Guide 12 · Evidence review
Content chunking: do you need to chunk content for AI?
No. Google says so directly: there is no requirement to break content into small pieces for AI to understand it, because its systems can find the relevant part of a long page. Our own citation data points the same way. The vendor cited most often across 45 AI answers publishes long comparison articles, not fragments.
What is content chunking supposed to do?
The idea comes from how retrieval systems are built. A retrieval pipeline splits documents into passages, turns each passage into a vector, and pulls back the passages closest to the question. From that, a piece of GEO advice follows: write short, self-contained blocks, one idea each, so every block can be retrieved on its own.
The first half is a fair description of how many retrieval systems work. The second half is an inference about what site owners should do, and the engines never asked for it. The splitting is done on the engine’s side, by the engine’s own rules, on whatever page it fetched.
What does Google say?
Google’s guide to optimizing for generative AI features is unusually blunt about it: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” It adds that Google’s systems can understand a page that covers several topics and show the relevant piece.
The same guide does ask for pages that are “well written and easy to follow”. That is an argument for clear headings and short paragraphs because readers need them. It is not an argument for cutting one good article into twenty thin pages.
What does our citation data show?
In the AI Plugin Leaderboard run of 17 September 2026 we read 45 answers from ChatGPT, Perplexity and Google AI Mode across five plugin categories. One publisher, FS Code, was a source in 18 of the 45, in all five categories and on all three engines. It sells a booking plugin and a social-posting plugin. What it publishes around them is long, structured comparison articles about neighbouring categories.
That is one publisher, and correlation is not a mechanism. But if engines preferred small self-contained fragments, the single most-cited vendor in the dataset would be an odd place to find long-form articles. Google’s own description of “query fan-out”, several related searches feeding one answer, also rewards a page that covers a topic thoroughly enough to match more than one of those searches.
Splitting one article into many short pages multiplies URLs, and each one has to be discovered, crawled and indexed on its own. On this site in September, 14 live pages sat in “Discovered – currently not indexed” for days. A page that is not indexed cannot be cited at all.
So how should a page be structured?
For the reader. Question-shaped headings where the reader would ask a question, a direct answer under each, and the evidence after it. That is the structure of this page and of every guide on this site, and it is ordinary good writing. It also happens to make a passage easy to lift, which is a side effect, not the goal.
What I would not do is rewrite pages that already work, split long guides into a series, or pay for a “chunk optimisation” pass. None of it has evidence behind it, and Google has said in writing that it is not required.
What would change my mind?
Our data is WordPress plugin categories only, one run of 45 answers, and it cannot separate page length from everything else FS Code does. Google’s guide speaks for Google Search only.
A controlled test would: the same content published as one long page and as short pages, with citation rates measured across engines over several re-runs. If short pages won consistently on ChatGPT or Perplexity, this page would change, and it would say so with the date.
How this was made: figures come from our own runs in the engines’ public interfaces or from the studies named above. Drafted with an AI assistant (Claude) and reviewed by Alauddin Chowdhury. Method and AI use.
Want to know what the engines read for your category?
Send me your wordpress.org slug. I will run your category on all three engines and show you which pages they actually cite, and in what format. Free, no call, usually within a day.