Here is a thing most operators still get wrong: they write a page as one continuous argument, where paragraph nine only makes sense because paragraphs one through eight set it up. That works for a human reading top to bottom. It fails completely in AI search, because ChatGPT, Perplexity, Gemini, and Google’s AI Overviews almost never see your whole page. They see a chunk of it. One section, torn loose from everything around it, evaluated entirely on its own.
If you understand chunking, a lot of confusing AI-visibility behavior suddenly makes sense — including why a page that ranks fine on Google still gets ignored by every AI engine.
The mechanic: retrieval happens at the passage level, not the page level
When an AI engine answers a question, it does not load your URL and read it. It runs a retrieval step against an index that was built by slicing pages into passages — chunks of roughly a few hundred words, usually broken at heading boundaries. Each chunk gets its own embedding (a numerical fingerprint of its meaning). When a user asks something, the engine matches the question against individual chunks, pulls the two or three best ones, and writes an answer from those.
Your page does not compete as a page. Each section competes as a standalone unit. The model ranking your “Pricing” section against a competitor’s never sees your introduction, your proof points, or the definition you established four headings earlier. It sees that one block of text and decides whether it answers the question cleanly enough to quote.
This is why heading structure matters more than people think. Those H2 and H3 boundaries are very often the literal cut lines for chunks — and it tracks with the data that 68.7% of AI-cited pages follow a strict H1→H2→H3 hierarchy. Sloppy nesting doesn’t just look bad; it produces chunks that start mid-thought and get scored as incoherent.
Why most pages chunk badly
Pull any section out of the middle of your best page and read it cold, the way the model will. You will usually find it leaning on context that isn’t there. “As we covered above.” “This approach.” “It.” An acronym you defined once in the intro and never again. A claim whose supporting number lives two sections up. Every one of those is a dependency on text the model did not retrieve.
A chunk with broken dependencies reads as vague or unsupported, so it loses to a competitor’s chunk that happens to be self-contained — even when your page, read whole, is the better resource. You’re not being beaten on quality. You’re being beaten on packaging.
The other failure is the dump: 1,400 words under a single H2 with no internal structure. That doesn’t become one strong chunk. It becomes a few mediocre ones, each a random slice with no clean beginning, no clear claim, no boundary the retriever can respect.
What to do this week
Read your top three pages section by section, out of order. Open each one, jump to a random H2, and read only that section as if it’s all you’ll ever see. If it depends on something earlier to make sense, it’s a broken chunk. This audit takes twenty minutes and is the single most useful thing you can do.
Make every section restate its own subject. The first sentence under each heading should name the entity and the claim in full — “Flat-rate AI SEO pricing runs $500 to $1,500 per month per client,” not “Our pricing is straightforward.” Kill cross-references like “as mentioned above.” Re-define acronyms the first time they appear in each major section, not just once at the top.
Front-load the answer inside every chunk, not just the page. A tight 40-to-60 word answer at the top of each section gives the retriever a clean, quotable unit — and it compounds with the fact that 44.2% of all LLM citations already come from the first 30% of a page’s text. Every section gets its own first 30%.
Break the long dumps. Any section over ~300 words without a subheading should be split at its natural seams. You’re not padding — you’re handing the retriever clean cut lines instead of letting it guess.
Need this done for you? Paris Roussos runs a flat-rate AI SEO service ($500–$1,500/mo per client, white-label for agencies) covering audits, schema and entity work, AI-visibility tracking, and content engineered to be cited by LLMs. Reach him at parisroussos@gmail.com.
Stop writing pages that only work read whole — write pages where every chunk can stand up alone and win on its own.