Content Granularity

Element Code: CO-031

What content granularity means

Granularity is the size of the pieces your content breaks into. A page is not one undifferentiated wall of text and it is not a hundred one-line fragments either. It is a set of sections, each sitting under a heading, each covering a coherent slice of the topic. Granularity is the check on whether those slices are the right size.

This matters more than it used to because of how AI answer systems read pages. Retrieval-augmented generation, the pattern behind ChatGPT browsing, Perplexity, Google AI Overviews, and similar tools, does not read your page top to bottom the way a person might. It splits the page into chunks, embeds each chunk as a vector, and retrieves the specific chunks that match a query. The unit it retrieves and quotes is the chunk, not the whole page. So the way you chunk your own content, through headings and paragraph structure, shapes what these systems can pull out and cite.

Why the wrong granularity hurts

Two failure modes, opposite directions.

Too coarse. A single 900 word section under one vague heading crams five distinct ideas together. When a retrieval system grabs that block for one of those ideas, it drags in four unrelated ones as noise. The relevance signal gets diluted, the chunk is less likely to be selected, and if it is selected the answer engine has to work harder to isolate the part that matters. Human readers suffer the same way: they cannot scan to the answer because the heading does not tell them what is inside.

Too fine. Chop the same content into fifteen tiny sections with one sentence each and you fragment ideas that belong together. A chunk that reads "It depends on your server configuration" is useless on its own because it has lost the question it was answering. Self-contained meaning disappears, and a retrieved fragment cannot stand as an answer without the context you scattered across neighboring blocks.

The target sits between them: sections that are each a complete, self-contained thought, readable and quotable without the rest of the page.

The granularity spectrum

How to detect granularity problems

1. Skim the headings alone. Read only your H2s and H3s, ignoring the body. If the outline reads like a coherent list of specific questions, granularity is probably fine. If headings are vague ("Overview", "More", "Details") or one heading clearly hides several topics, you have a problem.
2. Crawl and export the structure. Screaming Frog or Sitebulb can export every page's heading structure. Scan for pages with one giant H2 and no subheadings, or pages with a heading every two lines. Both extremes show up fast in a spreadsheet.
3. Read one section in isolation. Copy a single section out of context and ask: does this answer a question on its own? If it needs the paragraph above it to make sense, it is too fine. If it answers three questions at once, it is too coarse.
4. Test with an AI system. Ask ChatGPT or Perplexity a question your page should answer, and see whether it can cite a clean chunk from your page or whether it paraphrases vaguely. Weak retrieval often traces back to poorly bounded sections.
5. Watch section length distribution. There is no magic number, but sections that swing wildly from three words to a thousand usually signal that structure was an afterthought.

How to fix granularity, step by step

1. Map one idea to one section. List the distinct questions the page answers. Each becomes a section with a heading that states the question or the answer plainly.
2. Split coarse blocks. Where one section covers several ideas, break it apart and give each piece its own descriptive heading. "Setup" becomes "Install the plugin", "Configure the API key", "Verify the connection".
3. Merge orphan fragments. Where sentences got their own heading for no reason, fold them back into the section they belong to so the idea is whole again.
4. Make headings descriptive. A heading is the label an AI system uses to decide what a chunk is about. "Common causes" beats "Notes" every time.
5. Front-load the answer. Open each section with the direct answer, then elaborate. Retrieval systems and skimming humans both reward the section that leads with the point.
6. Keep each section self-sufficient. A reader or a model should be able to lift any one section out and still understand it. Repeat a noun instead of leaning on "it" or "this" that points three paragraphs back.

What good looks like

FAQ