Passage-Level Optimization: The Inverted Pyramid for AI

Most agencies still write pages like magazine articles: a clever lede, three or four meandering paragraphs, then finally the substance. That pattern was fine in 2018 when humans read the whole page. It is fatal in 2026 when AI engines extract 120-180 word chunks and stitch them into answers. The inverted pyramid is the single highest-leverage formatting change you can make for GEO.

This post is the operator manual for passage-level optimization. For the broader GEO context, see the full pillar. For schema specifics, see the schema deep-dive.

Why do AI engines extract passages instead of whole pages?

Because the synthesis layer (the LLM that composes the answer) has limited context window and tight latency budgets. Retrieving and embedding a whole 3,000-word page wastes tokens. The retrieval pipeline chunks each page into passages, indexes the chunks, retrieves the top-scoring chunks for a given query, and feeds only those chunks to the LLM. Typical chunk size is 120-180 words.

The implication: every passage on your page is competing for citation independently. A buried passage on page 2 of your blog post is irrelevant if the first passage on competitor X's page is cleaner. You're not optimizing a page — you're optimizing 15-25 passages on a page.

Google's own documentation backs the text-first part of this. Google Search Central states there are 'no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary' — and its actual advice is boring and structural: make sure 'important content is available in textual form' and that 'your structured data matches the visible text on the page.' There is no secret AI markup. If your best answer lives inside an image, a video with no transcript, or a script-rendered accordion, you have hidden it from the only thing that could have cited it.

Google also confirms that AI Overviews and AI Mode 'may use a query fan-out technique — issuing multiple related searches across subtopics and data sources.' That is the mechanism that makes passage-level work pay. The engine is not answering your one keyword; it is answering six sub-questions it invented, then assembling the result. A page built as one long argument has one way in. A page built as eight self-contained answer capsules has eight.

Passage ranking vs featured snippets vs AI extraction

Three overlapping ideas get conflated, so it's worth separating them. Passage ranking is Google's ability — rolled out across 2020-2021 — to rank a single relevant section of a broader page for a query, even when the page as a whole isn't about that query. A featured snippet is the extracted answer Google promotes to position zero at the top of the results. AI extraction is what generative engines do: pulling a self-contained chunk into a synthesized answer and citing it. All three reward the same underlying craft — a tightly scoped, front-loaded passage that answers one question on its own.

The distinction that matters in practice: passage ranking decides whether your section can surface at all, the featured snippet decides whether it's promoted above the fold, and AI extraction decides whether it gets cited inside an answer. Optimize the passage well and you improve your odds on all three at once — which is why the inverted pyramid pays off across classic SEO and GEO simultaneously. See get cited by Google AI Overviews for the AI-citation-specific version of this pattern.

What is the inverted pyramid pattern?

The inverted pyramid is a journalism structure where the most important information goes first, supporting facts come next, and context comes last. Applied to GEO, every H2 section opens with a 40-80 word answer capsule that fully answers the H2's question, then 1-3 supporting paragraphs add nuance, examples, and stats. It is the opposite of how marketing copy is usually written and the same as how news articles open.

• Sentence 1: complete answer in 15-25 words. Name the entity. Include a number if possible.
• Sentence 2: scope or qualifier. "For service businesses," "on transactional queries," "as of 2026."
• Sentence 3: supporting fact. A specific range, a comparison, a named tool.
• Optional sentence 4: an attributable source. Cite a real, linkable study (e.g. "per Ahrefs' AI search research") — never a benchmark you can't back.
• Total: 40-80 words. Below 40 = too thin. Above 80 = truncated.

What does a before/after rewrite look like for a dental practice?

Imagine the H2 "How much does a dental implant cost in Austin?". The bad version: "Great question — let's dive in! Dental implants are a serious investment, but the good news is that they can last a lifetime. The cost varies depending on several factors. Read on for our full guide." That's 38 words of nothing. AI engines will skip it.

The inverted-pyramid version: "A single dental implant in Austin costs $3,500-$6,000 including the post, abutment, and crown, with most practices clustered between $4,200 and $5,200. Full-mouth implant cases (All-on-4) run $24,000-$35,000 per arch. Pricing varies with implant brand (Straumann premium, Hahn mid-tier), bone-graft requirements, and whether sedation is included. Most Austin practices accept third-party financing via CareCredit or Cherry." That's 65 words with 7 specific numbers and 5 named entities. It gets cited.

What does it look like for an HVAC company?

Take the H2 "How long does a new HVAC system take to install?". The bad version: "It depends on the size of your home and the system you choose, but most installations take a few days. Our team works efficiently to get your new system up and running fast." Useless to the AI.

The inverted-pyramid version: "A standard residential HVAC replacement takes 6-10 hours for a single-stage system in a 2,000-square-foot home, or 1-2 days for a multi-zone system or full ductwork replacement. Same-day installs are common for direct one-for-one replacements with no electrical or duct modifications. Complex jobs — geothermal, dual-zone variable-speed, or new-construction installs — can run 3-5 days. Most contractors quote a fixed timeline upfront after a 30-minute site assessment." 72 words, 5 numbers, names the system types. Citable.

What does it look like for a personal injury law firm?

Take the H2 "How long does a personal injury case take to settle in Texas?". The bad version: "Every case is different. Some settle quickly, others take longer. A dedicated attorney will fight hard for the best possible outcome." That answers nothing.

The inverted-pyramid version: "Most Texas personal injury cases settle in 6-18 months from the date of injury, with simple soft-tissue claims often resolving in 4-9 months and cases involving surgery, disputed liability, or commercial defendants typically running 12-24 months. Cases that go to trial add another 12-18 months. Texas's 2-year statute of limitations sets a hard deadline for filing suit, but settlement negotiations can continue past the filing date." 72 words, 6 time ranges, names the statute. Citable, and useful.

Where on the page should citation-bait passages go?

Toward the top. A large share of citations tends to come from content high on the page — not because the top is magic, but because most pages put their best definitions there and chunking pipelines surface early, well-formed passages readily. Treat "front-load the answer" as the operating rule and don't make the engine dig for your best material.

Practical implication: put your highest-priority H2 answer capsules in the first 3-5 H2 sections of the page. Save backstory, history, and tangential context for the bottom half. Don't bury the lede.

And don't assume your ranking protects you. Ahrefs analyzed 863,000 SERPs and 4 million AI Overview URLs and found only 38% of AI-Overview-cited pages also rank in Google's top 10 for that query — down from roughly 76% a year earlier (March 2026). Citation and rank have decoupled. That cuts both ways: a strong ranking no longer guarantees you are the cited source, and a page that isn't ranking top-10 can still be pulled into the answer if its passage is the cleanest one available. Extractability is the variable you control.

How long should an extractable passage be?

Think in ranges rather than a single number, because the right length depends on the surface you're targeting. For a featured-snippet capture, the core answer wants to land in roughly 40-60 words — long enough to be complete, short enough for Google to lift verbatim. For broader informational queries, a slightly longer answer of up to about 100 words gives the engine more to work with. For AI-answer extraction, engines typically operate on 120-180 word chunks, so your 40-80 word answer capsule should sit inside a clean, self-contained chunk of that size.

There's a useful sanity check on those ranges. In Pew Research Center's study of 68,879 Google searches, the typical (median) AI summary was 67 words long, and 88% of summaries cited three or more sources (March 2025 data). Sit with that: the entire finished answer is about the length of one well-built capsule, stitched from three-plus sources. You are not competing to supply the whole answer. You are competing to supply one clean, quotable sentence or two inside it — which is exactly why a 40-80 word capsule is the unit of work.

The through-line is self-containment. Whatever the target length, the passage has to stand alone: no "as mentioned above," no pronouns pointing back at earlier paragraphs, no dependency on a caption or an image the extractor can't see. Front-load the answer, name the entity, and keep the supporting sentences tight — and the same passage becomes eligible for a snippet, a passage-ranking surface, and an AI citation without being rewritten three different ways.

How long should the page itself be?

Counterintuitively, shorter, tightly structured pages tend to extract more cleanly per passage. A focused 800-word Q&A page gives a chunking pipeline an unambiguous answer to pull; a sprawling 3,000-word guide dilutes any single passage's prominence. This isn't about absolute citation count — long pillar pages still earn more total citations because they contain more passages. It's about per-passage extraction efficiency, and it's a directional rule of thumb, not a precise ratio.

• Short Q&A pages (800-1,200 words): highest extraction rate per passage. Best for high-intent commercial queries.
• Mid-length cluster posts (1,500-2,500 words): balanced extraction + depth. The sweet spot for most cluster content.
• Long pillar pages (3,000-5,000 words): lower per-passage extraction but higher total citations because of passage count. Necessary for topic-authority signal but not the citation workhorses.
• Mega-guides (5,000+ words): diminishing returns on extraction. Only worth it for true pillar pages with strong internal linking.

What rhythm and sentence length works best?

Vary it. The most common AI-tell is a paragraph of 3-4 sentences of nearly identical length and rhythm — that's how LLMs generate by default, and Google's helpful-content classifier picks it up. Mix sentence lengths: a 22-word lead, an 8-word punch, a 14-word qualifier. Use occasional sentence fragments for emphasis. Write the way an experienced operator talks, not the way a junior writer thinks an expert sounds.

Avoid the em-dash as a comma replacement. AI overuses these. Avoid the bulleted list of 4 identical-length items — also an AI-tell. Use specific numbers, named tools, and direct comparisons. Write like a craftsperson who knows the work, not a content factory.

How do you audit existing pages for passage extractability?

Three-step audit. Step 1: pull your top 25 commercial pages. Step 2: for each page, check whether the first paragraph under each H2 is a 40-80 word answer capsule. Score yes/no. Step 3: rewrite the no's. Most pages written for human readers rather than AI extraction score poorly on the first pass — the marketing intro buries the answer. The goal after a rewrite sprint is 90%+ yes, on the premise that cleanly extractable passages are what the synthesis layer can actually cite.

• Audit metric 1: percentage of H2 sections whose first paragraph is 40-80 words. Baseline target: 90%+.
• Audit metric 2: percentage of first paragraphs that name the entity in the first 15 words. Baseline target: 85%+.
• Audit metric 3: average count of specific numbers per H2 section. Baseline target: 2 or more.
• Audit metric 4: ratio of question-formatted H2s. Baseline target: 80%+.
• Audit metric 5: presence of Article + FAQPage compound schema. Baseline target: 100% on Q&A pages.

Score each page against all five metrics. Pages scoring under 60% across the rubric are rewrite candidates. Pages scoring 80%+ are healthy. The work is straightforward — most pages take 45-90 minutes to rewrite once a writer has internalized the pattern. Budget 30-60 hours for a top-25-page rewrite sprint.

If you want this run for you, book a strategy call and we'll quote the audit + rewrite work. For dental practices specifically, see our dental SEO program. For the full GEO context, our GEO service covers the whole program.