TL;DR. AI engines do not quote pages, they quote paragraphs. Each paragraph has to stand alone when pulled out of context. The rule: 50–120 words, one idea per paragraph, no mid-paragraph references to "above" or "earlier." Diagnose by cutting one paragraph out of your page and pasting it into a blank doc. Does it still make sense? If not, that paragraph will have a hard time to be cited.
What chunking is
Chunks is a new term we use in GEO; they are basically just paragraphs. RAG (retrieval-augmented generation) systems break pages into chunks (usually by paragraph or heading boundary), embed each chunk as a vector, and retrieve the ones most relevant to a user query. The engine then quotes or paraphrases the retrieved chunks in its answer.
Chunks do not travel with the surrounding page. If your paragraph starts with "As we discussed above, this metric..." the engine sees a fragment with no anchor. It will not quote it. You are not cited.
The 4 chunkability rules
Rule 1: 50–120 words per paragraph - Too short (< 40 words): the chunk lacks enough context to be meaningful when retrieved. Too long (> 150 words): the chunk is below the retrieval-relevance threshold for most RAG systems; gets skipped.
Sweet spot: 50–120 words. Aim for 70–90 as the default.
Rule 2: One idea per paragraph - Each paragraph answers one question or makes one claim. If a paragraph introduces a new idea halfway through, split it.
Test: write a one-sentence summary of the paragraph. If the summary contains "and also" or "meanwhile," you have two ideas. Split.
Rule 3: No mid-paragraph pronoun references to "above" - Phrases to delete:
- "As mentioned earlier..."
- "As we discussed above..."
- "Continuing from the previous point..."
- "This ties back to..."
- "Building on the prior section..."
Replace with self-contained naming. Instead of "As we discussed above, this metric matters," write "Detection Rate is the primary branded signal. If AI engines cannot find your brand on queries that include your brand name, the problem is almost always SEO."
Rule 4: Lead with the claim - Put the main claim in the first sentence of the paragraph. The rest is justification, example, or caveat. This is the BLUF (bottom-line-up-front) pattern applied at paragraph level (more on BLUF in Module 5.3).
Anti-pattern:
"There are several factors to consider when evaluating AI visibility. Among these, one that frequently gets overlooked is..."
Pattern:
"Detection Rate is the single most important metric for diagnosing branded visibility. Below 90% = entity emergency."
The paste-into-a-blank-doc test
Open your priority page. Pick one paragraph. Cut it, paste it into a blank document. Read it cold.
Pass: you understand what it is claiming and why, without needing the rest of the page. Fail: the paragraph is dependent on context. Rewrite.
Repeat for every paragraph on the page. Every one has to pass.
The "chunk to H2" ratio
Aim for 1–3 body paragraphs under each H2. More than 5 and you should split the H2 into two. Less than 1 means the H2 is decorative, not functional.
Do this now:
Pick the worst-performing paragraph on your priority page (usually the longest one or the one with the most transitional phrases). Cut it, paste it into a blank doc, read it cold. Rewrite to pass the test. Ship today.
Start improving your AI visibility today with Rankscale.
Get started