AI Chunk Optimization

Element Code: TE-015

What chunking actually is

Retrieval-augmented systems, which is what sits behind ChatGPT search, Perplexity, Gemini grounding, and most AI answer engines, do not feed whole pages to the model. They split documents into passages, usually a few hundred tokens each, convert each passage into an embedding, and store those in an index. When someone asks a question, the system embeds the question, finds the closest passages, and hands only those passages to the model that writes the answer.

Google has worked at passage level for years in classic search too: passage-based ranking rolled out in early 2021, letting a single well-focused section rank when the page as a whole is a weak match. AI Overviews and AI Mode push the same idea further, they ground answers in specific retrieved passages and cite the pages those passages came from.

The practical consequence: the unit of competition is no longer the page, it is the chunk. Your page can be authoritative overall and still lose every retrieval because no individual slice of it answers anything on its own.

Why poorly chunked content loses

Splitting is mechanical. A chunker does not understand your narrative arc; it cuts where the structure tells it to, typically at headings and paragraph boundaries, or just every N tokens with some overlap. Three failure modes follow:

• Orphaned context. "This approach also works for larger sites" is meaningless when "this approach" was defined two sections earlier. The chunk embeds poorly and retrieves for nothing.
• Buried answers. If the direct answer arrives in sentence six, after warm-up prose, the chunk's embedding is diluted with filler and loses to a competitor whose first sentence is the answer.
• Split answers. A definition in one paragraph and its qualifying condition three paragraphs later can land in different chunks. The system retrieves half an answer, and the model either skips you or, worse, cites you saying something incomplete.

I have watched pages with genuinely better information get out-cited by mediocre competitors purely because the competitor wrote in self-contained blocks. It is annoying, and it is fixable.

Bad chunks vs good chunks

How to structure content for chunking

None of this requires new tooling. It is disciplined writing plus clean HTML:

1. Make headings carry the question. "How long does a site migration take" beats "Timing considerations". The heading often travels with the chunk and anchors its meaning.
2. Answer first, elaborate second. Open each section with the direct answer in one or two sentences, then add nuance. Journalists call it the inverted pyramid; retrieval systems reward it.
3. One idea per section, roughly 75 to 300 words. Long enough to be substantive, short enough that a mechanical splitter will not shear it mid-thought.
4. Kill long-range pronouns. Every "this", "it", or "the above method" that refers outside the current section is a context leak. Repeat the noun. It reads slightly redundant to you and perfectly clear to a machine reading one slice.
5. Keep claims and caveats adjacent. If a statement is only true under conditions, state the condition in the same paragraph.
6. Use semantic HTML. Proper h2/h3 hierarchy, real ul/ol lists, real tables. Heading elements are the most reliable split boundaries chunkers have. A div-soup page built from styled spans gives them nothing to cut on.
7. Lists and tables for enumerable facts. Steps, comparisons, and specs survive chunking better as structured elements than as comma-strung prose.

Chunk-friendliness by content element

How to detect the problem on your own pages

The core test costs nothing: read each section in isolation. Copy a single heading plus its paragraphs into a blank document. Does it make sense with zero surrounding context? Could it answer a question by itself? If not, that chunk is dead weight in a retrieval index.

Beyond the manual test:

• Screaming Frog: crawl and review heading structure per URL. Missing h2s, skipped levels, or 2,000-word gaps between headings all predict bad splits.
• Ask the engines: run your money queries through Perplexity, ChatGPT search, and Google AI Mode. Note who gets cited and open the cited pages. You will usually find tight, self-contained sections doing the winning.
• Embed it yourself: if you are technical, split your page with any open-source text splitter, embed the chunks, and query them with real user questions. Seeing your own content fail retrieval is uncomfortable and extremely motivating.
• Log files: check server logs for GPTBot, ClaudeBot, PerplexityBot and friends to confirm AI crawlers actually fetch the pages you care about.

FAQ