Content chunking

How does content chunking work in AI search?#

Content chunking works by splitting each fetched page into passages — often a heading and the text under it — that are indexed, retrieved and scored individually, so the engine can quote the one section that answers a question and ignore the rest of the page.

Perplexity describes the step in its architecture write-up: a content-understanding module splits each document into “self-contained spans” that are “individually retrieved and ranked at query time.” Google moved the same way in classic search in October 2020, saying it could now “better understand the relevancy of specific passages” in addition to the page as a whole, a change it expected to improve 7 percent of search queries (Google).

1. 1

   The page is fetched as HTML

   Most AI fetchers read raw HTML and don’t run JavaScript, and Claude’s fetcher can truncate long pages to a token budget.

   Your lever: Server-render the page, and put the most important sections early.

2. 2

   It's split into passages

   Headings, paragraphs, lists and tables mark the boundaries. Perplexity parses list- and table-heavy pages more formulaically.

   Your lever: Use real H2s and H3s, and real HTML tables for specs and comparisons.

3. 3

   Each passage is embedded and indexed

   The passage is represented on its own, without the paragraphs around it. See vector embeddings.

4. 4

   Passages compete for the question

   Retrieval and reranking score passages against the query, and the answer quotes the winners.

   Your lever: Answer each section’s question in its first sentence.

Why do chunks lose context?#

Chunks lose context because each passage is indexed and read on its own, so anything it relies on from elsewhere on the page — the product name, the time period, what “it” refers to — disappears when the passage is lifted out.

Anthropic’s contextual retrieval research gives the textbook example: the chunk “The company’s revenue grew by 3% over the previous quarter” is useless alone, because it doesn’t say which company or which quarter. Anthropic’s fix — adding a short generated summary of the surrounding document to every chunk before indexing — cut failed retrievals by 35%, by 49% when paired with keyword matching, and by 67% with reranking added.

You can’t count on every engine doing that repair for you. The safe move is to write passages that carry their own context:

How do you write content that chunks well?#

Write content that chunks well by giving every section one question-shaped heading, answering it in the first sentence, naming the subject instead of using pronouns, and giving each section enough substance to stand alone — very short sections earned the fewest ChatGPT citations in SE Ranking’s study.

1. One heading, one question. The heading is the chunk’s label, so make it the question the section answers.
2. Answer first. Growth Memo found 44.2% of ChatGPT citations came from the first 30% of a page, and the same logic holds inside a section. See answer-first content.
3. Name the subject in each section’s opening sentence, even if the previous section just did.
4. Don’t go too thin. SE Ranking found sections under 50 words averaged 2.7 ChatGPT citations, against 4.6 for sections of 120–180 words and 5.7 for longer ones.
5. Use real structure. HTML tables and lists are parsed as structure, and they split cleanly into passages.
6. Put key facts in short sentences. Claude’s citations quote at most 150 characters.

Common mistakes with content chunking#

The most common content chunking mistakes are sections that depend on the one before, headings too clever to say what’s below them, answers buried after a long preamble, and key facts hidden in images, tabs or scripts a fetcher never sees.

Related terms#

• Content & relevanceAnswer-first contentA writing structure that puts the direct answer to a section’s question in its opening sentence and adds context and evidence after it, so readers and AI retrieval systems can take the answer without reading further.Read the entry
• How LLMs answerRerankingA second retrieval stage in which a more precise model re-scores the top results of a first, faster search against the query and reorders them — the step that decides which few passages an AI answer engine actually reads and cites.Read the entry
• How LLMs answerVector embeddingsNumerical representations of text — lists of hundreds or thousands of numbers — arranged so that passages with similar meaning sit close together, which lets search systems match a query to content by meaning rather than by shared keywords.Read the entry
• RAGHow LLMs answerRetrieval-augmented generationA technique in which an AI system first retrieves relevant documents from an index and then gives them to a large language model to write its answer, so the response can cite current sources instead of relying only on what the model memorized in training.Read the entry
• Content & relevanceFeatured snippetA highlighted excerpt Google shows at the top of some results to answer the query directly, lifted from one ranking page and credited with a link — the original answer-extraction format and the direct forerunner of AI Overviews.Read the entry
• Content & relevanceScaled content abuseGoogle’s spam-policy term for producing many pages mainly to manipulate search rankings rather than to help people — whether by AI, templates or hand — and it targets the purpose and value of the content, not the use of AI itself.Read the entry