Why does chunk optimization matter to AI models?
A generative engine does not load the whole page into its answer. It indexes the page in fragments - most often single H2 or H3 sections - and, when a question arrives, retrieves the ones that look most relevant. Your fragment reaches the model without the article title, without the previous section and without the introduction.
That is why “as mentioned above, this parameter is crucial” is a loss in GEO - out of context it means nothing and cannot be quoted. The same goes for a section that never repeats the topic it discusses, relying on a pronoun pointing at a heading two screens up.
The reverse also holds: a section written as a standalone answer can be cited even when the rest of the page is average. In AI Search the chunk is the unit of competition - not the page.
What will you see in the report?
You get a section-by-section autonomy assessment, flagging the ones that stop making sense without the whole page.
Plus two complementary metrics, computed algorithmically and shown next to the score but excluded from it: the citability of individual sections - the odds that this particular one gets picked as a source, built from heading-to-intent similarity, the self-containment of the first sentence, the share of declarative sentences, position on the page and heading depth - and the density of citable fragments across the text, meaning what percentage of words sit in passages fit to be quoted.
Separately we list headings that recur among the Top 10 competitors, marking the ones you do not have - a ready list of sections to add.
Sample recommendations
Fragments of a report from an audit of an electronics store category page.
Problem: The section and its H3 headings have zero citable density — empty fragments go into the index.
### Check related categories ### Looking for a specific brand? ### Want to know more?
Add 30–50 words under each heading with category names, brands and guide titles.
Problem: The product card carries no model name, so once cut out of the listing it stops meaning anything.
Product type: mouse. Scroll wheel: yes. Transmission: wired. €34.90.
Put the full commercial name at the top of the card — a chunk has to stand without the rest of the page.
How do we measure chunk optimization?
The model examines each section separately on four points: whether it opens with an answer, whether it repeats the main topic at least twice, whether it avoids pointing at other fragments (“above”, “as we wrote earlier”), and whether its length suits the content type.
That last criterion depends on the content type: an FAQ answer should be short, an encyclopedia entry may be long:
Optimal section length depends on the page type
| Content type | Section length |
|---|---|
| Article | 200-500 words |
| Encyclopedia entry | 400-1500 words |
| Comparison | 150-400 words |
| Landing, listing, product | 80-300 words |
| Tool, calculator | 50-200 words |
| FAQ | 30-150 words |
What raises and what lowers the score?
Raises
- A section opening with a declarative sentence that is itself an answer
- The main topic named explicitly rather than replaced by a pronoun
- Length matched to the page type, not to the article norm
- A heading that reads like a user question, not like a slogan
- Facts and numbers inside the section, not in a separate summary at the end of the page
Lowers
- References such as “above”, “as we already mentioned”, “in the previous part”
- A section opening with a rhetorical question or a connective (“However”, “That said”)
- Opening with a pronoun that points at nothing (“This means that...”)
- A wall of text with no section breaks
- Thirty-word sections in an article, or 900-word answers in an FAQ
Frequently asked questions
Do I have to split the text into short sections?
No, the length has to match the content type. An encyclopedia entry can run 1500-word sections, an FAQ answer 50. Artificially cutting an article into 100-word pieces lowers the score just like a wall of text does.
Do phrases like “as mentioned above” really hurt?
Yes, and specifically so. The section reaches the model without its neighbours, so the reference points at something that is not there. Such a fragment stops being quotable even when it is the strongest material on the page.
Does passage citability affect the dimension score?
No, it is a complementary metric shown alongside the score. It tells you which sections have the best odds of being selected as a source, but it does not enter the final result.