ContextBolt SEO Free for 7 days. Keywords, SERPs, backlinks and AI visibility, inside Claude. SEO data inside Claude. Start free trial
Guide · Content Chunking for SEO

Content Chunking for SEO: What You Actually Control

Search for content chunking and you land in the middle of an argument. Search Engine Land and Semrush have guides on how to do it. Ahrefs calls it overrated. iPullRank wrote a refutation of what it calls misinformation about chunking. Google has told you to stop.

I read all of it, including the primary sources each side cites, to find out whether a year of chunking advice should change how a page gets written. The two camps that disagree most loudly recommend almost the same page. They disagree about what to call it.

Quick answer
  • You cannot control how a system chunks your page. Boundaries are set by the retrieval pipeline, and your headings are not consulted.
  • Google says don’t do it. Danny Sullivan, January 8, 2026, on the record. Google’s AI docs say no special optimizations are necessary.
  • The two loudest opponents recommend the same work. Ahrefs calls it overrated and prescribes atomic content. iPullRank calls it essential and prescribes atomic passages.
  • The one test worth running is free. Delete the rest of the page and read the section alone. The one number worth watching afterwards is whether AI answers start naming you, which ContextBolt SEO tracks.

What is content chunking, and who is arguing about it?

Chunking is the step where a retrieval system splits a document into smaller pieces before it stores or searches them. Models have context limits, and scoring a paragraph against a question is more precise than scoring a whole page. Every retrieval-augmented system does it, and that part is not controversial.

The argument starts when the word crosses from engineering into marketing. In an AI pipeline, chunking is something the system does to your document. In SEO advice, “content chunking” means something you do to your document, hoping to influence the first thing. Those are different activities wearing one name, and almost all of the argument traces back to that.

WhenWhoPosition
August 21, 2025Despina Gavoyannis, AhrefsChunk optimization is overrated. Engineers control chunking, not you
January 8, 2026Danny Sullivan, Google”So we don’t want you to do that”
January 15, 2026Mike King, iPullRankPassages are the atomic unit of meaning. Sullivan is wrong
OngoingSearch Engine Land, Semrush, LumarHow to chunk your content

Does Google chunk your pages?

Google’s public answer is that whatever it does internally is not your problem.

On the Search Off the Record podcast published January 8, 2026, Danny Sullivan took the advice head on. As Search Engine Land reported it, he repeated himself for emphasis: “So we don’t want you to do that. I was talking to some engineers about that. We don’t want you to do that.” His reasoning was the usual Google reasoning. Content built for a ranking system rather than a reader tends not to age well.

Google’s documentation says the same. Its AI features guidance states that “You don’t need to create new machine readable files, AI text files, or markup to appear in these features” and that “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

None of that says a smaller unit does not exist. It says you are not required to manufacture one. When passage ranking rolled out, Google was specific that it was not indexing passages on their own. It still indexed pages, still weighed whole-page signals, and added the ability to consider a passage as one more ranking factor. The guidance then was the same as Sullivan’s five years later: creators had nothing special to do.

What do the two sides recommend?

Ahrefs published “SEO Chunk Optimization is Overrated” in August 2025. Despina Gavoyannis argues that chunking happens inside model pipelines, driven by token limits and efficiency, and that none of it responds to your headings or paragraph lengths. Different systems use different strategies, and your carefully built section may get cut in half regardless. Then she recommends an atomic content workflow: self-contained sections that deliver a complete answer on their own.

Mike King published a refutation a week after Sullivan’s remarks. He argues that passages are the atomic unit of meaning across modern architectures, cites Google’s passage-indexing patents and a run of research papers, and shows with embeddings that splitting a paragraph covering several topics improved relevance scoring by 19.24%. Then he recommends structuring content into atomic, self-contained passages.

Side by side, the difference disappears.

AhrefsiPullRank
Can you control the chunk boundary?NoNot the point
Does the smaller unit matter?Not to youYes, enormously
What should you write?Self-contained atomic sectionsSelf-contained atomic passages
What should you stop doing?Chasing chunk sizesWriting multi-topic blobs

One calls the practice overrated and the other calls it essential, and they hand you the same instruction. Google objects to fragmenting a page into bite-sized pieces, which is not what either of them told you to do.

What are the chunk sizes people keep quoting?

They are real, they are published, and they are applied to the wrong thing.

If you build on OpenAI’s file search, the numbers are in the retrieval documentation. Every file is split into 800-token chunks with 400 tokens of overlap between consecutive chunks, and a response carries at most 10 results by default, with 50 as the ceiling. Those are defaults you can change, on your own uploaded documents, in your own vector store. They describe what happens when you hand a system a PDF, not what happens when ChatGPT cites your blog post. Nobody outside those companies knows the retrieval settings behind web search.

You can see how small the unit gets by the time it reaches an answer. Anthropic’s web search documentation specifies that each citation returns cited_text carrying “Up to 150 characters of the cited content”. That is roughly one sentence. The retrieval unit might be 800 tokens, the citation unit is about a sentence, and your page is neither.

What changed in 2026?

One thing changed, and it pushes selection further out of your hands. Anthropic’s newer web search versions add what the documentation calls dynamic filtering. With basic web search, every result loads into the model’s context. With web_search_20260209 and later, the assistant writes and runs code that filters the results first, so only the relevant content reaches the context window. The stated reason is token cost on search-heavy requests.

So there is now a programmatic pass between fetching your page and the model reading it, written on the fly for that specific question. There is no fixed chunk size and no stable boundary. If your mental model was “the system splits my page into pieces and picks one,” it is out of date. A filter looking for the part of your page that answers a question finds it faster when such a part exists.

What do you control?

Not the boundary, the size, or the strategy, which differs per system and changes without notice. You control one thing, and it is what all three camps are circling.

Whether a passage survives removal: Take any section of your page, cut it out, and read it with nothing around it. If it still answers a question completely, every system that wants to quote it can. If it opens with “this” or “as we saw above” or “the third reason”, nothing can lift it, because the meaning lives in the paragraph you did not include.

That test costs nothing, needs no tool, and takes ten seconds per section. A few habits follow from it.

  • Answer the H2 in the first two sentences under it: if the heading asks a question, the answer should not arrive in paragraph four.
  • Repeat the subject instead of using a pronoun at the start of a section. It reads slightly heavier to a human and makes the section liftable.
  • Give one section one job: King’s 19.24% number came from splitting a paragraph that was doing several jobs at once, the one manipulation with published evidence behind it.
  • Skip setup paragraphs that exist only to introduce the next paragraph. They are unquotable by construction and readers skip them anyway.
  • Do not fragment the page into a hundred tiny headings. That is what Google objected to, and it makes a page worse for the humans who still make up most of your traffic.

None of that is new. It is the answer engine optimization checklist, reached from a different direction.

How do you know whether it worked?

Almost none of the chunking discourse is measured. Both sides argue from mechanism, and the mechanism is not the outcome. The outcome is whether an assistant names you when someone asks the question your page answers.

That is checkable. Ask an assistant the question, read the answer, note who got cited, and do it again next month. Google will not help here: its generative AI report in Search Console counts appearances and withholds the questions, as I went through in AI search prompts in Search Console.

ContextBolt SEO is a hosted MCP server. You paste one URL into Claude Code, Claude Desktop or Cursor, ask in plain language, and every check saves to your SEO Dashboard. Two of its tools bear on this. ai_answer runs the question live in ChatGPT, Gemini, Claude, and Perplexity and reports what came back and who was cited. ai_visibility reads a historical index of Google AI and ChatGPT answers to show how often a brand shows up over time. It is $35 a month with 1,000 credits included, and there is a 7-day free trial with the card upfront and nothing due today.

Those tools tell you whether you are cited. They cannot tell you which passage did it, because no assistant reports that, whatever a pitch says.

What to do this week

Check first that the sections reach a crawler at all. If the page builds its content in the browser, an assistant that does not run JavaScript receives none of it, and no amount of good chunking helps. Our free GEO audit shows you the text that arrived.

Then pick your three most important pages and run the removal test on every H2 section. Cut each section into a blank document and read it cold. On my own pages the failure rate was higher than I expected, and every failure was the same: a section that opened by referring to something above it.

Fix those openings and stop there. Do not go looking for a token count, split the page into fragments, or buy a tool that promises to optimize your chunks, because the thing it optimizes is set by a pipeline that has never read your markup. Then ask an assistant the question and see who it names.

Where next

Frequently asked questions

Something else on your mind? Get in touch.

What is content chunking in SEO?

Chunking is splitting a page into self-contained passages so a retrieval system can lift one piece without the rest. In AI engineering it is a mechanical step done by the system. In SEO it has come to mean writing sections that still make sense when quoted alone.

Does Google want you to chunk your content?

No. Danny Sullivan said so plainly on the Search Off the Record podcast published January 8, 2026. Google's own AI features documentation adds that there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations necessary.

Can you control how an AI system chunks your page?

No. Chunk boundaries are set by the retrieval pipeline, not by your headings. OpenAI's file search splits documents into 800-token chunks with 400-token overlap by default. Your markup is not consulted. What you control is whether a passage survives being cut out.

Is chunking the same as passage ranking?

Not quite. Passage ranking is Google considering a section of a page as an extra ranking signal while still indexing whole pages. Chunking in the AI sense is a retrieval pipeline splitting a document before embedding it. They share an assumption that the useful unit is smaller than a page.

So what should you actually do about chunking?

Write sections that answer one question completely, then test them by deleting the rest of the page and reading what is left. If a paragraph needs the one above it to make sense, no system can quote it. See answer engine optimization for the full checklist.