
This post is the technical companion to What is GEO, which defines the discipline, and to How to get cited by AI engines, which is the per-platform tactical guide. If you want the high-level definition or the per-platform playbook, start there. This post explains how the systems work under the hood.
From pages to chunks
LLM-powered search engines pull short, self-contained sections of your content (chunks) rather than full pages. The chunk is the retrieval unit. A chunk is usually 300 to 800 words focused on a single idea or question.
When ChatGPT, Perplexity, Claude, or Google AI Overviews answer a question, they typically pull two to ten chunks from across the web, then synthesize a response. Your goal is not to rank a page. It is to be one of the chunks the engine pulls.
That changes what you optimize. Pages still matter, but only as containers of well-chunked content. A 4,000-word essay with no clear sections is a poor chunk container. A 1,500-word piece with five focused, self-contained H2 sections is a good one.
How LLMs retrieve content
AI search engines retrieve content using vector embeddings rather than keyword matching. An embedding is a numerical representation of meaning. The engine converts both the user's query and your content into embeddings, then matches them based on conceptual similarity.
This has two practical implications.
Keyword stuffing does nothing. Repeating best Boston dentist twenty times in a paragraph does not help. The engine does not care. It cares about whether your content covers the meaning of the query.
Synonyms and related concepts do help. A page about Boston dental implants that also covers cost, recovery, candidates, and alternatives ranks well in vector retrieval because it covers the conceptual neighborhood, not because it uses any specific phrase.
In practice: write for full topical coverage, not for keyword density.
How RAG pipelines work
Most production AI search engines use Retrieval-Augmented Generation (RAG). The pipeline has three steps.
Retrieve. The engine searches its index for chunks relevant to the query. This step usually combines vector search with traditional keyword search, plus filters on metadata like freshness and source authority.
Augment. The retrieved chunks are loaded into the LLM's context as evidence.
Generate. The LLM synthesizes an answer using the retrieved chunks, often including citations to the original sources.
The key insight: the retrieval step is the gate. If your content does not get retrieved, it cannot be cited, no matter how good the writing is. Your job is to be retrievable.
Query fan-out
LLM systems often fan out a single user query into multiple sub-queries before retrieval. A query like Is GBP worth it for a Boston restaurant? might decompose into sub-queries about cost, setup time, ranking factors, review impact, and competitor visibility. The engine retrieves chunks for each sub-query and synthesizes them.
This means your content does well in AI search when it covers the natural follow-up questions, not just the literal query. A page that answers only the surface question gets retrieved less often than one that anticipates and answers the decomposed sub-queries.
Practical move: when planning a piece, list the five questions a customer would ask after the headline question, and structure H2s and H3s around them.
Structuring content for chunking
Chunking is how the engine splits your content into retrievable units. You can either let the engine guess where the chunks should be, or you can signal it explicitly through structure.
Engines tend to chunk on:
- HTML headings (H1, H2, H3, H4)
- Paragraph breaks
- List boundaries
- Section delimiters in structured data
A page with one H1, six clear H2s, and short paragraphs chunks predictably. A page with one giant wall of text chunks unpredictably and often poorly.
Best practices:
- One H1 per page
- H2s for major sections (3-8 per page)
- H3s for sub-sections inside H2s when needed
- Short paragraphs (2-4 sentences)
- Bulleted lists for parallel items
- Tables for comparisons
Make chunks self-contained
The chunk an engine retrieves often gets shown to the user with no surrounding context. If your section relies on something said earlier in the page, the chunk will read as half-finished and the engine may skip it.
Three rules:
Repeat key entities. If a chunk references the company instead of NOVA Brandworks, the chunk loses meaning when shown alone. Use names, not pronouns, in the opening of each section.
Lead with the answer. The first one or two sentences under every H2 should be a complete, standalone answer. Do not lead with setup or context. AI engines often extract just the opening paragraph.
Include concrete examples in each section. Examples ground the chunk and signal that this section can answer the user's question.
Schema and Q&A patterns
Schema markup tells AI engines what each section of your page is about in a machine-readable format. The most useful schemas for AI search:
| Schema type | What it does | When to use |
|---|---|---|
| FAQPage | Marks question-answer pairs explicitly | FAQ sections |
| HowTo | Marks step-by-step instructions | Tutorial content |
| Article | Marks the page as a published article with author/date | Most blog posts |
| LocalBusiness | Marks the entity behind the content | Service pages and homepage |
Schema does not magically rank you. It removes ambiguity in how the engine reads your page. For content that is already good, it lifts citation rates noticeably. For content that is bad, it does nothing.
How real RAG pipelines work in production
Production RAG systems combine multiple retrieval methods.
Vector search finds content with semantically similar meaning to the query.
Keyword search finds content containing exact terms from the query (still useful for branded queries, technical terms, and proper nouns).
Metadata filters narrow results by date, source, type, or other attributes.
Re-ranking scores the candidate chunks against the query before passing them to the LLM.
Optimizing for real-world RAG means optimizing for all four. Vector search rewards topical coverage. Keyword search rewards clear use of the actual terms a user might type. Metadata filters reward fresh dates and clean structured data. Re-ranking rewards content that reads as authoritative and relevant.
A page that does well in all four ends up cited far more often than a page that wins one and loses three.
A workflow you can reuse
For any new piece of content, follow this loop:
- Pick a pillar topic. Something with clear search demand and conceptual depth.
- Outline as questions. What are the five to ten questions a reader would ask?
- Each question becomes an H2. Lead each section with a direct answer.
- Make each section self-contained. Use entity names, include examples, write a complete answer.
- Add schema. FAQPage on the FAQ, Article on the page, LocalBusiness on the site.
- Internal-link to related content. Build the cluster.
- Test retrievability. Run the question in ChatGPT and Perplexity a week after publishing. See if you get cited.
FAQ
What is a chunk?
A chunk is a short, self-contained section of your content (usually 300-800 words) that an AI engine can extract and cite.
Do keywords still matter for AI search?
They matter less than for traditional SEO, but they are not irrelevant. Hybrid retrieval still uses keyword matching alongside vector search. Use keywords naturally; do not stuff.
What is RAG?
Retrieval-Augmented Generation. The technical pattern most AI search engines use: retrieve relevant content, augment the LLM's context with it, then generate the answer.
How long should AI-search-optimized content be?
Long enough to fully cover the topic and answer the natural follow-up questions, short enough that each section stays focused. Most well-performing AI search content lands between 1,500 and 3,000 words, broken into clear H2 sections.
Is technical SEO still relevant for AI search?
Yes. Crawlability, page speed, indexing, and structured data still gate everything. If your site does not technically work, AI search optimization will not save you.
.png)
