These are writing and layout patterns for pages that AI engines (Google AI Overviews/AI Mode, ChatGPT search, Perplexity, Copilot, Claude) quote and link to. Every pattern also helps human readers, and none replaces indexability (see the seo SKILL).
1. What the evidence supports
| Source | Finding | Implication |
|---|---|---|
| Aggarwal et al., GEO, KDD 2024 (10K queries; also tested on Perplexity) | Cite sources, add quotations, add statistics raised visibility 30–40%. At rank 5: citations +115%, quotations +100%, statistics +98%. At rank 1 the same edits lost 20–30%. Keyword stuffing and "authoritative" tone: no gain | Add evidence, especially on pages that aren't already #1 |
| Puerto et al., C-SEO Bench, NeurIPS 2025 | Most content-rewriting tricks are ineffective or harmful, and gains shrink as competitors adopt them. Traditional SEO is significantly more effective | Don't "rewrite for AI". Rank first |
| Chen et al., 2025 (arXiv 2509.08919) | AI search strongly favors earned media (third-party, authoritative) over brand-owned pages. Big-brand bias | Get covered, reviewed and cited elsewhere (off-site) |
| Citation Selection to Citation Absorption, 2026 (21K citations) | High-influence cited pages are longer, more structured, and rich in definitions, numerical facts, comparisons and procedural steps | Structure plus facts |
| Microsoft Ads, "Optimizing content for AI search answers" (2025-10) | Title/description/H1 alignment; descriptive H2/H3; Q&A pairs ("lifted word for word"); lists and tables; self-contained 1–2 sentence answers. Avoid walls of text, answers hidden in tabs, facts only in images or PDFs, decorative symbols | Snippable sections |
| Bing Webmaster (2025-12, 2026-02) | Near-duplicate URLs get clustered and one is chosen. Freshness via IndexNow. Evidence, examples and cited sources | Consolidate, stay fresh |
| Google, "Optimizing for generative AI features" (2026-05) | No special requirements. Non-commodity content and a unique point of view win. llms.txt, chunking, "AI rewriting" and special schema are unnecessary | Be genuinely useful and original |
2. Patterns
2.1 Answer first (every page and every question-style section)
The first paragraph after the h1 (and after each h2 phrased as a question) answers in 1–2 self-contained sentences, ≤ 320 characters, naming the entity instead of using a pronoun.
<!-- ❌ -->
<h1>Welcome to the future of builds</h1>
<p>We're passionate about helping teams move faster. It's built for you.</p>
<!-- ✅ -->
<h1>Acme: incremental TypeScript builds for monorepos</h1>
<p>Acme is an open-source build tool that caches TypeScript compilation per package, cutting CI builds in a
200-package monorepo from 14 minutes to 90 seconds. It is MIT licensed and runs on Node 22+ and Bun.</p>2.2 Numbers with units, dates and sources
<p>Median mobile JavaScript is 646 KB per page
(<a href="https://almanac.httparchive.org/en/2025/page-weight">HTTP Archive Web Almanac, 2025</a>).
Acme's landing page ships 142 KB, measured on 2026-09-10 with Lighthouse 13.</p>- Every claim that could be doubted gets a number, a unit, a date or a link to a primary source (docs, papers, standards, official statistics). No links to aggregator blogs.
- Roughly 1 outbound primary-source link per ~300 words of factual prose.
- Own benchmarks: state the method (hardware, versions, runs) next to the number.
2.3 Quotations from named people
<blockquote cite="https://example.com/interview">
<p>"We replaced three CI jobs with one Acme cache and cut our monthly runner bill by 38%."</p>
<footer>— <cite>Dana Li, Staff Engineer at Example Corp</cite></footer>
</blockquote>Real, attributable and permissioned. No invented testimonials.
2.4 Question headings + Q&A blocks (plain HTML, no FAQPage schema)
<section aria-labelledby="q-pricing">
<h2 id="q-pricing">How much does Acme cost?</h2>
<p>Acme is free for open-source projects. The Team plan costs $12 per seat per month billed annually, and includes remote caching.</p>
</section>Use the questions real users ask (support tickets, sales calls, search console queries). Keep the answers visible, not collapsed. If you use , set open for the key facts.
2.5 Comparison and spec tables
<table>
<caption>Acme vs. tsc --build (200-package monorepo, cold CI, 2026-09)</caption>
<thead><tr><th scope="col">Metric</th><th scope="col">Acme 3.2</th><th scope="col">tsc 5.9</th></tr></thead>
<tbody>
<tr><th scope="row">Cold build</th><td>6 min 10 s</td><td>14 min 2 s</td></tr>
<tr><th scope="row">Warm build (1 file changed)</th><td>1.5 s</td><td>38 s</td></tr>
<tr><th scope="row">License</th><td>MIT</td><td>Apache-2.0</td></tr>
</tbody>
</table>Real The exact same name everywhere: Product home: h1 (what + for whom) → answer paragraph → primary CTA → 3–5 benefit sections, each an h2 claim + a measured proof → comparison table → pricing summary in text → named customer quotes → FAQ-style h2 questions → footer fact block. Docs page: h1 = task or concept → one-sentence definition or outcome → prerequisites list → numbered steps with code → expected output → troubleshooting h2 questions → "Updated" date → Article: h1 → byline + updated date → TL;DR answer (2–3 sentences) → sections with question or claim headings → data with sources → quotes → conclusion restating the answer → author bio linked to a ProfilePage. AI engines lean on third-party sources. Maintain accurate profiles (GitHub, LinkedIn, Crunchbase, Wikipedia where notable), get listed in credible comparisons and directories, publish benchmarks others can cite, and answer in communities. with
, never a grid of divs. Compare fairly and name what the competitor does better.
2.6 Definitions and steps
with one action per step and the expected result.2.7 Entity clarity
, h1, og:site_name, JSON-LD Organization.name and WebSite.name, the llms.txt H1, footer. Keep a "What it is / Who it's for / Pricing / License / Status" fact block on the home or about page. Link official profiles through sameAs.2.8 Freshness
that matches JSON-LD dateModified and sitemap lastmod.next-metadata.md §9).2.9 Visible text, not pictures of text
2.10 Readable chunks
3. Page templates
.md mirror.4. Don't
speakable spam.5. Off-site (outside the codebase, tell the owner)
6. Measure
utm_source=chatgpt.com, and from perplexity.ai, copilot.microsoft.com, claude.ai.void geo reports these heuristics as info findings (answer-first, citation density, question headings, tables, entity consistency). They are review prompts, never CI failures.Sources