Comparison structure is the strongest content signal for AI citation (d = 0.43). First-person writing is the strongest negative signal (d = -0.30 to -0.37). This guide covers every page-level feature, content format, and optimization tactic backed by two peer-reviewed studies totaling 29,849 queries and 10,293 position-matched pages.
AI search engines do not read content the way humans do. They parse pages, identify extractable units, and decide in real time whether your page answers the user's question well enough to cite. The pages they choose look nothing like a typical "SEO-optimized" article built for Google's blue links.
Two studies ground every recommendation in this guide. Lee (2026a) analyzed 19,556 queries across 8 verticals and 4 AI platforms, crawling 4,658 pages. Lee (2026c) extended that work with a position-controlled study of 10,293 pages, 66 features, and 250 queries across 3 AI platforms, isolating page-level effects by comparing equally ranked pages. The combined evidence reveals a clear hierarchy of what matters, and most of it has nothing to do with keywords or backlinks.
The Bottom Line: Writing for AI citation is not about gaming an algorithm. It is about structuring information so clearly that an AI model can extract, attribute, and cite it with confidence. For a broader strategic framework, see our generative engine optimization guide.
🏗️ THE PAGE-LEVEL FEATURES THAT PREDICT AI CITATION
Lee (2026c) tested 66 page features across all four Google position bands (positions 1-3, 4-7, 8-12, and 13-20). Six features predicted citation in ALL FOUR bands, meaning the pattern held regardless of where a page ranked on Google. (The earlier study, Lee 2026a, identified 7 features using a different methodology without position control.) These are the features you can control.
| Rank | Feature | Effect Size (Cohen's d) | Direction | Model Importance |
|---|---|---|---|---|
| 1 | Comparison structure | d = 0.43 (medium) | Positive | Strongest signal |
| 2 | Query term coverage | d = 0.42 (medium) | Positive | High |
| 3 | Blog/opinion tone | d = -0.30 to -0.37 (medium) | NEGATIVE | 8.1% importance |
| 4 | Primary source score | d = 0.27 (small-medium) | Positive | 4.1% importance |
| 5 | Word count | d = 0.20 (small) | Positive | 7.1% importance |
| 6 | Subheading depth (H3s) | d = 0.19 (small) | Positive | 5.9% importance |
A quick note on what "Cohen's d" means: it measures how big the difference is between cited and not-cited pages. A d of 0.43 means there is a clear, visible gap between the two groups. In social science, anything above 0.3 is considered a "medium" effect. Anything above 0.2 is "small but real."
What about the features everyone recommends?
Several commonly recommended AI optimization features showed zero effect within position bands:
| Feature | Common Advice | What the Data Showed |
|---|---|---|
| Page speed | "Make your site faster!" | Not significant in ANY position band (p > 0.39) |
| Author bylines | "Add author bios for trust!" | No consistent effect |
| Readability scores | "Write at an 8th grade level!" | No effect |
| Review schema | "Add Review markup!" | No effect within bands |
| Product schema | "Add Product markup!" | No effect within bands |
Page speed is the most dramatic reversal. It appeared to be the most important factor in earlier cross-domain analysis (65.8% of the model). After controlling for which website a page came from, the effect completely disappeared. Zero signal.
The Bottom Line: Page structure is the most valuable thing you can add beyond ranking well on Google. Content quality features (statistics, technical depth) do not add anything beyond what structure already captures, because well-structured pages tend to also be high-quality pages.
📊 COMPARISON FORMAT DEEP DIVE: WHY d = 0.43 ACROSS ALL INTENTS
Comparison structure is the single strongest content signal for AI citation. Pages that compare things (side-by-side tables, "Product A vs Product B" breakdowns, feature grids) get cited more than pages without comparison elements. And here is the surprise: this effect holds for ALL query types, not just comparison queries.
Even when someone asks a simple informational question or a "best of" question, pages with comparison structure get cited more. The likely reason: comparison tables are packed with specific, extractable facts that AI models can easily pull out and cite. A comparison table is pre-segmented data. The model does not need to parse three paragraphs of narrative to find the answer.
What makes a comparison table AI-extractable
Not all tables work equally well. The difference between a table that gets cited and one that gets ignored is specificity and self-containment.
| Table Quality | Example Cell | AI Extractability |
|---|---|---|
| Vague labels | "Good," "Fair," "Poor" | Low: no concrete data to cite |
| Specific values | "$49/mo," "14-day trial," "99.9% uptime" | High: discrete facts the model can attribute |
| Missing context | Raw numbers with no units | Low: ambiguous without interpretation |
| Self-contained cells | Each cell answers a question independently | High: extractable without reading the full table |
Use semantic HTML (<table>, <thead>, <th>, <td>) rather than CSS grid layouts or image-based tables. AI crawlers parse HTML tables directly, and clean table markup increases content-to-HTML ratio, which itself predicts citation (cited pages: 0.086 median ratio vs. 0.065 for non-cited).
Table types and when to use them
| Table Type | Best For | Query Intent Served |
|---|---|---|
| Feature comparison | Product or service evaluation pages | Discovery (31.2% of autocomplete queries) |
| Spec table | Technical or product detail pages | Informational (61.3% of autocomplete queries) |
| Pricing grid | Any page discussing costs | Discovery, Validation |
| Pros/cons matrix | Review or recommendation pages | Validation (3.2%), Comparison (2.3%) |
| Decision matrix | Advisory or strategy pages | Discovery, Informational |
The Bottom Line: If you publish any content that compares products, tools, services, or approaches, put the comparison in a table with specific values in every cell. Narrative comparisons buried in paragraphs are dramatically less extractable. For a deeper look at what gets cited and why, see What Gets You Cited by AI, Explained.
📝 LISTICLE FORMAT: 13.75 LIST SECTIONS ON CITED PAGES
The Sellm 400K-page study of AI-cited content found that cited pages contain a median of 13.75 list sections per page, roughly 17 times more than typical uncited pages. This finding aligns with the broader pattern: content broken into discrete, labeled units is more extractable than continuous prose.
But not any list format works. The difference between a high-citation listicle and a low-citation one is data density per entry.
A listicle that says "10 Best CRMs for 2026" with one vague paragraph per entry is not a citation magnet. A listicle where each entry includes pricing, feature specs, a "Best for" recommendation, and honest pros/cons is. Each entry in a data-rich listicle functions as a self-contained answer to a potential query.
What makes a listicle data-rich
| Element | Purpose | Example |
|---|---|---|
| Specific pricing | Enables cost comparison queries | "$29/mo starter, $79/mo pro, custom enterprise" |
| Quantified features | Provides extractable specs | "Up to 10,000 contacts, 5 pipelines, 50GB storage" |
| "Best for" framing | Matches discovery intent | "Best for: agencies managing 10+ client accounts" |
| Pros/cons per entry | Gives evaluation signals | "Pro: unlimited users. Con: no mobile app yet." |
| Star or numerical rating | Quick comparative signal | "4.7/5 based on 2,300 reviews" |
Structure each entry as: "Best for" statement first, then a two-sentence overview, then key specs as a bulleted list, then pros/cons. Each entry built this way becomes a self-contained, extractable unit the AI model can cite independently.
"Best for" framing matches 31.2% of all queries
Discovery intent accounts for 31.2% of autocomplete queries, and these queries follow a predictable pattern: "best [product] for [use case]." Content that explicitly uses this framing gets cited because it directly matches the query structure.
| Instead of This | Write This |
|---|---|
| "Notion is a versatile productivity tool." | "Best for: teams of 5 to 50 that need docs, wikis, and project management in one workspace." |
| "HubSpot offers many CRM features." | "Best for: small businesses that want free CRM with optional paid upgrades for marketing automation." |
| "This protein powder is high quality." | "Best for: post-workout recovery for strength athletes training 4+ days per week." |
The Bottom Line: Lists are not inherently better than narrative. Lists with embedded data are. Each list entry should function as a standalone, citable answer to a potential query.
❓ FAQ FORMAT: MATCHING THE CHATBOT INTERFACE
FAQ sections occupy a unique position in the format hierarchy. They present information in the exact structure that AI chatbot interactions use: question followed by answer.
Lee (2026a) found that 61.3% of autocomplete queries are informational, and users phrase these as questions: "How does X work?" "What is the difference between Y and Z?" FAQ sections with properly implemented FAQPage schema provide content pre-structured into the same question-answer units the AI model needs to generate its response.
The schema markup matters. FAQPage schema is the only schema type consistently significant across all 4 position bands in Experiment M. Pages with FAQ schema receive approximately 2x more recrawls from AI-affiliated bots. More crawls mean a fresher index entry, which compounds the citation advantage.
How to write FAQ questions AI will use
The questions need to match real query patterns, not generic questions you wish people would ask.
| Question Source | Why It Works | Example |
|---|---|---|
| Google Autocomplete | Reflects real search behavior | "How long does [process] take?" |
| People Also Ask boxes | Algorithmically validated questions | "Is [product] worth the price?" |
| AI chatbot testing | Shows what follow-ups the model generates | "What are the downsides of [approach]?" |
| Reddit/forum threads | Natural language phrasing | "Has anyone tried [solution] for [problem]?" |
Structure each answer in two to three sentences that directly address the question, followed by one to two sentences of supporting context. AI models favor concise, self-contained answers over long narrative responses that require parsing.
The Bottom Line: FAQ sections with schema are the only content format that simultaneously matches the dominant query type (informational, 61.3% of autocomplete queries), provides a machine-readable structure, and attracts more frequent AI bot crawls. Triple advantage.
🗣️ WRITING TONE: WHY FIRST-PERSON HURTS AI CITATION
Blog/opinion tone is the strongest NEGATIVE signal in the entire study (d = -0.30 to -0.37, 8.1% model importance). Pages written in first person ("I think," "in my experience," "I tested") get cited less. AI platforms prefer pages that read like reference material, not personal blogs.
This does not mean personal experience is worthless. It means the way you frame information determines whether AI treats your page as a citable source or as an opinion piece. The same information written in an objective tone ("testing showed that...") versus a personal tone ("I found that...") gets treated differently by retrieval systems.
How to fix your writing tone
| First-Person (Lower Citation) | Third-Person Objective (Higher Citation) |
|---|---|
| "I tested 15 CRMs and found HubSpot was fastest to set up" | "Testing across 15 CRM platforms showed HubSpot had the fastest onboarding time at under 15 minutes" |
| "In my experience, the free plan is too limited" | "The free plan limits users to 500 contacts and 1 pipeline, which constrains teams managing more than 50 active deals" |
| "I think this tool is best for small teams" | "This tool is best suited for teams of 3 to 10 people managing fewer than 100 client accounts" |
The shift is subtle but measurable. Each rewrite does three things: removes the first-person framing, adds specific numbers, and makes the claim independently verifiable. AI models can cite "testing showed X" as a factual statement. They struggle to cite "I found X" because it is an opinion.
Claude is the most tone-sensitive platform. It penalizes pure marketing language by a factor of 0.8x and boosts content with honest limitations by a factor of 1.7x. Honest limitations sections satisfy Claude. Recent dates and fresh data satisfy Perplexity. Clean semantic HTML with schema markup helps across all platforms.
Include honest limitations sections
This is the tactic most marketers resist, and one of the most powerful. Content that acknowledges limitations, trade-offs, and situations where a product or approach is not the right fit signals objectivity.
Write limitations with specificity:
- Weak: "There are some limitations." (not citable)
- Strong: "This approach struggles with datasets larger than 10GB." (citable)
- Weak: "Can be expensive." (not citable)
- Strong: "Pricing starts at $49/month per seat, which adds up for teams larger than 10." (citable)
The Bottom Line: The page that says "here is when NOT to use our product" gets more AI citations than the page that says "our product is perfect for everyone." For technical implementation details, see our technical SEO for AI citations guide.
🔄 RETROFITTING EXISTING CONTENT: THE PRIORITY CHECKLIST
You do not need to start from scratch. The fastest path to AI search visibility is retrofitting pages you already have. Your published pages already have indexed authority, traffic data, and content substance. Retrofitting structure is faster than creating new substance.
Lee (2026a) identified 7 page-level features that statistically predict AI citation. Every one can be retrofitted without rewriting from scratch.
The 7-step retrofit framework (priority order)
| Priority | Retrofit Action | Effect | Time Per Page |
|---|---|---|---|
| 1 | Add self-referencing canonical tag | OR = 1.92 (nearly 2x citation odds) | 2 minutes |
| 2 | Switch to correct schema type | Product OR = 3.09; Article OR = 0.76 (hurts) | 15 to 30 minutes |
| 3 | Improve content-to-HTML ratio to 0.08+ | Cited median 0.086 vs non-cited 0.065 | 30 to 60 minutes |
| 4 | Add FAQ section with FAQPage schema | FAQPage OR = 1.39; 2x more recrawls | 20 to 40 minutes |
| 5 | Front-load key findings | 44.2% of ChatGPT citations come from first 30% of content | 15 to 30 minutes |
| 6 | Add comparison tables | d = 0.43, strongest content signal | 15 to 30 minutes |
| 7 | Update dateModified in schema and visible text | Freshness signal for index-based platforms | 5 minutes |
Schema type selection matters more than schema presence
Schema markup presence alone has an odds ratio of just 1.02 (p = 0.78), meaning it is statistically meaningless. But the TYPE of schema changes everything:
| Schema Type | Odds Ratio | Recommendation |
|---|---|---|
| Product | 3.09 (p < 0.001) | Use on any page with product information |
| Review | 2.24 (p < 0.001) | Use on pages with genuine review content |
| FAQPage | 1.39 (p < 0.05) | Use on any page with Q&A content |
| Organization | 1.08 (p = 0.35) | No measurable effect |
| Breadcrumb | 0.99 (p = 0.97) | No measurable effect |
| Article | 0.76 (p < 0.05) | CAUTION: negative association |
Article schema on a non-editorial page can suppress AI visibility. If your page contains product data, comparisons, or FAQs, use the schema type that matches your actual content. Reserve Article schema for genuinely editorial content.
Content-to-HTML ratio: common problems and fixes
| Problem | Typical Ratio Impact | Fix |
|---|---|---|
| Inline CSS on every element | Drops 15 to 25% | Move to external stylesheet |
| 5+ nested wrapper divs | Drops 10 to 20% | Simplify to semantic HTML |
| Mid-content ad blocks (3+) | Drops 10 to 15% | Reduce to 1 to 2 placements |
| Large inline SVGs or base64 images | Drops 5 to 15% | Use external image files |
| Cookie/consent banners in HTML | Drops 5 to 10% | Load via deferred JavaScript |
Retrofit audit scoring system
Score each page against the 7 priorities to batch-process your content library:
| Check | Points | Pass Criteria |
|---|---|---|
| Self-referencing canonical present | 20 | Canonical URL matches page URL exactly |
| Correct schema type (not just Article) | 20 | Product, Review, FAQPage, or HowTo present |
| Content-to-HTML ratio >= 0.08 | 15 | Measured via Screaming Frog or manual check |
| FAQ section with FAQPage schema | 15 | Minimum 3 questions with structured data |
| Key finding in first 30% of content | 10 | Core insight in paragraphs 1 to 3 |
| At least 1 comparison table | 10 | Semantic HTML table with headers |
| dateModified updated within 90 days | 10 | Schema and visible date reflect recent update |
| Total possible | 100 |
Pages scoring below 40 with high traffic (1,000+ visits/month) are your highest-ROI retrofit candidates. To check your current pages, try the free AI visibility scanner. For a professional audit of your full content library, see our AI SEO services.
The Bottom Line: Retrofitting gives you the same structural advantages as new content in a fraction of the time, with the added benefit of existing authority and traffic.
🎯 CONTENT FORMAT BY INTENT TYPE: THE STRATEGIC MAPPING
Format selection is not a creative decision. It is a strategic one driven by query intent. Lee (2026a) identified five query intent types, and each draws citations from completely different source types and formats.
| Intent Type | Query Share | Winning Format | Schema to Use | Example Query |
|---|---|---|---|---|
| Informational | 61.3% of autocomplete queries | FAQ sections, spec tables, comprehensive guides | FAQPage, Product | "How does containerization work?" |
| Discovery | 31.2% of autocomplete queries | Comparison tables, "Best for" listicles | Product, Review | "Best project management tool for startups" |
| Validation | 3.2% | Pros/cons lists, case studies, honest reviews | Review | "Is Webflow worth it for small business?" |
| Comparison | 2.3% | Side-by-side comparison tables | Product | "Shopify vs WooCommerce for dropshipping" |
| Review-seeking | 2.0% | Detailed reviews with ratings | Review | "Notion reviews from real users" |
Informational and discovery together account for 92.5% of all queries. If you are choosing where to invest format effort, start with FAQ sections and comparison tables. Build your page targeting discovery intent, then add an FAQ section targeting informational intent. This covers the vast majority of the query space with a single page.
What source types win by intent
| Intent Type | What Gets Cited | What Gets Ignored |
|---|---|---|
| Informational | Tutorials, Wikipedia, .gov/.edu, comprehensive guides | Product pages, marketing copy |
| Discovery | Review aggregators, listicles, comparison pages | Single-product pages, brand sites |
| Validation | Reddit (web UI), brand sites, case studies | Generic articles without real opinions |
| Comparison | Publisher reviews, media sites | Brand sites (conflict of interest) |
| Review-seeking | YouTube, TechRadar, Reddit | Manufacturer pages |
A comparison page will never get cited for an informational query, no matter how well-optimized it is. Intent mismatch is the most common reason good content gets zero AI citations.
Review aggregators: how to compete
Review aggregators (G2, Capterra, TrustRadius) earn 16 to 21% of discovery citations across AI platforms. You cannot outcompete them on review volume. But you can win on specificity. A review aggregator covers "Best CRM Software" generically. Your comparison page can target "Best CRM for Real Estate Teams Under 20 People" with far more specific recommendations.
The Bottom Line: Format without intent matching is wasted effort. Match the format to the intent, or accept that your content will be skipped regardless of quality.
📋 THE COMPLETE OPTIMIZATION CHECKLIST
Before publishing any content intended for AI citation, verify these elements:
| Checkpoint | Target | Priority |
|---|---|---|
| Comparison or spec table with specific values | At least 1 semantic HTML table | Critical |
| FAQ section with FAQPage schema | 3 to 5 questions from real query patterns | Critical |
| Key finding in first 30% of content | Core insight in paragraphs 1 to 3 | Critical |
| Third-person objective tone | No "I think," "in my experience" framing | Critical |
| "Best for [Use Case]" framing | Explicit use-case recommendations per item | High |
| Pros/cons with quantified trade-offs | Specific numbers in every evaluation | High |
| Honest limitations section | Acknowledge when the product is not the right fit | High |
| Correct schema type | Product, Review, or FAQPage (not Article on structured pages) | High |
| Self-referencing canonical tag | Canonical URL matches page URL exactly | High |
| Word count around 2,000 | Comprehensive but not bloated | Medium |
| H3 subheadings throughout | Roughly twice as many as a typical page | Medium |
| Content-to-HTML ratio >= 0.08 | Clean semantic HTML, minimal wrapper divs | Medium |
| Internal-to-external link ratio at least 2:1 | Excessive outbound links drop OR to 0.47 | Medium |
| dateModified updated after real edits | Freshness signal for Perplexity and Gemini | Lower |
❓ FREQUENTLY ASKED QUESTIONS
Does traditional SEO still matter if I am optimizing for AI citation?
Yes, but with important caveats. Google rank itself predicts AI citation at just 7.8% for ChatGPT and 6.8% for Claude at the query level. However, Experiment M showed that pages in Google's top 3 get cited 57.2% of the time versus 13.5% for positions 13 to 20. Google rank is the gateway (roughly 30% of prediction); page content is the selector (roughly 70%). The foundational practices of SEO (clean HTML, proper schema, logical structure) overlap with what AI crawlers need.
How long does it take for content changes to show up in AI citations?
ChatGPT and Claude perform live page fetches, meaning they see your current content in real time. Perplexity and Gemini use pre-built indices that update on their own schedules. Expect changes reflected in ChatGPT and Claude responses within hours, but allow two to four weeks for Perplexity and Gemini to re-index.
Why does comparison structure help even for non-comparison queries?
The d = 0.43 effect held across ALL four position bands and ALL intent types, not just comparison queries. The likely explanation: comparison tables are dense with specific, extractable facts. A feature grid has discrete values in every cell. AI models find it easier to pull a concrete data point from a table cell than to parse three paragraphs of narrative for the same information.
Should I remove Article schema from all my pages?
Not from genuinely editorial pages (op-eds, news commentary, personal essays). Article schema is appropriate there. The OR = 0.76 finding means it hurts citation odds on pages that could use a more specific type. Product comparison pages, FAQ resources, service pages, and review content should switch to Product, Review, FAQPage, or HowTo schema.
Can I optimize one page for multiple query intents?
Yes, with clear structural separation. A page can serve both informational and discovery intents if it includes distinct sections for each: an FAQ section for informational queries and a comparison table for discovery queries. Together, those two intents cover 92.5% of the query space. The key is using explicit formatting (headers, schema, table structures) so the AI model can extract the relevant section for each query type.
How many pages should I retrofit at once?
Start with your top 10 highest-traffic pages and work through all 7 retrofit priorities on each one before moving to the next batch. Batching by page rather than by priority across all pages is more efficient because you only open each page's code once. Most teams can retrofit 5 to 10 pages per week once the process is established.
What is the single highest-impact change for a page that currently gets zero AI citations?
Add a comparison table with specific values and switch from Article schema to Product or FAQPage schema. These two changes address the strongest positive signal (comparison structure, d = 0.43) and remove the strongest negative schema signal (Article, OR = 0.76) in a single editing pass.
Does first-person writing always hurt, or only in certain contexts?
First-person writing (d = -0.30 to -0.37) consistently reduces citation probability across all position bands. However, the effect is about framing, not content. The same information rewritten in third-person objective tone ("testing showed that...") performs better than first-person ("I found that..."). Review-seeking queries (2.0% of queries) are the one context where personal experience content has some advantage, but even there, specific data points outperform subjective opinions.
📚 REFERENCES
Lee, A. (2026a). "Query Intent, Not Google Rank: What Best Predicts AI Citation Behavior." AI+Automation Research, Preprint v5. DOI: 10.5281/zenodo.18653093. Dataset: 19,556 queries across 8 verticals, 4 AI platforms, 4,658 crawled pages.
Lee, A. (2026c). "I Rank on Page 1: What Gets Me Cited by AI? Position-Controlled Analysis of Page-Level and Domain-Level Predictors of AI Search Citation." AI+Automation Research. Paper: aixiv.science/abs/aixiv.260403.000002. Dataset DOI: 10.5281/zenodo.19398158. 10,293 pages, 66 features, 250 queries, 3 AI platforms.
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). "GEO: Generative Engine Optimization." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. DOI: 10.48550/arXiv.2311.09735. (Note: specific content-level features do not replicate on production platforms.)
Sellm, J. (2024). "Large-Scale Analysis of AI-Cited Web Content: Structural Features of 400K Pages." Preprint.
Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). "Detecting Hallucinations in Large Language Models Using Semantic Entropy." Nature, 630, 625-630. DOI: 10.1038/s41586-024-07421-0