Chunking is the step in every AI search pipeline where your page gets split into the passages the retriever actually scores. A page that reads beautifully end-to-end can still lose at retrieval because its chunks straddle topics, drop the question's keywords by the third sentence, or open with a context-setting paragraph that has no answer in it. Chunking is invisible to the human reader but it is the largest single lever you have over whether your URL gets cited.
How chunking works in practice
Most production retrievers chunk on a sliding window, usually 200 to 500 tokens per chunk with some overlap between chunks to preserve context across boundaries. The chunker tries to break on natural boundaries — headings, paragraph breaks, list items — but falls back to hard token cuts when paragraphs are too long. Chunks that span two unrelated topics tend to embed weakly for both, and chunks that bury the answer past the first 150 tokens often lose to a competitor whose answer is at the top of its chunk.
What strong chunks look like
A strong chunk has one topic, one definition-shaped opening sentence, and supporting evidence (numbers, examples, named entities) in the same passage. Lists with a parent heading chunk well because each item already has a context and a payload. Question-and-answer blocks chunk well for the same reason. Long uninterrupted prose chunks badly: even when it is great writing, the retriever cannot tell which two sentences are the answer.
Common chunking failures on marketing pages
Three patterns recur. The first is the windup: a hero paragraph that frames the topic but never states the answer, followed by sub-sections that finally do. The retriever often grabs the windup chunk and loses to a competitor whose first chunk is the answer. The second is the wall: a 600-word continuous paragraph that gets chopped on token boundary rather than meaning boundary. The third is the kitchen-sink section: a heading like 'Features' followed by twelve unrelated paragraphs that produce twelve weak chunks instead of one strong one.
How to chunk-proof a page
Run a tactical pass on each high-priority page. For every heading on the page, ask whether the first paragraph beneath it answers a single, literal buyer question. If it does not, rewrite it so it does. Cap paragraphs at three or four sentences. Use H3s liberally so the retriever has natural break points. Move FAQ-style sections close to the topic they address, not to the bottom of the page. Then re-test which passages from the page get retrieved for your priority queries.
How AskRanker's gap analysis uses chunks
When AskRanker compares your page to competitors for a buyer question, it does the comparison at the chunk level, not the page level. It identifies which chunks from competing pages are winning the query, what those chunks contain, and which of your chunks are closest. The Execute playbook then proposes paragraph-level edits to the chunks that are nearly winning, rather than asking you to rewrite the whole page.