There's a real debate happening about whether you should deliberately structure content into distinct, self-contained chunks for AI retrieval, or whether that's over-engineering a problem the models already solve on their own. It's not a vibes debate. Someone actually ran the test.
This debate matters more than its slightly technical framing suggests. It's really a proxy fight over a bigger question: how much of the work of being understood by an AI system falls on the publisher, versus how much the model is expected to absorb on its own. Get the answer wrong in either direction and you either waste effort restructuring content that was already fine, or you leave real relevance on the table assuming a model will do work it was never architected to do.
The claim: Google says don't worry about chunking
The skeptical position, publicly associated with comments from Google's Danny Sullivan, is that content creators shouldn't need to manually chunk their pages for AI systems. The implicit argument: modern language models are sophisticated enough to parse and retrieve relevant passages from ordinary, human-written prose without publishers restructuring everything into rigid, chunk-friendly blocks. It's a reasonable-sounding position, and it matches the general "just write naturally, the model will figure it out" advice that's floated around GEO discourse all year.
It's also an easy position for a search engine to prefer, whether or not that's the intent behind it. Telling millions of publishers they don't need to change anything is a much smaller ask than telling them their content architecture needs a structural rework. That doesn't make the position wrong on its face, models genuinely have gotten better at handling messy, unstructured prose, but it's worth noticing that the low-effort answer and the search engine's preferred answer happen to be the same answer here, which is exactly the kind of alignment that deserves a second look rather than an automatic pass.
The rebuttal: a 19.24% relevance lift
Mike King's response, published on iPullRank in January 2026 and updated later that month, doesn't argue the point rhetorically. He ran a test. He took a paragraph covering two distinct subjects, the kind of dense, multi-topic writing that's common on enterprise pages, and split it into two separate, single-subject chunks. Then he scored both versions using a cosine-similarity method against target queries.
| VERSION | ML RELEVANCE SCORE | DATA-PRIVACY SCORING |
|---|---|---|
| Single combined paragraph | Baseline | Baseline |
| Split into two single-subject chunks | +19.24% | +1.29% |
The second metric, the smaller 1.29% lift in data-privacy scoring, is easy to skip past, but it points at a second, less obvious benefit of chunking. A cleanly separated unit is also easier to evaluate for what it does and doesn't disclose, which matters for enterprise content that touches compliance, security, or regulated claims. A tangled paragraph covering two subjects at once makes it harder for any system, human or machine, to audit exactly what's being asserted about which subject. Chunking isn't just a relevance play. It's a legibility play that happens to pay off on more than one axis at once.
A 19.24% relevance improvement from a structural edit alone, no new information added, nothing removed, just separating two ideas into two retrievable units, is a hard number to argue past with a general reassurance that models "handle it fine." They may handle it. They handle a properly separated chunk better.
King's post also makes the architectural case, not just the empirical one, which is what turns this from a single test result into a durable argument. He points to Infini-attention, Ring Attention, MemWalker, and Recursive Language Models, four distinct approaches to how modern systems manage and retrieve long context, and notes that all four are built around operating on discrete, retrievable units rather than ingesting a whole document as one undifferentiated block. That's not a preference these systems could be talked out of with better prompting. It's how the underlying retrieval mechanics are constructed, which means content that happens to already be chunked well has a structural advantage baked in before any relevance scoring even happens.
What SEO Week's data adds
King's own conference, SEO Week 2026, ran alongside this debate and added corroborating signal from other speakers. Krishna Madhavan, representing Microsoft and Bing, put the retrieval-era stakes plainly: "visibility belongs to content that AI can trust, understand, and ground." Scott Stouffer went further: "if you're not retrieved, you do not exist to AI," a blunt restatement of the shift we've written about before, that ranking and retrieval are increasingly separate games with separate rules. Jeff Coyle added the practical framing: "most failures start with eligibility," meaning a huge share of AI-visibility problems are decided before relevance is even evaluated, at the access and structure layer.
Christian Ward's line from the same event is the one that reframes the whole chunking argument: "no one wants to go to your website. The vast majority just want the answer." If that's the operating assumption, and a separate 38-million-data-point study presented at the same event found AI visibility is fragmented, personalized, and location-based rather than a single stable ranking, then chunking stops being a nice-to-have formatting choice and becomes the mechanism by which any of your content gets selected as "the answer" at all.
That fragmentation finding matters more than it might look at first. A single stable ranking is forgiving of sloppy structure, because there's one target to hit and time to iterate toward it. A fragmented, personalized, location-varying visibility picture is much less forgiving, because there's no single page state to optimize toward; the content has to hold up as a retrievable, self-contained unit across many different retrieval contexts at once. That's a much stronger argument for deliberate chunking than "it might help a little," and it's the argument the 38-million-data-point study is quietly making underneath its headline finding.
Who's actually right
Both sides have a defensible piece of this. The skeptical position is correct that you don't need to butcher natural writing into robotic bullet fragments; a well-organized page with clear paragraph breaks already chunks reasonably well by accident. King's rebuttal is correct that deliberate, tested chunking measurably outperforms accidental chunking, and that the retrieval architectures underneath modern AI search, Infini-attention, Ring Attention, MemWalker, Recursive Language Models, are explicitly built to work on passages, not full documents. The debate isn't "chunk or don't." It's "chunk on purpose or chunk by accident," and the data says purpose wins by a measurable margin.
It's also worth being honest about the limits of a single test, even a well-designed one. King ran this on one paragraph, scored against a specific set of target queries, and the 19.24% figure is a snapshot of that particular comparison, not a universal constant you should expect to reproduce exactly on every page you edit. What it does establish, reliably, is direction: deliberate single-subject structuring measurably beats accidental multi-subject density, on the architectures currently doing the retrieving. That's enough to change a content-editing habit, even without treating the specific percentage as gospel for every page.
How to structure content for both
One more practical distinction worth making before the close: chunking for retrieval is not the same project as writing shorter content overall. A page can be long, thorough, and comprehensive while still being internally chunked well, each section a clean, self-contained unit a retrieval system can lift on its own. The mistake to avoid is treating this debate as an argument for terse pages; it's an argument for well-bounded sections inside pages of whatever length the topic actually deserves.
Write for a human first, and then do the deliberate second pass King's data argues for: find the paragraphs on your highest-value pages that quietly cover two ideas at once, and split them into two, each with its own clear heading or clean topic sentence. That's the same discipline behind the schema and structure work we run in technical audits, extended one layer deeper into the prose itself. Pair it with the access-layer check our audits already run, and treat retrieval, not just ranking, as a number worth tracking the way our citation research treats mention rate. We ran exactly this paragraph-level chunking pass for a computer vision platform alongside its broader GEO engagement, splitting dense product-spec paragraphs into single-subject chunks across its highest-traffic pages, and the pages that got the treatment started surfacing in AI answers the ones next to them didn't. The debate is settled by the test, not the comment thread.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.