What is content chunking?
Content chunking is organizing a page into self-contained sections, each built around one question or idea, so that any single section makes sense when read on its own. In writing, chunking means splitting concepts "into small pieces or 'chunks' of information to make reading and understanding faster and easier" [5]. In AI search, it means writing so a system that pulls one passage out of your page gets a complete, correct answer.
A chunked section on a page about heat pumps might look like this: the heading "How long does a heat pump last?", then a first sentence that answers it with a range and the main factor, then two or three sentences of detail. An unchunked version buries the same fact in the middle of a long paragraph about installation, after a pronoun that only makes sense if you read the paragraph before.
Why does content chunking matter for AI search?
Content chunking matters because AI answer engines retrieve and quote passages, not whole pages. When a system grounds an answer in search results, it links specific sentences of its answer to the specific sources that support them [7]. A passage that needs the rest of the page to make sense is harder to match to a question and harder to quote accurately.
Google describes its own systems as able to "understand the nuance of multiple topics on a page and show the relevant piece to users" [1]. The system finds the piece. Chunking makes sure the piece it finds is complete.
Does Google recommend content chunking?
No. Google does not recommend breaking content into small pieces for AI, and has said so twice in 2026. Google's guide to generative AI features in Search states: "There's no requirement to break your content into tiny pieces for AI to better understand it" [1].
Earlier, on the Search Off the Record podcast published January 8, 2026, Google's Danny Sullivan addressed the practice of producing "bite-sized" content for large language models: "We don't want you to do that. I was talking to some engineers about that. We don't want you to do that" [2]. He warned that tactics aimed at pleasing an LLM "may or may not have worked" and "may not carry through" as systems improve to "reward content written for humans" [2].
Google's objection is to fragmentation, not to structure. Clear headings and direct answers are long-standing good writing. What Google is warning against is cutting a page into thin, isolated snippets, or spinning separate pages for every variation of a question, which its guide links to the scaled content abuse policy [1].
What is the difference between writer-side and system-side chunking?
Writer-side chunking is how you structure your page; system-side chunking is how an AI retrieval system splits your text into pieces for its index, and you cannot control the second one. Most guides treat these as one thing. They are not, and confusing them leads to bad advice.
| Writer-side chunking | System-side chunking | |
|---|---|---|
| Who does it | The publisher | The search or AI system |
| What it decides | Headings, section scope, answer placement | Where text is cut into retrieval units |
| Unit | A section answering one question | A block of tokens, often with overlap |
| Who sets the size | The writer | The system's developers |
| Can you see it? | Yes | No |
System-side chunk sizes are engineering settings. OpenAI's retrieval documentation, for example, lets developers set a maximum chunk size and an overlap in tokens, with a sample configuration of 1,200 tokens per chunk and 200 tokens of overlap [6]. Google, Microsoft and OpenAI do not publicly document how their consumer search products split public web pages.
So there is no "ideal chunk length" for a writer to hit. Any guide that tells you to write sections of exactly N words to match an AI system's chunk size is guessing at a number the system's owner has not disclosed, and that number can change without notice. What a writer can control is that each section is complete wherever the system happens to cut.
How is content chunking different from passage ranking?
Content chunking is something a writer does to a page, while passage ranking is something Google does when it ranks one: Google scores individual passages within a page for a query while still indexing the page as a whole [3]. Google announced passage ranking in October 2020, and it went live for US English queries on February 10, 2021 [3]. Google initially called it "passage indexing" and corrected the name, because pages are still indexed whole [3].
Passage ranking is a useful precedent. It showed that Google could surface one good section from a long page years before AI Overviews existed, which is why Google says publishers do not need to cut pages up for its AI features [1].
Where did content chunking come from?
Content chunking comes from cognitive psychology and technical writing, and the idea of retrieving passages rather than documents comes from information retrieval research going back decades. Its use as an SEO and AI search term is recent, and Google pushed back on the AI-specific version in January 2026 [2].
1956: chunks in human memory
The psychologist George Miller's 1956 paper "The Magical Number Seven, Plus or Minus Two" described how people recode information into larger meaningful units, or chunks [4]. Technical writers and web usability specialists later borrowed the word for breaking text into scannable units [5].
1980s to 1990s: passage retrieval
Information retrieval research began returning relevant passages instead of whole documents. John O'Connor wrote on answer-passage retrieval in 1980, and Gerard Salton, James Allan and Chris Buckley argued in 1993 that long documents often contain several topics, so retrieving excerpts is more useful [4].
2020 to 2021: retrieval-augmented generation and passage ranking
Retrieval-augmented generation made passages the explicit unit an AI model retrieves [4], and Google's passage ranking reached US search results on February 10, 2021 [3].
2026: Google warns against bite-sized content
As "chunking for LLMs" spread through AI search advice, Google's Danny Sullivan said in January 2026 that Google does not want content made bite-sized for machines [2], and Google's May 2026 guide said there is no requirement to break content into tiny pieces [1].
How do you chunk content without fragmenting it?
To chunk content without fragmenting it, keep the page whole and make each section stand alone: one question per heading, the answer in the first sentence, and no references that depend on earlier sections. A practical check:
- Read each H2 section in isolation. Copy one section into a blank document. If it no longer makes sense, it is not self-contained.
- Replace backward pronouns. "It", "this" and "they" at the start of a section usually point to something above. Repeat the noun.
- Put the answer first. The first sentence under a question heading should answer it completely.
- Keep one idea per section. If a section answers two questions, split it under two headings on the same page, not across two pages.
- Keep tables and lists labeled. A table needs a heading or caption that says what it compares, so it still makes sense when lifted alone.
- Don't spin out thin pages. Separate URLs for every variation of a question is the pattern Google ties to scaled content abuse [1].
Every entry in this glossary is written this way: each H2 is a question, and its first sentence is the answer.
Does content chunking increase AI citations?
No published controlled study that Search Insider has found shows content chunking, on its own, increasing AI citations. The case for it is mechanical: AI systems quote passages, so self-contained passages are easier to quote. Guides promoting chunking, including those that cite Google's skepticism, generally recommend it on that reasoning and on experience rather than measured results [8].
Treat chunking as good writing that removes an obstacle, not as a ranking tactic. A clearly structured page with nothing new to say will still lose the citation to a page that holds the original fact.