Obility Editorial · · 4 min read
The engine does the cutting
A popular piece of AEO advice says to rewrite your pages as short, self-contained chunks, often with a target word count per section, because AI assistants are said to prefer bite-sized content. The premise is half right. Answer engines do work with pieces of pages rather than whole pages. The conclusion, that you should fragment your content to match, does not follow from it.
Microsoft describes the mechanism plainly. In an October 2025 post on its advertising blog, Krishna Madhavan writes that assistants like Copilot break content into smaller, structured pieces, a process the post calls parsing, and that those pieces are evaluated for authority and relevance and then assembled into answers. Google has ranked at this level for years. In October 2020 it announced that its systems could understand the relevance of specific passages as well as the overall page, and said the change would improve 7 percent of search queries across all languages.
Where a page is split, how large each piece is and how the pieces are scored are decided by the engine. They differ between engines, and none of the companies publishes them. No engine has released a recommended section length, so a word-count target for “AI chunks” is a guess about a system nobody outside can see.
Google asks you not to write a second version for machines
Google has addressed the advice directly. On the 8 January 2026 episode of its Search Off the Record podcast, Google’s Danny Sullivan said that turning content into bite-sized chunks because LLMs like them is something Google does not want people to do. He said he had discussed it with engineers, and that Google does not want people producing two versions of their content, one for the LLM and one for everyone else.
His argument rests on durability. He allowed that some people may be seeing a small advantage today, then pointed out that the systems may change, and that they have tended to move toward rewarding content written for people. Work done only to please one ranking system can be stranded when that system improves.
Microsoft’s post leans the other way on format. It recommends concise, self-contained phrasing in lists, Q&As and tables, while warning against turning every line into a bullet. Read side by side, the two companies disagree about layout. They agree about something more useful, which is what a good passage contains.
A passage fails when it loses its subject
Anthropic’s September 2024 write-up on what it calls contextual retrieval shows the failure clearly. Its example is a chunk from a company filing that reads “The company’s revenue grew by 3% over the previous quarter.” Lifted out of its document, the sentence no longer says which company or which quarter, so a search about that company’s growth can miss it. Anthropic’s fix runs on the retrieval side: it adds a short description of the source document to each chunk before indexing. In its tests that cut failed retrievals by 49 percent, and by 67 percent when combined with reranking.
Two lessons follow. The people who build retrieval systems treat lost context as a serious enough problem to engineer around, and their repair is to put the missing subject back. A passage that already names its subject does not depend on that repair being in place. Keep the boundary in view, though: the write-up describes a technique for retrieval systems in general. It does not say how Claude’s web search, or any other assistant, processes your site.
Microsoft’s list of writing mistakes points the same way. It names overloaded sentences that pack several claims into one line, and labels such as “next-gen” with no context, which it says leave the system unsure how to classify the content.
Fix the sentence before you touch the layout
The practical change is small, and it helps human readers as much as machines. Take the paragraph on your product or pricing page that answers the question buyers ask most, and read it alone, as if it were the only text an assistant had. Check that it names the product instead of saying “it” or “our platform”, that each claim keeps its conditions (the plan, region or company size it applies to) in the same sentence, and that every number sits next to what it counts.
Here is an illustration with a made-up product. “It integrates with all major tools and is priced per seat” fails that test. “Acme Payroll connects to QuickBooks and Xero and is priced per employee per month” passes, and costs a human reader nothing. Headings help in the same way when they state what the section answers. Sections can stay as long as the subject needs.
What to avoid is the step Google asked people to skip: splitting one coherent explanation into a dozen fragments, or publishing a stripped “AI version” of a page. Fragments tend to repeat the subject without adding facts, and a second version doubles the pages you have to keep accurate.
In Obility, Context Manager records product names, descriptions, intended audiences and approved claims, which gives writers one reference for the exact naming that self-contained sentences depend on. Content Agents keeps a person reviewing factual claims, brand language and sources before anything is published, and that review is the natural place to read key passages on their own.
A clear passage can still lose to a better source
Self-contained writing removes one reason a strong page goes unused. It does not make an engine choose you. Selection still depends on whether the page is in the index that engine searches, and on whether a third party is the more credible source for the claim in question.
Measure it the slow way. Record which passages you rewrote and when, then compare the cited sources for the same buying prompts across repeated observations before and after. Do not credit a rewrite for a change seen in a single answer. If several weeks show no movement in citations, you still have a clearer page, which is the outcome Google says it wants in the first place.