The Anatomy of an AI-Citable Page

L4 Tactics: What the data says about how to structure a page AI will actually cite

Post 7 in the AI Visibility Framework series.

Post 6 mapped what to do at each layer. This post goes deep on Layer 4, aka the page itself. What makes AI retrieval systems choose your page over the other 8 in the grounding context.

A reminder that this isn’t theory. Every spec below traces to a measured finding in our research database. 13 on-page specifications, each with the data behind it. No, trust me, bro, here.

If you build pages that follow these specs, you are engineering for citation. If you don’t, you’re hoping.

Why L4 is different from SEO on-page optimization

Traditional on-page SEO optimizes for ranking. L4 optimizes for citation: being selected from the retrieval pool and having your sentences extracted into the AI’s answer.

The distinction matters because the systems work differently. Google ranks pages. AI retrieval systems rank passages. Your page can rank #1 on Google and still never get cited because the passages inside it aren’t extractable.

SEO ranking factors explain only 4-7% of AI citations (Profound, Feb 2026). The other 93-96% is determined by content structure, entity density, formatting, and how well your passages serve the model’s generation needs.

Some of that can happen naturally with an SEO-only focus because there is overlap. But the “GEO is just SEO” narrative fails when you look under the hood. There are distinct AI pick-ups. Why not optimize for both?

Here’s the full anatomy.

ZONE 1: The First 30%

This is where citation lives or dies.

44.2% of all ChatGPT citations come from the first 30% of page text. Content buried in paragraph 12 is 2.5x less likely to be cited (@Kevin_Indig / Gauge, 1.2M citations, p=0.0). This isn’t a preference. Oh, no, it’s architectural. OpenAI’s embedding system (Matryoshka Representation Learning) front-loads critical semantic information into the first vector dimensions. During fast retrieval at web scale, vectors get truncated. If your core thesis is below the fold, it may be literally cut from the candidate set before the model evaluates it (@cyberandy / WordLift, Mar 2026).

What goes here:

Your primary finding, thesis, or answer. Not background. Not “in this article we’ll explore…” Not a history of the topic. The answer. Show the goods.

Lay out the data point or original insight that makes your page worth citing. If you ran a study, the result goes here. If you have a framework, the framework goes here. If you’re comparing options, the comparison verdict goes here.

Entity-dense opening. Cited text has 20.6% entity density vs 5-8% in normal English. “The top CRM platforms for mid-market teams are Salesforce, HubSpot, and Pipedrive, each serving different use cases” beats “There are several popular CRM options available for businesses of various sizes.”

Every sentence in Zone 1 should be extractable on its own. No “as mentioned above.” No “this approach.” Each sentence carries its own meaning clearly.

ZONE 2: Answer Capsules

72.4% of ChatGPT-cited posts use this structure: a question as an H2 heading, followed immediately by a self-contained answer of roughly 120-150 characters (@Kevin_Indig / SEL, Nov 2025). Confirmed at scale — 78.4% of citations containing questions come from H2 headings (Gauge, 1.2M citations).

The model treats the H2 as the user’s question and the following paragraph as the answer. Entity echoing – repeating the entity from the heading in the first word of the answer – is a measurable citation signal.

The spec:

Write H2s as questions your audience actually asks. Not rhetorical questions. Not keyword-stuffed headings. Direct questions.

Follow each H2 immediately with a declarative answer. One to two sentences. No hedging. “X is defined as…” not “X could potentially be considered…”

The median cited sentence is 10 words. The maximum observed is 17 words. Zero mid-sentence fragments — AI extraction respects sentence boundaries (Shashko / Bright Data, 42,971 citations, 6 platforms). Write short, declarative sentences for your key claims.

No links inside the answer capsule. Over 90% of cited capsules contained zero links.

Example:

H2: What is the average conversion rate from AI search traffic?

AI search traffic converts at 14.2% compared to Google organic at 2.8% — a 5.1x difference (Exposure Ninja, Mar 2026). Claude users convert the highest at 16.8%, followed by ChatGPT at 14.2% and Perplexity at 12.4%.

That’s an answer capsule. Question heading. Declarative answer. Specific entities and numbers. Under 150 characters for the core claim. Extractable without any surrounding context.

ZONE 3: Self-Contained Sections

AI retrieval systems don’t read pages. They chunk them.

iykyk

Google’s Vertex AI Search uses a default chunk size of 500 tokens (~375 words). Each chunk is evaluated independently for relevance to the query. If a section depends on a previous section for meaning, it fails when extracted alone (@dejanseo / Dejan AI, Dec 2025 — 7,060 queries, confirmed by Google Vertex AI documentation).

Each Gemini grounding query has a budget of roughly 2,000 words total, distributed by relevance rank: the #1 source gets 531 words, #2 gets 433, and #3 gets 378. An 800-word page gets 50%+ grounding coverage. A 4,000-word page gets 13%.

The spec:

Each section between H2s should be extractable without context from the surrounding sections. If you reference “the approach above” or “as we discussed,” that section fails chunk-level retrieval.

Optimal section length: 120-180 words between headings for ChatGPT citation (SE Ranking, 129K domains). 100-150 words for AI Mode (slightly narrower). Sections under 50 words = 70% fewer citations. Over 150 words see diminishing returns.

Density beats length. Always. A focused 800-word page outperforms a comprehensive 4,000-word guide for AI grounding coverage. The 4,000-word guide might rank better on Google — but ranking and citation are different systems.

Structured pages produce 2.3x the sentence-match rate of unstructured pages (CI28). 98.1% of cited pages had lists. Structure isn’t cosmetic – it’s how the retrieval system parses your content.

ZONE 4: The Voice That Gets Cited

Not all writing styles get cited equally.

The citation sweet spot is a subjectivity score of 0.47 on a 0-1 scale (@Kevin_Indig / Gauge, 1.2M citations). Not pure facts (0.1). Not pure opinion (0.9). Analyst voice – facts with applied analysis. “The data shows X, which means Y for teams doing Z.”

Readability is bimodal, not linear (Shashko, 11,672 sentences). AI cites both very simple content (Flesch 90-100) and very technical content (Flesch below 30) at roughly equal rates. The dead zone is Flesch 50-59 — corporate jargon, hedged language, committee prose. Only 5% of citations come from that range.

The spec:

Match readability to query intent. Consumer/informational content: write at Flesch 70-100 (plain language). Technical/specialist content: write at Flesch below 40 (precise terminology). Never write in the corporate middle. Stand out, but do it correctly.

Use definitive language. Cited text is 2x more likely to contain “is defined as” or “refers to” than non-cited text. Declarative beats hedged. “X is Y” beats “X could potentially be considered Y in some contexts.” This is literally defined as the best X article ever written.

Content should resemble what the model would answer, not what users search for. Embedding models are shifting from encoding queries to encoding answers — content that already looks like an LLM response gets higher similarity scores in retrieval (McGill NLP, Mar 2026). Write like an answer, not like a page. Reminder: I’m stating the facts here based on data and how you can work with them to optimize your content for retrieval. I am not saying this is the best way to write purely as a human. Your call. My personal approach is a blend.

ZONE 5: Technical Signals

The invisible layer that determines whether AI even considers your page.

Schema decisions:

This one is counterintuitive. Generic schema markup (Article, Organization, BreadcrumbList) actively hurts AI citation 🤯 — 41.6% citation rate vs 59.8% for pages with no schema at all. Only attribute-rich schema with full specifications outperforms: Product/Review schema with pricing, ratings, and specs = 61.7% (Growth Marshal, 730 citations, Feb 2026).

ChatGPT doesn’t parse JSON-LD at all (@dejanseo / DEJAN). The schema value is Google-ecosystem only. If your schema is CMS-default boilerplate, it may be suppressing your citation rate. Either implement an attribute-rich schema with complete data or remove the generic markup entirely.

Author schema specifically: must include full Person schema with name, jobTitle, affiliation, sameAs links to LinkedIn/Twitter/Wikipedia, and knowsAbout fields. A bare name-only author tag is generic schema and may hurt, per the same study.

Meta descriptions as AI advertisement:

AI platforms use title + description + URL to decide whether to fetch the page at all. Meta descriptions are not a Google ranking factor but directly influence AI citation (Profound + @iPullRank, Feb 2026). Your meta description is the advertisement to the LLM — write it for the model, not just for human searchers. Welcome back, old friend.

Semantic URLs with high keyword similarity get 11.4% more citations. URL slugs of 17-40 characters correlate with peak citation rates (Semrush, 5M URLs). Longer, descriptive slugs outperform keyword-only slugs. Do NOT change existing URLs (catastrophic for search)…. apply to new content only.

Freshness cadence (platform-specific):

One refresh schedule doesn’t fit all platforms. ChatGPT citations skew 458 days fresher than organic results. 76.4% of most-cited pages updated within 30 days. AI Mode is the opposite — median cited page is 2.2 years old, with 52.8% being 2+ years old (Shashko, Mar 2026).

For ChatGPT and Perplexity targets: monthly refresh minimum. For AI Mode targets: invest in depth and authority, not freshness. Content updated within 3 months = 2x citation rate vs outdated content for ChatGPT specifically (SE Ranking, 129K domains). But for AI Mode, the same refresh has only 28% lift.

Update: magnitude matters too. Adding 31-100% new content = +8 Google ranking positions (p=0.026). Minor tweaks (0-10% change) are wasted effort. Moderate updates (11-30%) are actively harmful – worse than not updating at all (RepublishAI, 14,987 URLs).

Server response time:

Pages that respond slowly get a 499 (client closed request) from AI platforms – they don’t have a persistent index, they fetch in real time. First Contentful Paint under 0.4 seconds = 3x more ChatGPT citations vs FCP over 1.13 seconds (SE Ranking). This is a technical factor with zero precedent in traditional SEO playbooks and requires log file analysis to diagnose (@iPullRank, Feb 2026).

The Compound Effect

No single spec produces citation. They compound.

A page with strong Zone 1 (front-loaded thesis) but weak Zone 3 (sections that depend on each other) may get retrieved but not cited. A page with perfect answer capsules but generic schema may be suppressed before the capsules are evaluated.

Pages scoring 0.70+ on a combined quality index AND hitting 12+ of these specifications achieve a 78% cross-engine citation rate (Kumar & Palkhouski, UC Berkeley, 1,100 URLs, 3 engines). Below that threshold, citation drops steeply.

The priority stack is based on correlation strength with citation:

1. Metadata + freshness signals (r=0.68)
2. Semantic HTML structure (r=0.65)
3. Structured data — attribute-rich only (r=0.63)
4. Evidence and citation density (r=0.61)
5. Authority and trust signals (r=0.59)
6. Internal linking (r=0.57)

This is the anatomy. Every zone, every spec, every data point traces back to a measured finding. The next post in this series covers L3 tactics: how to get your brand into the third-party sources AI actually cites for category queries.

This article was originally published on X by Aaron Haynes. Aaron is the CEO of Loganix, a visibility + SEO platform for brands and agencies.

Sources referenced:

@Kevin_Indig / Gauge, Feb 2026 (S16). 1.2M citations, ski ramp, 5 characteristics. @Kevin_Indig / SEL, Nov 2025 (S1). 72.4% answer capsule rate. @cyberandy / WordLift, Mar 2026 (S14). Embedding architecture, MRL front-loading. @dejanseo / Dejan AI, Dec. 2025 (S13). Grounding budget, density beats length. Shashko / Bright Data, Mar 2026 (CI28). Sentence-level citation, 10-word median. Growth Marshal, Feb 2026 (CI30). Generic schema hurts. Profound + @iPullRank, Feb 2026 (S8). Meta descriptions for AI. @deaborysenko / DEJAN (CI30). ChatGPT doesn’t parse JSON-LD. Semrush, 5M URLs, Jan 2026 (S17). Schema presence + URL slug length. SE Ranking, 129K domains, Nov 2025 (CI33). ChatGPT citation factors. SE Ranking, Dec 2025 (CI34). AI Mode citation factors. RepublishAI, 14,987 URLs, Mar 2026 (SEO1). Content refresh magnitude. Kumar & Palkhouski / UC Berkeley, Sep 2025 (S30). GEO-16 framework. McGill NLP, Mar 2026 (S19). Answer-shaped embeddings. Google Vertex AI Search docs, Feb 2026 (S7). Chunk-level retrieval