If you’ve been in SEO long enough, you’ve seen this movie.

A new system shows up. A new “optimization” trend forms around it. Then Google (or someone adjacent) says: please stop doing that. And half the industry hears “it doesn’t work,” while the other half hears “it works so well they’re begging us not to.”

That’s basically where chunking sits right now.

Googler, Danny Sullivan, has explicitly pushed back on the idea of turning content into “bite-sized chunks” as a tactic for LLM-driven surfaces.

And to be fair, he’s warning against a real failure mode: teams creating two versions of content (one for humans, one for machines), then eating the operational cost when the system shifts.

But there’s a second reality that doesn’t go away just because we wish it would:

AI search systems increasingly select, extract, and recombine passages, not pages. That means the unit of “being understood” is often smaller than the webpage you publish.

So the actual question isn’t “should we chunk?” It’s: should we make our content resilient when it gets pulled apart?

Chunking is one of those terms that got overloaded.

In Retrieval Augmented Generation (RAG) pipelines, chunking is the act of breaking content into retrievable components so the system can fetch only the most relevant passages for a query.

This isn’t hypothetical. Google has discussed passage-level behavior for years (often described as “passage indexing” / passage ranking).

In the SEO community, “chunking” usually means structuring content so passages and statements perform better when retrieved, without turning the page into a chopped-up mess.

But Mike King reminds us, chunking isn’t “chopping paragraphs into smaller paragraphs and using more headings and hoping for the best.”

That distinction matters, because most chunking advice fails by confusing “format” with meaning.

Here’s the honest answer: yes, sometimes, and also, it can absolutely waste your time.

A paragraph that covers two ideas is often a worse candidate than two paragraphs that each cover one idea. That’s not “writing for robots.” That’s just making each passage a clean unit of meaning.

Modern systems evaluate relevance at the passage level, and chunking lets you intentionally shape content where relevance is actually computed.

Sullivan’s warning is basically about incentives: when people optimize for a perceived system quirk, they create fragile work that doesn’t age well.

And you can see the failure mode in the wild: Pages that read like an exploded FAQ. Chunks that are missing context, so they’re easy to retrieve and easy to misinterpret. Layouts that prioritize extraction over persuasion.

Chunking helps when it produces clarity.
Chunking hurts when it produces fragmented language.

The practical move is to stop thinking “How do I break this up?” and start thinking:

How do I structure this so each passage survives retrieval without losing meaning?

Because even if models get bigger context windows, the system still has constraints: cost, latency, routing, summarization, and selection. Structure remains important because it creates units of meaning.

That matches what we see in AI search interfaces: your page is less a destination and more a dataset.

So you need structure that serves: Humans (clarity, trust, action), and Machines (extractability, coherence, unambiguous claims). Not two versions. One version that survives both.

The goal is not “more better units of meaning”.

Short, focused paragraphs reduce the chance the model extracts the wrong part, or skips you because your answer is buried.

A good paragraph can be quoted out of context and still be correct.

Think of headings as retrieval scaffolding. LLMs and readers both use hierarchy to understand relationships between ideas.

If everything is an H2 (or worse) everything looks like flat.

Don’t hide the lead. Put the definition, constraint, or key recommendation early, then expand with nuance.

This is good writing.

Lists, steps, tables, “common mistakes,” “when to do X vs Y.” These formats reduce ambiguity.

If you want to be cited, make it easy to quote without distortion.

If your brand voice hates those phrases, fine. You can still add the section with different clear direct langue.

Pop-ups, modal CTAs, and disjointed carousels can pollute what the model sees, even if users close them.

Schema still matters. But it’s not a magic bullet.

Google’s own guidance continues to emphasize structured data as useful, especially when it matches visible content.

Don’t forget this: Structure and clarity first; Structured Markup reinforces, it doesn’t rescue.

Carolyn Shelby drives this point home: “This is why poorly structured content – even if it’s keyword-rich and marked up with schema – can fail to show up in AI summaries, while a clear, well-formatted blog post without a single line of JSON-LD might get cited or paraphrased directly.“

Here’s a simple rule that won’t age badly:

Optimize for meaning that survives extraction. Not for “LLM tricks.”

Guardrails:

Chunking is what systems do to content. Your job is to make sure your meaning doesn’t break when that happens.

So yes, structure for passage-level content is good for the same reason you do most good SEO infrastructure work:

That’s the only kind of optimization worth defending.

One more thing: AI is only as good as it’s operator, and if you are reading this newsletter, you are better than most!

Till next time,

PS: Let me know what you think of this issue, or anything else here: [email protected]