In-Page Information Architecture: Structuring Content for Three Audiences

Site architecture has a mature body of research. Topic clusters drove one HubSpot client from 500 to 190,000 monthly visitors. Crawl depth beyond three clicks degrades crawl rates by 33% (Botify, 6.2B requests). Internal linking A/B tests show +25% organic uplift (SearchPilot). These findings shape how practitioners connect pages to each other.

Content strategy has its own evidence base. The content engineering discipline covers what to build: content models, metadata, structured data, and governance systems. Neither body of work covers the layer between them, which is how a page organizes its own informationwithin the page.

The gap matters because three distinct consumers parse in-page structure, each with different processing logic.

Human readers scan in predictable patterns shaped by visual hierarchy. NN/g eye-tracking across multiple studies (2006-2019, cumulative n=200+) found that 79% of users scan any new page. Only 16% read word-by-word. The layer-cake pattern, where users fixate on headings and skip body text, is the most effective scanning behavior. It only occurs when pages provide sufficient structural cues. Without headings, bolding, and whitespace, users fall into the F-pattern: a failure state where reading efficiency collapses.

Search crawlers extract text from rendered HTML, segment it into passages, and classify entities. Passage-level indexing (patent US20160078102) evaluates individual passages within a page independently of surrounding content. A well-structured page with clear section boundaries gives the passage indexer cleaner input.

AI retrieval systems extract at sentence-level granularity. Shashko's analysis of 42,971 AI citations found a median cited sentence length of 10 words and a hard ceiling at 17 words. The grounding budget per query is approximately 2,000 words (DEJAN, 7,060 queries). A page must carry its highest-value sentences in self-contained form, because AI systems lift the exact sentence and write no paraphrase.

All three consumers converge on the same block, and four independent studies measured it at the same size. iPullRank defines a semantic unit as a 50-150 word block capturing a single concept with explicit subject-predicate-object structure. The size range aligns with multiple independent findings about how machines process content.

The convergence is notable because these studies measured different things. SE Ranking measured AI citation rates across 129,000 domains. Shashko measured sentence-level extraction patterns across six AI platforms. iPullRank measured cosine similarity improvements from restructuring content. AirOps measured citation likelihood across their 2026 dataset. They all arrived at the same structural unit size.

The semantic unit is more than a search optimization concept. It maps onto how a person reads in the layer-cake scanning pattern. Users fixate on a heading to decide whether to read the block below it, so each heading-plus-block is a decision point. A self-contained block gives an accurate preview in the heading and delivers on it in the body. It borrows no context from the block before it. The human scanner and the AI extractor can both process that block on its own.

NN/g paragraph attention data puts numbers on the decay. 81% of paragraph 1 reaches 81% of users, and paragraph 2 reaches 71%. Paragraph 3 reaches 63%, and paragraph 4 reaches 32%. The drop from the first paragraph to the fourth is nearly 50 percentage points. Front-load the most important information inside every semantic unit, and not only at the top of the page.

The unit has that size because the retrieval systems have a budget. They operate under hard constraints, and those constraints set how much of a page reaches the model. The AI citation research covers the retrieval pipeline in detail. Here the focus is on what these constraints mean for page-level structure.

Grounding plateaus at approximately 540 words / 3,500 characters regardless of page length. The architectural question is not "how long should the page be" but "which 540 words will the model see."

Sentence-level extraction specifics from Shashko's 42,971-citation study across six platforms paint the granular picture. The median cited sentence is 10 words. The maximum is 17 words, a hard ceiling with nothing longer cited in the entire dataset. 92.4% of citations fall between 6 and 20 words. Position bias is significant: the mean cited position is 34.9% down the page, with the 75th percentile at 48.8%. The top third of the page produces a disproportionate share of citations.

A related structural pattern validates the extraction mechanics at the paragraph level. Adam Gnuse (Saltbox Solutions, via Search Engine Land, November 2025) found an "answer capsule" in 72.4% of ChatGPT-cited blog posts. The capsule is a declarative statement of 120-150 characters, about 20-25 words, placed directly after a question-based H2. About 91% of these capsules held no link. The capsule is the source paragraph, and the pipeline extracts a shorter sentence out of it. That matches Shashko's sentence-level findings of a 10-word median and a 17-word maximum. The capsule carries the self-contained context, and the extraction pipeline takes the single most information-dense sentence inside it.

Kevin Indig and Nicole Johnson (Growth Memo, March 2026, 1.2M ChatGPT responses, 98K citations across 7 verticals) extended this principle to opening paragraphs. Declarative intro language (the "[X] is [Y]" form in the first paragraph) was the only universal writing signal across all 7 verticals, with a +14% aggregate citation lift. Entity types predict citation in different directions. DATE and NUMBER entities are universal positive signals. PRICE entities are universally negative, and Knowledge Graph-verified entities reduce citation rates to 0.81x. A page built for entity richness should favor temporal and quantitative entities over brand and product names.

The structural gap between pages is stark. Pages with lists, tables, and headings reached a 91.3% sentence-match rate. Unstructured pages reached 39.3%.

That 2.3x advantage comes from extractable sentences, not from better content. A well-researched paragraph buried in a wall of text is invisible to the extraction pipeline. The same sentence, with a heading above it and whitespace around it, becomes a citation candidate.

Heading count is vertical-specific. The Indig/Johnson study found that 3-4 headings performs worse than zero headings in every vertical studied. CRM and SaaS content peaks at 20 or more headings. Healthcare inverts, and zero headings performs best there. Match the heading density to what the citation-earning pages in your own vertical carry.

The machine constraints have a human counterpart, and researchers measured it decades earlier. Eye-tracking gives the empirical foundation for an in-page architecture decision. These patterns describe measured behavior rather than style preference, and they decide whether a reader processes the content at all.

NN/g eye-tracking studies (multiple rounds, 2006-2019, cumulative n=200+) established two primary scanning patterns. The layer-cake pattern is optimal: users fixate on headings and subheadings, and skip the body text between them. NN/g calls it the most effective way to scan a page. The pattern appears only where a page carries a clear visual hierarchy: distinct headings, varied formatting, and whitespace. The F-pattern is a failure state: users read the first line fully, scan partway through the second, then skim vertically down the left margin. It occurs when pages lack formatting cues. Reading efficiency collapses.

Combined scannable + concise + objective formatting produces+124% usability improvement (Morkes & Nielsen, 1997):

• Scannable text (headings, bullets): +47%
• Concise text (half word count): +58%
• Objective tone (non-promotional): +27%

The sample was small by behavioral science standards, but the directional findings are consistent with all subsequent NN/g research.

First impressions compound the structural stakes. Users form reliable aesthetic judgments within 50 milliseconds (Lindgaard et al. 2006, 1,000+ academic citations, replicated by Google/University of Basel 2012). Visual complexity and prototypicality affect perception at 17ms. The reader judges the structure above the fold before reading the first word. A dense wall of text triggers a negative aesthetic judgment before content quality can register.

The connection to NavBoost: these scanning patterns directly feed Google's most important ranking signal. If visual structure causes a user to stay and engage (goodClicks, lastLongestClicks), the page accumulates positive NavBoost signal on a rolling 13-month window. If poor structure causes a quick return to the SERP (badClicks), the page accumulates negative signal. In-page architecture is not a UX concern separate from ranking. It is a ranking input.

Scanning and extraction reward the same containers, and the citation data prices them. Specific formats produce measurably different AI citation rates. The format is the container, not the content. The same information, restructured, produces a different citation outcome.

The gradient from data tables (2.5x) to opinion pieces (18%) is a format effect, not a quality effect. Opinion pieces can be brilliant. Data tables can be trivial. The difference is extractability: tables present discrete, labeled data points that AI systems can lift directly. Opinion prose requires the model to identify the claim, which adds a processing step that reduces selection likelihood.

Previsible's analysis of 5,000 prompts (Ana Fernandez, January 2026) adds granular structural benchmarks. Cited pages carry a header every 100-200 words, against one header per 400 or more words on uncited pages. Lists appeared on 63% of cited pages and tables on 39%, and the table share rises for competitive and buying queries. FAQ sections appeared on 47%, mostly for factual and informational prompts. 58% of cited pages used an interrogative header.

When an LLM used the source content in a response, it converted that content into a list 76% of the time. The structure of the original made no difference. Five formats almost never earned a citation. Opinion pieces, storytelling without structure, blogs without intermediate headers, pages with more images than text, and pure conversion landing pages.

Clarity correlates with AI citation more strongly than depth, authority, or comprehensiveness. Structural clarity is the property that makes content parseable by a human scanner and by a machine extractor at the same time. The AI citation research covers the full pipeline from retrieval to citation. Here, the takeaway is narrower. Clarity is a structural property first. You build it with a heading hierarchy, semantic units, data tables, and format diversity. Clarity is an architecture decision before it is a writing quality.

Format diversity buys extraction targets, and it costs page length. Progressive disclosure is how a page buys depth without paying that cost. The pattern shows the core content first and reveals the detail on demand. NN/g research confirms it improves learnability, efficiency, and error reduction. The failure condition is more than two disclosure levels, where users lose orientation. Tooltips represent a single disclosure level, well within safe bounds.

A glossary tooltip is one implementation of the principle. It defines a domain term in context, through a hover or tap popover. Baymard Institute validated this pattern across 4,400+ usability test sessions (25 rounds, Think Aloud protocol). Definitions served in tooltips on desktop or tappable links on mobile improved comprehension without adding page length. The tested examples, B&H Photo for video resolution and Crutchfield for audio terminology, demonstrate the pattern across different product vocabularies.

NN/g timing research sets two implementation constraints. A 200ms open delay prevents accidental activation during normal cursor movement. A 150ms close delay prevents premature dismissal when the user moves the cursor into the tooltip to click a link.

WCAG 1.4.13 adds three requirements. The tooltip must dismiss on the Escape key. The user must be able to enter it without closing it. It must stay visible until the user dismisses it.

The SEO mechanics are straightforward. Google indexes tooltip content present in rendered HTML via its Chromium-based rendering pipeline, including content rendered through the native Popover API. The content sits in the DOM whatever its visual state. Google gives hidden content less weight than visible content. John Mueller has said so for tabs and accordions, and the same logic applies here. Tooltip definitions contribute to entity understanding and page-level semantics without carrying the full weight of visible body text.

The Information Gain patent (US20200349181A1, granted June 2024) provides one more conceptual connection. A tooltip definition plausibly cuts the effort side of the trade the patent describes, because it makes the content self-contained and removes the external lookup. The patent defines information gain as "the amount of valuable information learned minus the amount of effort it took to learn."

The connection is inferential. No study has tested whether a tooltip affects information gain scoring.

The largest research gap in this area: no published A/B test measures tooltip impact on engagement metrics (time-on-page, bounce rate, scroll depth, conversion). The UX case for tooltips rests on usability testing observations rather than on quantitative engagement data. Any site that runs the pattern at scale should close that gap.

Most of the structure above already has a written standard, drafted for a different reason. Web accessibility compliance is a legal obligation and a structural advantage at once. Many WCAG requirements produce the exact properties that serve search and AI retrieval.

The ARIA paradox: pages with ARIA averaged 57 errors compared to 27 on pages without ARIA (WebAIM Million 2025). ARIA does not cause errors. Complex implementations tend to be more broken. The finding is a caution against adding ARIA attributes as a checkbox exercise. Semantic HTML that needs fewer ARIA overrides produces better outcomes than ARIA layered on top of non-semantic markup.

The structural overlap between WCAG compliance and in-page architecture is concrete:

• Semantic heading hierarchy (h1-h6) creates the layer-cake scanning pattern that both users and passage-level indexing depend on.
• Alt text provides entity context for image understanding and multimodal retrieval.
• Document language aids Google's language classification pipeline.
• Keyboard navigation structure implies logical content ordering.
• Color contrast improves readability, affecting scanning efficiency and time-on-page, feeding back into NavBoost behavioral signals.

A correlation finding: WCAG-compliant sites show 23% more organic traffic and 27% more keywords (SEMrush/ AccessibilityChecker.org, 2025, n=10,000). The finding is a correlation and not a cause. A site that invests in accessibility usually invests in the other structural quality signals as well: semantic HTML, a proper heading hierarchy, and clean markup. Those signals together produce the traffic differential.

The business case extends beyond search. Click-Away Pound (2019) found 69% of disabled consumers abandon inaccessible sites. 4.9M disabled online shoppers represent GBP 17.1B/year in lost purchasing power. The European Accessibility Act (EAA) enforcement began 2025. In the US, UsableNet tracked 4,187 digital accessibility lawsuits in 2024, with projections trending upward for 2025.