Lead with the answer. AI Overviews and Perplexity extract from the first 1-2 sentences after a heading. The rest of the section provides reinforcement, not the quote. This is the cheapest content change you can ship for AEO and GEO: rewrite the first paragraph under each H2, leave everything else alone, watch citation rates climb. A 2025 Authoritas study found a 38% higher likelihood of appearing in AI Overviews when clear entity signals appear within the opening paragraph.
The inverted pyramid is back, and machines read it now
Print journalists have led with the lede since the 1860s, when telegraph operators truncated wire copy that ran long. The first sentence carried the news; supporting detail trailed. AI extraction runs the same playbook — the model needs a self-contained answer it can quote without yanking context from three paragraphs down.
A Semrush analysis of 1.6 million featured snippets and a Portent study of paragraph snippets converged on the same range: 40-60 words, two to three sentences, with 45 words landing on SERPs 21% more often than any other length. Snippets longer than 55 words appeared in only 8% of cases. The shape Google's extractor wants is the shape an LLM wants: short, factual, scoped to one claim.
That convergence is the point. You're not optimizing for two different surfaces. You're writing one paragraph that satisfies Google's snippet selector and every LLM's source picker at the same time.
Before and after: a buried answer vs an answer-first rewrite
Here's a paragraph from a typical product help page about session cookies. The answer is in there. Find it.
Session management is one of the most discussed topics in modern web development, and there are many considerations to weigh when deciding how to implement it for your application. Cookies have been used since the early days of the web, and their behavior has evolved alongside browser standards. After much deliberation within our engineering team, we settled on a model that balances performance with security. The session cookie expires after 14 days of inactivity, at which point the user is signed out.
The answer ("14 days of inactivity") is the last clause of the last sentence, 67 words deep. An extractor scanning the first 60 words sees process narrative and gets nothing quotable. Now the rewrite:
Session cookies expire after 14 days of inactivity. Each request resets the timer; once a user goes 14 consecutive days without hitting the API, the cookie is invalidated and they're signed out. The 14-day window applies to both web and mobile clients.
Forty-three words, three sentences, the fact in sentence one. An LLM asked "how long do sessions last on this product?" can quote the first sentence verbatim with full context. The buried version forces the model to summarize, and summaries get attributed less often than direct quotes.
Question-based H2s match how AI Overview retrieval works
AI Overviews are triggered by queries phrased as questions. The retrieval system scores candidate passages partly on heading-query similarity. A topic-style H2 like "Session expiration" matches loosely; a question-style H2 like "How long do session cookies last?" matches the user query almost token-for-token.
Topic heading:
Session expiration
Session cookies expire after 14 days of inactivity...
Question heading:
How long do session cookies last?
Session cookies expire after 14 days of inactivity...
The body paragraph is identical. The question heading wins because it mirrors the actual search string. Lily Ray's 2025 analysis at Amsive found content with consistent heading hierarchy was 40% more likely to be cited by ChatGPT — and question-formatted headings sit at the top of that hierarchy.
A working pattern: pull the top 10-20 questions from your support ticket system, your sales call transcripts, or AlsoAsked. Use the exact question phrasing as your H2. Answer it in the next 40-60 words. Skip the throat-clearing.
The 1-3 sentence answer rule, and why longer kills extraction
The first paragraph after each heading should be 1-3 sentences, 40-60 words, factual, scoped to the heading question. Anything longer gets summarized by the LLM (citation drops). Anything shorter lacks the context for a standalone quote (extraction skips it).
Three checks before shipping a paragraph:
Could this be quoted alone? Strip the surrounding context. Does the paragraph still make sense? If sentence one references "this approach" or "the above", you've failed the standalone test.
Is the noun phrase repeated from the heading? "How long do session cookies last?" → "Session cookies expire after..." The subject of the answer matches the subject of the question. This is mechanical and it works.
Is there a number, date, or proper noun in the first sentence? Concrete facts get extracted. "Generally fast" doesn't. "Under 200ms p95" does.
What goes after the answer paragraph? Elaboration, edge cases, counter-examples, code snippets, links. The model can use these to ground its answer, but they're not the quote.
Speakable schema is a red herring — don't bother
Speakable JSON-LD lets you mark sections of a page as suitable for text-to-speech playback by Google Assistant. It's been in beta since 2018, restricted to US English, and Google's December 2025 docs still list it as beta with no general-availability timeline.
John Mueller confirmed in late 2025 that Google is pruning structured data types that aren't pulling weight, and Speakable appears on community shortlists of candidates for removal in the January 2026 cleanup. Time invested in Speakable markup is time not invested in answer-first prose, which works across every AI system regardless of schema support.
The leverage is in the writing, not the markup.
Verify before shipping
The isitready.dev audit flags pages where the first paragraph after each H2 misses answer-first structure, runs longer than 60 words, or buries the noun phrase from the heading. It also checks question-style heading coverage and entity signals in opening paragraphs — the same signals Authoritas tied to a 38% lift in AI Overview inclusion. Run it against your top 20 indexable pages, fix the first paragraphs the report flags, re-crawl, watch citations climb.