Where Structure Actually Works — and Why Brands Usually Get Only One of Four Decisions Right

By /

By Manuel Hürlimann for GaryOwl.com | Published: April 13, 2026 | Updated: April 13, 2026
Expertise: Digital Authority Engineering | AI Pipeline Analysis | Structural Authority
Time to read: 22 minutes
Series: Operative Article 5 — Glossary

From One Decision to Four — Reframing Structural Authority

Article 4 in this series — “Six Types of Authority AI Systems Actually Measure” — ended with a forward link titled “Where Structure Actually Works.” That link made a specific promise: not a promise to map every authority type to every pipeline stage, but a promise to answer one question. Where, in the technical chain that turns a web page into an AI citation, does structural authority actually take effect?

The question matters because Structural Authority is the authority type most brands think they understand and most brands get wrong. The mistake is almost always the same: treating structure as one decision — the HTML hierarchy, the heading levels, the FAQ markup — when in fact structural authority operates as a cascade of four distinct decisions, each with its own peer-reviewed evidence base, each with its own failure mode, and each determining a measurable fraction of whether a page becomes a citation or stays invisible. A brand can do the HTML right and still be invisible, because the HTML is only one of four structural decisions and the other three are silently breaking.

This article establishes the four-stage cascade. It introduces the five peer-reviewed benchmarks that have, over the past eighteen months, quantified how each stage of the cascade constrains Retrieval-Augmented Generation performance. And it reframes Structural Authority as an original DAE synthesis: not a property of HTML, but a chain of structural decisions that spans parsing, extraction, segmentation, and markup preservation. Every step of that chain has a measured effect size. Every step is a place where content architecture either holds or fails.

“Structure is not one thing. It is four things pretending to be one — and the pretending is why brands optimize for one decision and lose on the other three.” — Manuel Hürlimann, Digital Authority Engineering

📌 Key Insights — What This Article Establishes

  1. Structural Authority operates as a four-stage cascade, not as a single decision. The four stages are parsing quality, parsing robustness, retrieval granularity, and markup preservation. Each has its own peer-reviewed evidence base and its own failure mode.
  2. The parsing ceiling is real and measured. Zhang et al. (ICCV 2025) document that even the best available parsing tools leave at least a 14% F1-score gap compared to structured ground truth, across 8,498 Q&A pairs in seven domains.
  3. Smaller models are more sensitive to parsing failures — at least on arithmetic tasks over financial tables. Hui et al. (NeurIPS 2024) found that on the FinHybrid subset of UDA, clean parsing improved GPT-4-Turbo accuracy by ~5.7% and Llama-3-8B accuracy by ~15%. The pattern weakens on other subsets (PaperTab), so the “smaller models benefit more” claim is a domain-specific observation, not a general law. As smaller models spread through production pipelines, the question of whether this pattern generalizes becomes operationally important.
  4. Retrieval granularity is a first-order determinant of citation quality. Chen et al. (EMNLP 2024 main) showed that segmenting documents into atomic propositions rather than passages improves retrieval Recall@5 by up to 12 absolute points and downstream answer quality by up to 7.5 Exact Match points.
  5. HTML markup carries semantic information that plain-text extraction destroys. Tan et al. (WWW 2025) demonstrated that preserving HTML structure during retrieval produces statistically significant improvements of up to 4.5 Exact Match points over plain-text baselines, with the largest gains on multi-hop reasoning tasks.
  6. Chunking strategy is context-dependent, not universally valuable. Qu et al. (NAACL 2025 Findings) found that expensive semantic chunking does not consistently outperform simple fixed-size chunking — except on multi-topic documents, where the gap widens to over 12 F1 points.
  7. No single structural investment is sufficient. The four stages are multiplicative. A page that optimizes for one stage and neglects the other three performs near the bottom of the achievable range. Content architecture is a first-order concern precisely because it has to succeed in four places at once.

📌 First Publication: The Four-Stage Structural Authority Cascade

This article, first published in April 2026 by Manuel Hürlimann on GaryOwl.com, is the first piece in the DAE series to treat Structural Authority as a four-stage cascade with peer-reviewed empirical backing for each stage. The cascade synthesis — parsing quality, parsing robustness, retrieval granularity, and markup preservation as four distinct decisions with multiplicative effects — is an original DAE contribution built from five peer-reviewed benchmarks spanning ICCV, NeurIPS, EMNLP, WWW, and NAACL. The ceiling-vs-floor framing of parsing effects, the argument that parsing robustness scales inversely with model size, and the identification of four multiplicative structural investments as a first-order concern are original syntheses within the Digital Authority Engineering framework.

📌 Navigate the DAE Framework

📌 Reading Guide

  • If you read one section: Read “Beyond the Parser.” It contains the three peer-reviewed findings that most GEO and AI-SEO guidance still ignores.
  • If you are a strategist (9 minutes): Read “The Four Stages of Structural Authority” + “Four Investments, Not One.” That gives you the full operational picture.
  • If you already understand structural multidimensionality: Skip directly to “The Parsing Ceiling and the Parsing Floor” for the quantified effect sizes, and then to “Beyond the Parser” for the retrieval-granularity and HTML-markup evidence.

📌 Glossary: Key DAE Terms in This Article

The Four Stages of Structural Authority — A New DAE Synthesis

Article 4 established that AI systems measure at least six distinct types of authority, each through a different mechanism. Among those six, Structural Authority (#67) was defined as the authority type that governs whether a page is technically accessible to AI systems and whether its content can be parsed into useful, retrievable chunks. That definition contains a hidden complication, which this article now makes explicit: “accessible” and “parsable” and “retrievable” and “structurally preserved” are not the same property, they happen at different points in the pipeline, and a brand that conflates them will invest in the wrong one.

The four stages of the cascade are:

StageDecisionPrimary Peer-Reviewed SourceEffect Size
AParsing Quality — how well extraction converts the fetched document to clean textZhang et al. (ICCV 2025) ≥14% F1 ceiling gap vs. ground truth
BParsing Robustness — how well the language model tolerates imperfect parsingHui et al. (NeurIPS 2024) ~5.7% to ~15% accuracy gain on FinHybrid (weaker on other subsets)
CRetrieval Granularity — how text is segmented into retrievable unitsChen et al. (EMNLP 2024) +12 Recall@5, +7.5 downstream EM
DMarkup Preservation — how much structural signal survives retrieval-to-generation handoffTan et al. (WWW 2025) +4.5 EM on multi-hop tasks

Every one of these four stages has been quantified in peer-reviewed research over the last eighteen months. Every one of them has a failure mode that cannot be repaired further downstream. And every one of them is multiplicative with the others, which means the total effect of getting structure right is not the sum of four small improvements but their product.

Structural Authority, in the DAE sense, is the authority type that operates across this cascade. It is the only authority type that spans so many pipeline decisions, which is why it is the most misunderstood and the most operationally consequential of the six. The rest of this article develops the cascade stage by stage, brings in the empirical evidence, and ends with the operational consequence for brands that want to invest in structure without wasting effort on the wrong decision.

Stage 2: The Mechanical Dimension That Comes Before the Cascade

Before the four-stage cascade can operate at all, the page has to arrive at the parser. This is the Stage 2 problem from the broader pipeline model (Article 6 in this series develops the full four-stage pipeline) — the question of whether the crawler can reach the page and receive its content intact. Four technical properties determine this: discoverability, accessibility, renderability, and crawler-specific handling.

Discoverability is the question of whether the crawler knows the page exists. Crawlers work from a graph: they start with pages they already know and follow links to pages they have not yet seen. Sitemaps accelerate this by declaring page sets in bulk. Internal linking architecture determines which pages get revisited and which get orphaned. A page that is not in the sitemap, not linked from any already-crawled page, and not otherwise announced will not be fetched — not because the crawler decided against it, but because the crawler has no way to know it exists.

Accessibility is the question of whether the crawler can make the HTTP request and receive a non-error response. This is where robots.txt configuration, server availability, rate limiting, and crawler-specific blocking come into play. A page that returns 200 OK for a browser user-agent but 403 Forbidden for GPTBot is, for the purposes of AI citation, not accessible. A surprising number of sites are in this state by accident, because they adopted a blanket bot-blocking rule early in the AI-crawler debate and never revised it. Operators who have not audited their robots.txt in the last twelve months should treat this as an open question until they verify it.

Renderability is the question that destroys more sites than the other three combined. A modern web page is often delivered as a near-empty HTML shell with a large JavaScript bundle, and the actual content is assembled in the browser after the JavaScript executes. This is called client-side rendering, and it is the default behavior of many React, Vue, and Angular applications. For a human user with a browser, it works fine. For a crawler, it depends entirely on whether the crawler executes JavaScript. Most AI crawlers — at least the ones that dominate the citation layer — do not. They fetch the initial HTML response, read what is there, and move on. What is there, for a client-side-rendered site, is an empty div. The content that the operator wrote, edited, and structured never enters the pipeline at all.

This failure mode is operationally important because it is invisible from the inside. A site owner looking at their own pages sees the fully rendered content, because their browser executed the JavaScript. The gap between what the owner sees and what the crawler sees is the gap between being cited and being invisible, and for many modern sites it is the single largest structural authority problem.

The fix at Stage 2 is to ensure that the content that matters is present in the initial HTML response, before any JavaScript executes. This can be done through server-side rendering, static site generation, or incremental static regeneration — all standard features in modern frameworks like Next.js, Nuxt, SvelteKit, and Astro. The architectural pattern does not matter much; what matters is that when the crawler requests the page, the response contains the content.

Crawler-specific handling is the fourth property and the most variable one. Different AI systems use different crawlers with different capabilities, user-agents, and fetch behaviors. The details — including the distinction between classical citation-producing crawlers, agentic Chromium-based browsers, and agentic browser extensions — are the subject of Article 7 in this series. For the purposes of this article, the point to hold is simpler: Stage 2 is not a single switch. It is a set of conditions that have to be true simultaneously, and any one of them being false makes the others irrelevant.

Once Stage 2 has succeeded and the page has arrived at the parser with its content intact, the four-stage cascade begins. That is where the peer-reviewed evidence has concentrated, and that is where the rest of this article focuses.

The Parsing Ceiling and the Parsing Floor — Quantifying Stages A and B

Two peer-reviewed benchmarks, published in 2024 and 2025, quantify Stage A (parsing quality) and Stage B (parsing robustness) from opposite ends of the measurement. They frame the question differently, and the difference is informative.

Zhang et al. (ICCV 2025): The Parsing Ceiling

Zhang et al. (ICCV 2025) , in “OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation,” introduced OHRBench — a benchmark of 8,498 question-answer pairs across 1,261 documents from seven real-world domains (textbook, law, finance, newspaper, manual, academia, administration) — to measure the gap between machine-parsed documents and a perfectly-structured ground truth. For each document, the team compared state-of-the-art parsing pipelines (pipeline-based OCR, end-to-end OCR, and vision-language models such as GPT-4o and Qwen-VL) against the ground-truth reference. The result: a performance gap of at least 14% in F1 score, even when the most advanced tools are used. No parser in the test field reached ground-truth performance.

The question they asked was how much RAG performance degrades when the system works from the parsed version instead of the ground truth. Some parsing approaches produced much larger gaps than the 14% floor. None produced no gap. Important context for the transfer to HTML web content: this measurement was made on PDFs processed through OCR-based pipelines — the direction of the finding generalizes to HTML, but the specific magnitude is context-dependent (see Honest Limitation).

The framing of this result matters. Zhang et al. did not measure the difference between good parsing and bad parsing. They measured the difference between the best available parsing and perfect input. The gap they found is the parsing ceiling — the best that the current tool landscape can achieve, compared to what perfect structural input would allow. A site that does everything right at the Stage A editorial level is still constrained by this ceiling, because no amount of HTML hygiene can compensate for the fact that the parser is working from an imperfect reconstruction.

Hui et al. (NeurIPS 2024): The Parsing Floor and the Model-Size Robustness Curve

Hui et al. (NeurIPS 2024) , in the Datasets and Benchmarks Track paper “UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis,” framed the question from the opposite end. Their UDA benchmark comprises 29,590 expert-annotated question-answer pairs across 2,965 documents in six subsets spanning finance (FinHybrid, TatHybrid), academic literature (PaperTab, PaperText), and knowledge-base Q&A (FetaTab, NqText). For the parsing-strategy comparison in particular, the experiment narrowed to two subsets (FinHybrid and PaperTab) and two models (GPT-4-Turbo and Llama-3-8B), comparing five parsing pipelines against raw-text and ground-truth baselines. On the FinHybrid subset, this comparison reveals what we can call the parsing floor — the cost of skipping structural parsing compared to doing it properly — and it shows the largest effects for small models requiring exact numerical extraction from financial tables.

Hui et al. added a detail that reshapes the operational picture — but with a scope caveat that deserves explicit naming. Their parsing-strategy experiment tested only two subsets of the UDA benchmark (FinHybrid, which covers arithmetic Q&A over financial tables in S&P 500 earnings reports, and PaperTab, which covers extractive Q&A over academic tables) and only two models (GPT-4-Turbo at the high-capacity end and Llama-3-8B at the low-capacity end, with no intermediate sizes). On the FinHybrid subset — where exact numerical extraction is rewarded — clean parsing produced a 5.7% relative Exact-Match improvement for GPT-4-Turbo and approximately 15% for Llama-3-8B. Larger models, with greater parametric capacity, could reconstruct missing information through contextual inference; smaller models, with less capacity, depended more directly on clean input signals. The authors frame this in their own words: “the much smaller Llama-3-8B offers a significant 15% enhancement, suggesting that compact models with a limited capability of parsing table layouts may benefit more from enhanced parsing.”

But on PaperTab, this pattern partially inverts. For several conditions, raw-text parsing even outperformed well-parsed input. Hui et al. themselves note that “in the PaperTab dataset, where completely accurate information is less critical, GPT-4-Omni and raw-text parsing could even outperform well-parsed tables.” The finding, in other words, is strongest when exact numerical extraction is required — and weaker or absent when the task tolerates approximate information.

Model CategoryParametric CapacityAccuracy Gain from Clean Parsing (FinHybrid subset)
High-Capacity (e.g., GPT-4-Turbo)Very high~5.7% (relatively robust)
Mid-to-Low Capacity (e.g., Llama-3-8B)Limited~15% (more sensitive)

Scope note: These figures are drawn specifically from the FinHybrid subset of UDA, which tests arithmetic Q&A over financial tables. On the PaperTab subset (extractive academic Q&A), the pattern weakens or partially inverts. The two-model comparison across one subset does not constitute a fully validated “robustness curve”; it is a directional observation that motivates Investment 2 below as a DAE synthesis, not as a direct empirical conclusion.

📌 Core thesis: The Model-Size Robustness Curve

As organizations deploy smaller, cheaper models for cost reasons, the value of Stage A parsing quality grows rather than shrinks. The operational instruction follows: write as if the reading model is small. Content that is dense with explicit entity mentions, explicit claim statements, and explicit reference structures survives parsing imperfections that would have broken more implicit text.

Put the two papers together and the shape of the Stage A / Stage B problem becomes clear. Parsing quality has a floor (Hui et al.: approximately 5.7–15% cost of getting it wrong on FinHybrid arithmetic Q&A, weaker on other subsets) and a ceiling (Zhang et al.: ≥14% gap between the best available parsing and structured ground truth, measured on PDFs; direction transfers to HTML but magnitude is context-dependent — see Honest Limitation). The space between the floor and the ceiling is the space in which Stage A editorial work actually operates. Content architecture cannot fully close either gap, but it can determine where in that space a given page lands. A page with strong editorial structure lands near the top of the achievable range. A page with weak structure lands near the bottom. No page escapes the ceiling entirely, because the ceiling is a property of the parsing layer itself, not of the content.

This is only the first half of the cascade. Two more peer-reviewed findings show that even when parsing has succeeded, the subsequent structural decisions — how text is segmented into retrievable units and how much of its original structure survives retrieval — carry their own measurable effect sizes.

Beyond the Parser: Why Granularity and Markup Matter

The parsing layer is where most operator intuition stops. Once the text has been extracted, the assumption usually goes, the hard structural work is done. Three peer-reviewed papers show that this assumption is wrong. Two post-parsing decisions — how the extracted text is segmented into retrievable units, and how much of its markup is preserved when retrieval hands text to the language model — produce effect sizes comparable to or larger than the parsing layer itself.

Chen et al. (EMNLP 2024 Main): Proposition-Level Retrieval as a Structural Decision

Chen et al. (EMNLP 2024 Main Conference) , in a paper titled “Dense X Retrieval: What Retrieval Granularity Should We Use?”, asked a deceptively simple question: when a document corpus is indexed for retrieval, what should the unit of indexing be? Traditional practice uses passages of approximately 100 words. Chen et al. tested an alternative: atomic propositions — the smallest self-contained units of meaning that can stand alone as factual statements.

The experiment decomposed the entire English Wikipedia into 256.9 million atomic propositions, compared against 41.4 million traditional passage units — a 6.2× increase in the number of indexable units. The team then tested retrieval performance across five open-domain QA benchmarks (Natural Questions, TriviaQA, WebQuestions, SQuAD, and EntityQuestions) using both unsupervised retrievers (SimCSE, Contriever) and supervised ones (DPR, GTR). Downstream QA performance was tested with LLaMA-2-7B and with Fusion-in-Decoder.

The numbers are striking. With unsupervised retrievers, proposition-level indexing improved average Recall@5 by +12.0 absolute points (from 34.3 to 46.3 with SimCSE) and average Recall@20 by +10.1 points. Even with supervised retrievers already trained on passage-level data, gains persisted: +2.6 to +2.8 Recall@5 with GTR and DPR. More consequentially, downstream answer quality followed retrieval quality: feeding proposition-level retrieval to LLaMA-2-7B yielded +2.7 to +4.1 Exact Match points on average across the five datasets, and with Fusion-in-Decoder the gains reached +2.1 to +7.5 EM points. The finding has been influential enough that both LangChain and LlamaIndex have integrated proposition-level chunking as standard options since late 2024.

Granularity LevelUnitRetrieval (Recall@5)Downstream EM (LLaMA-2-7B)
Passage~100 wordsBaselineBaseline
SentenceSingle sentenceSlight improvementModerate
PropositionAtomic fact+12.0 (SimCSE)+2.1 to +7.5

The operational implication for content architecture is subtle but important. Proposition-level indexing works best when the source text is written in a way that makes atomic propositions easy to extract — when sentences are self-contained, when claims are not buried in multi-clause reasoning, when key facts are stated directly rather than implied. In other words, the writing style that performs best in a proposition-indexed RAG system is the writing style that DAE’s AI-Priority-Zone principle has been advocating for: short, self-contained, fact-dense passages that can be lifted out of their surrounding context without losing meaning. Chen et al. did not test AI-Priority-Zone content directly, but the architecture of their finding aligns with it precisely. A page written in dense, chunkable, self-contained units is a page that survives proposition-level indexing with the most information intact.

Tan et al. (WWW 2025): HTML Markup as Machine-Readable Structure

Tan et al. (WWW 2025) , in a paper titled “HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems,” asked what happens when retrieval preserves HTML markup rather than stripping it to plain text before passing retrieved chunks to the language model. The question matters because standard RAG pipelines do strip markup — the operating assumption has been that HTML tags are formatting noise that the model does not need. Tan et al. tested this assumption directly.

The experimental setup used six QA benchmarks (ASQA, HotpotQA, Natural Questions, TriviaQA, MuSiQue, ELI5) with 400 queries each for a total of 2,400 test cases. The retrieval source was Bing search results, so the corpus is real web content rather than synthesized or cleaned documents. Tan et al. built a two-step block-tree pruning method that reduced HTML down to approximately 6% of its original token length while preserving structure — headings, tables, lists, emphasis, and nested relationships. They compared this against plain-text baselines produced by standard extraction.

With Llama-3.1-70B, the results were statistically significant on most benchmarks at p < 0.05: +4.50 EM on HotpotQA, +3.25 Hit@1 on ASQA, +2.75 Hit@1 on Natural Questions, and +0.75 EM on MuSiQue versus the best plain-text baseline. The largest gains appeared on multi-hop reasoning tasks, where the language model had to connect information across multiple retrieved documents. The structural cues in HTML — which heading a fact sits under, which table row a number occupies, which list item links to which concept — turned out to carry information that the model could use to disambiguate and connect evidence.

MethodInformation DepthHotpotQA Gain (EM)
Plain Text ExtractionLinguistic context onlyBaseline
HTML-Preserving (HtmlRAG)Linguistic + structural context+4.50

A methodological caveat is worth naming explicitly, because it has been raised in post-publication review of the paper. HtmlRAG’s pipeline includes a pruning step that the plain-text baselines do not have, which could inflate the apparent advantage. The authors’ ablation studies suggest that the structural preservation itself, not just the pruning, produces the improvement — but the fairness of the comparison is an open question. What remains uncontested is the direction of the effect: HTML markup is not formatting noise. It is structural information that the language model can read when the pipeline shows it, and that the language model cannot read when the pipeline has stripped it away.

The operational implication for content architecture is blunt. Every HTML tag that a careful author writes — every <h2>, every <table>, every <blockquote>, every <li> — is a piece of machine-readable semantic structure. When the retrieval pipeline converts that structure to plain text (which most pipelines still do by default), that information is destroyed. When the pipeline preserves it, the information flows through to the generation step and affects the answer. The author does not control which pipelines preserve markup — that is a decision made by platform operators. But the author does control whether the markup is there to preserve in the first place. A page written in flat paragraph walls has no markup to preserve, even in a pipeline that would preserve it. A page written with clean, meaningful HTML structure carries preservable information into every pipeline that is willing to carry it forward.

Qu et al. (NAACL 2025 Findings): Chunking Is Context-Dependent

Chen et al. showed that changing the unit of retrieval from passages to propositions produces large, consistent gains. Qu et al. (NAACL 2025 Findings) add an important nuance: not every change to chunking strategy produces gains, and the gains depend on the type of content being chunked.

Qu et al., in “Is Semantic Chunking Worth the Computational Cost?”, compared three chunking strategies across 10 document retrieval datasets and 5 evidence retrieval datasets: fixed-size chunking (200 words), breakpoint-based semantic chunking, and clustering-based semantic chunking. The comparison was run across three embedding models to control for retriever effects.

The headline finding is that semantic chunking does not consistently outperform fixed-size chunking on natural documents. On HotpotQA, fixed-size chunking produced 90.59 F1@5 while breakpoint semantic chunking produced only 87.37 F1@5. On several other natural-document benchmarks, the difference was similarly small or inverted. Given that semantic chunking requires roughly 100× more computation than fixed-size chunking, Qu et al. concluded that the overhead is not justified by consistent gains.

But a second finding from the same paper inverts the picture for a specific content type. On artificially “stitched” multi-topic documents (where short documents on unrelated topics were combined into longer composites), semantic chunking produced dramatic improvements: 81.89 vs. 69.45 F1@5 on MIRACL, and 63.93 vs. 43.79 on Natural Questions stitched — gaps of 12 to 20 F1 points. When documents contain multiple disconnected topics, breaking them at topic boundaries matters enormously. When documents contain a single coherent topic, it barely matters at all.

Document TypeSemantic Chunking AdvantageStrategic Recommendation
Topically focusedNegligibleFixed-size chunking is sufficient
Sprawling / Multi-topic12–20 F1-point gainSemantic chunking required

📌 Strategic implication: What You Write Determines Which Chunking Strategies Benefit You

A page that covers one tight topic with linear reasoning will perform well under any reasonable chunking strategy, because the chunking decision is nearly trivial. A page that mixes several loosely-connected topics in a single document — the sprawling, comprehensive “ultimate guide” format that dominates SEO content — will perform well only under semantic chunking, and most production RAG systems do not use semantic chunking because of its cost. The content architecture that DAE recommends — focused, topically-concentrated articles with clear structure — is not just a readability choice. It is a choice that makes the content robust to whatever chunking strategy the downstream pipeline happens to use.

A methodological caveat to Qu et al. should be noted. The paper has two authors from Vectara, Inc., a commercial RAG platform, and the finding that “simple chunking works fine” aligns with a commercial narrative of pipeline simplification. However, the methodology is transparent, the evaluation uses standard public benchmarks, and the paper was peer-reviewed through the ACL review process at NAACL. The nuanced finding about stitched documents also partially cuts against a purely commercial interpretation. The paper is usable as peer-reviewed evidence, with the vendor affiliation noted.

The Cascade Is Multiplicative, Not Additive

The four stages of the cascade are not independent. Their effects compound, which means a page that optimizes for one stage and neglects the others performs at the bottom of the achievable range, not somewhere in the middle.

Consider the arithmetic. Zhang et al. document a parsing ceiling that constrains every page regardless of content quality: at least 14% F1 is lost between the best parsing and perfect input. Hui et al. document a parsing floor that varies by model and by subset: approximately 5.7% to 15% accuracy is gained by doing parsing right on FinHybrid arithmetic Q&A, with smaller effects elsewhere. Chen et al. document a retrieval-granularity effect: up to 12 Recall@5 points and 7.5 EM points from proposition-level over passage-level indexing. Tan et al. document a markup-preservation effect: up to 4.5 EM points from HTML-preserving retrieval over plain-text retrieval. Qu et al. document a chunking-strategy effect that ranges from negligible (on coherent documents) to 20 F1 points (on multi-topic documents).

These are not five numbers that can be added. They are five separate effects that operate in sequence, and a page either succeeds or fails at each one. A failure in an early stage cannot be compensated for by optimization in a later stage. If Stage A produces semantic noise because the parser cannot read the page, Stage C (proposition-level indexing) merely breaks that noise into smaller, more precise fragments of wrong knowledge. A page that fails at Stage A (bad parsing because of poor HTML structure) has nothing left to optimize at Stages B, C, and D — the downstream pipeline is working with degraded input. A page that succeeds at Stage A but fails at Stage C (the content is written as a sprawling multi-topic wall that no reasonable chunking strategy can segment cleanly) delivers correctly-parsed garbage to the retrieval layer. A page that succeeds at Stages A, B, and C but fails at Stage D (the HTML structure was never there to preserve, because the page was written in flat paragraphs) loses whatever markup information the pipeline would have carried forward.

The multiplicative characterization is a conceptual derivation from this cascade logic, not a directly measured interaction effect. No single study measures all four stages simultaneously to test whether their effects compound multiplicatively rather than additively. The cascade model predicts multiplicative behavior — each stage gates what the next stage can do — and empirical verification of the compound effect awaits a future benchmark that tests all four stages in a single experiment.

📌 Core thesis: The Four Stages Are Multiplicative

No single structural investment is sufficient. Content architecture is a first-order concern not because one structural decision matters, but because four structural decisions matter, and they matter in sequence. A site that treats Structural Authority as a single checkbox — “we have clean HTML” — is investing in one-quarter of what actually determines citation outcomes.

Four Investments, Not One — The Operational Consequence

The entire preceding argument collapses into an operational statement: Structural Authority requires four simultaneous investments, and none is a substitute for the others.

Investment 1 — Parsing-friendly HTML structure. This is the classic content-architecture work: clean heading hierarchy, meaningful section boundaries, tables that are actually <table> elements rather than CSS-styled divs, lists that are <ul> or <ol>, FAQ sections with <h3> questions and answer paragraphs directly below. The purpose of this investment is to minimize parsing ambiguity so that Stage A produces clean output. Zhang et al. (ICCV 2025) is the empirical backing: parsing has a ceiling, but well-structured HTML lands closer to the ceiling than poorly-structured HTML.

Investment 2 — Model-robust content density. This investment is new in the DAE vocabulary and is inspired by — not directly concluded from — Hui et al.’s (NeurIPS 2024) observation that smaller models benefited disproportionately from clean parsing on the FinHybrid subset, which tests arithmetic Q&A over financial tables. The generalization from this two-model, single-subset observation to a broader content-architecture principle is a DAE synthesis, not a direct empirical conclusion from the cited study. Model-robust content is content that is dense with explicit entity mentions, explicit claim statements, and explicit reference structures — not content that relies on implicit reasoning that a large model could have recovered from noisy input. A smaller model that receives model-robust content survives parsing imperfections that would have broken a more implicit text. The operational expression is: write as if the reading model is small. This DAE synthesis awaits broader empirical testing across more model sizes and more domain types before it can be stated as a validated general principle.

Investment 3 — Proposition-ready content structure. Chen et al.’s (EMNLP 2024) proposition-indexing finding has a content-side implication: content that can be decomposed into atomic, self-contained claims survives proposition-level retrieval better than content that embeds its claims in dependent-clause reasoning. This is the DAE AI-Priority-Zone principle stated empirically. Every <blockquote> infobox, every key-statistics box, every stand-alone fact-laden paragraph is a proposition-ready unit. Every multi-sentence argument where the main claim requires context from the preceding and following sentences is a unit that proposition retrieval will struggle with.

Investment 4 — Markup-preservation-ready structure. Tan et al.’s (WWW 2025) HtmlRAG finding has a content-side implication that is more direct than the others. Every HTML tag that conveys meaning — heading levels, table organization, list membership, emphasis — is a piece of machine-readable information that survives structure-preserving retrieval and is destroyed by plain-text extraction. For a markup-preserving pipeline, a semantic <strong> tag functions as an attention anchor — a machine-readable signal that a particular concept matters enough to mark. The operational instruction is: use HTML semantically, not decoratively. A <strong> tag that marks an actually-important concept is an attention anchor for the model. A <strong> tag used for visual emphasis on a non-important word is noise. The difference matters only when the pipeline preserves markup — which is not yet the default, but is becoming more common as HtmlRAG and similar architectures propagate through production systems.

📌 Three immediate actions for Structural Authority

  1. Audit your rendering architecture. Fetch your pages with curl using a GPTBot user-agent string and verify that the content you care about is in the initial HTTP response. If it is not, no downstream investment will compensate.
  2. Rewrite your flat paragraph walls as self-contained sections. Every section should be independently understandable. Every key fact should survive being lifted out of its context. This is the Investment 3 bar.
  3. Use HTML tags semantically, not decoratively. Audit existing pages for <strong>, <em>, <table>, and heading usage. Every tag should carry meaning that a markup-preserving retrieval pipeline could read. This is the Investment 4 bar.

The four investments are multiplicative. A page with perfect HTML structure (Investment 1) but sprawling multi-topic content (failing Investment 3) delivers clean parse output that the chunker cannot segment cleanly. A page with focused, proposition-ready content (Investment 3) but decorative-rather-than-semantic markup (failing Investment 4) leaves information on the table when the pipeline would have used it. A page that writes for a large model and assumes implicit reasoning (failing Investment 2) breaks silently when the content is served to a smaller model at lower cost.

📌 Sequencing guidance for limited budgets

The multiplicative character of the four investments means that all of them must eventually be in place — but the operational order in which they are addressed is not arbitrary. Investment 1 (parsing-friendly HTML structure) should be addressed first, because it has the broadest evidence base and because it is the prerequisite for all downstream stages: if parsing fails, nothing else matters. Investment 3 (proposition-ready content structure) should come second, because it has the largest measured effect sizes on retrieval (+12 Recall@5) and because it aligns with writing practices that are already standard DAE discipline through the AI-Priority Zone principle. Investment 4 (markup-preservation-ready structure) follows as third priority, because it builds on Investment 1 — clean HTML is the precondition for meaningful semantic markup. Investment 2 (model-robust content density) is the most demanding because it requires editorial consistency across the lifetime of many articles, and its benefits become most pronounced as smaller models proliferate in production pipelines. An operator starting with a limited budget should address Investments 1 and 3 as the first wave, then Investments 4 and 2 as the second wave. This is sequencing guidance, not priority ranking — the multiplicative logic still holds, and none of the four can be permanently skipped without cost.

All four investments are in the author’s control. None of them require privileged access to the platform or the pipeline. They are structural choices that can be made in the content editor, verified in the source HTML, and maintained through editorial discipline.

From One Authority Type to the Full Map — What Comes Next

This article has done one thing at length: it has established that Structural Authority is not one decision but four, and it has brought in the peer-reviewed evidence that quantifies each one. It has not mapped the other five authority types to their pipeline stages. That map — the full coordinate system in which every authority type from the Article 4 taxonomy lands in a single primary stage, with two types (Structural and Network) operating multidimensionally — is the subject of Article 6, “Mapping the Six Authority Types to the AI Pipeline.” Article 6 takes the four-stage logic developed here and generalizes it: it introduces the four-stage pipeline model for the full taxonomy, presents the complete mapping table, and shows how the taxonomy and the pipeline read together as rows and columns of the same diagnostic framework.

The reason to separate the two articles is that they answer different questions. This article answers “why does Structural Authority require four investments?” Article 6 answers “where does every authority type take effect, and how does the map become a diagnostic tool?” Both are operational, both build on the Article 4 taxonomy, and both are necessary for the full DAE picture.

In an era in which AI systems decide what information is visible, structural authority is no longer a technical footnote. It is the necessary condition for machine-verifiable expertise.

Living Lab Disclosure — Where GaryOwl.com Stands on the Four Investments

GaryOwl.com is a living lab for the Digital Authority Engineering framework, and this article is subject to the same disclosure standard as every other article in the series. At the time of writing (April 12, 2026), GaryOwl.com is positioned on the four Structural Authority investments as follows.

Investment 1 (parsing-friendly HTML structure): Substantially in place. Articles use clean heading hierarchy, semantic table elements, native list elements, and FAQ sections with explicit question-answer structure. The WordPress Astra theme produces compliant HTML output, and article templates enforce structural consistency across the series.

Investment 2 (model-robust content density): In progress. Articles are written with explicit entity mentions and explicit claim statements, but systematic measurement of model-size sensitivity has not been performed on GaryOwl.com content. The Q2 scorecard article will address this gap by testing selected articles against smaller open-source models to verify that the intended information survives lower-capacity inference.

Investment 3 (proposition-ready content structure): Substantially in place. The AI-Priority-Zone principle (infoboxes, key-statistics boxes, stand-alone fact paragraphs) has been applied systematically since Article 3. Every DAE article contains at least 10 blockquote-based infoboxes designed as self-contained proposition units.

Investment 4 (markup-preservation-ready structure): Partially in place. HTML markup is used semantically (headings for hierarchy, strong for genuine emphasis, tables for tabular data, lists for enumerations) but the site has not been tested in a markup-preserving retrieval pipeline. Whether the structural information survives any particular AI system’s extraction is an external variable that the site operator does not fully control.

The site has not yet been tested at scale against the four-stage cascade framework this article presents. The Q2 scorecard article later in the series will address that gap with systematic measurement across multiple AI platforms. For now, the honest statement is that Structural Authority is the one authority type for which GaryOwl.com has the most complete internal picture, and it is also the one for which the external measurement problem is hardest.

📌 Honest Limitation — What This Article Demonstrates and What Remains Open

The parsing ceiling was measured on PDFs, not HTML. Zhang et al.’s ≥14% F1 gap was established on 1,261 documents that were primarily PDFs processed through OCR-based pipelines. HTML web content is parsed differently. The direction of the finding transfers to the HTML context (parsing quality imposes a ceiling on RAG performance), but the specific 14% value does not. The magnitude of the HTML parsing ceiling is an open empirical question.

The cascade model is structural, not causal. The four-stage cascade is a framework for thinking about where structural decisions operate in the pipeline. It is not a claim that every page’s failure can be diagnosed to a single stage, or that the stages are strictly sequential in every system. Real RAG pipelines have feedback loops, reranking stages, and generation-time retrieval adjustments that the four-stage cascade abstracts away.

The five papers cover different RAG architectures. Zhang et al. and Hui et al. test document-parsing pipelines. Chen et al. tests open-domain QA retrieval over Wikipedia. Tan et al. tests web-document retrieval via Bing. Qu et al. tests chunking strategies across mixed benchmarks. Transferring the specific numeric values to a particular AI system’s production pipeline requires interpretive judgment that this article does not attempt. What transfers is the structural direction, not the percentages.

Vendor affiliations exist and are named. Of the five peer-reviewed papers cited, three have industry co-authors: Chen et al. (Tencent AI Lab), Tan et al. (Baichuan Intelligent Technology), and Qu et al. (Vectara). All three went through peer review at top-tier venues and use transparent, publicly-reproducible methodology. Industry co-authorship on peer-reviewed work is normal and is not the same as vendor-published white papers.

This article operates on the content-architecture layer of a larger structural problem. The analogous problem on the enterprise-ontology layer — the gap between vendor claims of “semantic reasoning” and what informal property graphs actually provide — is analyzed by Nicolas Figay in “Everyone Has an Ontology Now. Almost Nobody Has an Ontology” (LinkedIn, April 2026). Figay’s question — “what can your system actually prove?” — applies equally to content architecture. The two analyses are independent but structurally parallel: both name a gap between what is promised (semantic reasoning, structural authority) and what operates technically (governed property graphs, parsed text, stripped markup). DAE addresses the content-layer instance of this pattern. The enterprise-ontology instance remains outside this article’s scope.

The model-size robustness finding generalizes from a narrow evidence base. Hui et al.’s parsing-strategy experiment tested only two models (GPT-4-Turbo and Llama-3-8B) on two subsets (FinHybrid and PaperTab), and the ~5.7% vs ~15% accuracy gap was observed only on FinHybrid arithmetic Q&A over financial tables. On PaperTab’s extractive academic Q&A, the pattern partially inverted — for some conditions, raw-text parsing even outperformed well-parsed input. No intermediate-sized models (Llama-3-70B, Qwen-1.5-32B, Mixtral-8x7B) were included in the parsing comparison, so the characterization of a “model-size robustness curve” is strictly a two-point observation, not a validated continuous relationship. The DAE cascade uses this finding to motivate Investment 2 (model-robust content density), but the empirical basis is narrower than a general “robustness curve” claim would require. Broader replication across more model sizes and more document types would strengthen or qualify this claim.

Forward Link — What Article 6 Will Show

This article gives you the structural argument for one authority type, in full peer-reviewed depth. The next article gives you the full map. Article 6 — “Mapping the Six Authority Types to the AI Pipeline” — takes the four-stage cascade developed here and extends it into a complete coordinate system: four pipeline stages, six authority types, one table that shows where each type takes effect, and the diagnostic logic that turns the map into an operator’s tool.

→ Article 6 link will activate on publication.

In an era in which AI systems increasingly decide which information becomes visible, structural authority is no longer a technical footnote at the bottom of a content brief. It is the necessary condition for machine-verifiable expertise — the precondition on which every other authority signal depends. A page that fails the four-stage cascade cannot be rescued by reputation, citation count, or editorial depth. A page that passes the cascade gives every other authority signal the chance to register at all. This is the practical meaning of structural authority: not a set of SEO checkboxes, but the gate through which machine-legible expertise passes — or does not.

Frequently Asked Questions

Why is this article only about Structural Authority when Article 4 introduced six types?

Because Article 4’s forward link was specifically titled “Where Structure Actually Works” — which is a promise about structural authority, not about all six types. Article 6 in this series is the full mapping article that covers every type. This article is the focused treatment that the Article 4 forward link promised, and it is also the first piece in the series to treat Structural Authority with the peer-reviewed depth that the topic requires. A single article trying to do both would dilute the Structural Authority argument.

Is client-side rendering really that bad for AI citations?

For pages whose content has to reach AI crawlers, yes. Most AI citation crawlers do not execute JavaScript and therefore see only the initial HTML response. The details vary by crawler, but the safe default assumption is that client-side-rendered content is invisible to the crawlers that produce the majority of current AI citations. The full picture, including the distinction between classical citation crawlers and agentic browser-based fetchers, is the subject of Article 7 in this series. The conservative operational stance is to ensure that content which has to be cited is present in the initial server response, regardless of what else the site does client-side.

How do I test whether my page is Stage 2 accessible?

The minimum test is to fetch the page with curl using a user-agent that identifies as a major AI crawler and inspect whether the content you care about is in the response body. If it is, Stage 2 is working for that page and that crawler. If it is not — if the response is an empty shell or a loading placeholder — Stage 2 is failing and no amount of downstream work will compensate. More thorough testing procedures (multiple crawler user-agents, JavaScript-disabled browser inspection, server log analysis) are the subject of Article 7.

Does the Zhang et al. parsing ceiling mean content architecture is pointless?

No — it means the opposite. The ceiling exists regardless of what the operator does, but where within the achievable range a given page lands depends heavily on content architecture. A page with strong editorial structure performs at the top of what the current parsing layer can produce. A page with weak structure performs at the bottom. The ceiling makes content architecture more important, not less, because it is the one variable the operator controls that the parsing layer cannot paper over.

Why do smaller language models benefit more from good parsing?

Hui et al. (NeurIPS 2024) observed this on the FinHybrid subset of the UDA benchmark, which tests arithmetic Q&A over financial tables. On that subset, the accuracy gain from well-parsed input was ~5.7% for GPT-4-Turbo and ~15% for Llama-3-8B — nearly a 3× amplification, but observed on only two models and one subset. On the PaperTab subset (extractive academic Q&A), the pattern weakened or partially inverted. Their paper does not fully explain the mechanism, but the likely explanation is that larger models have more parametric capacity to recover from noisy input through contextual inference, while smaller models depend more directly on clean input signals — at least when the task requires exact information. The operational implication, treated as a DAE synthesis rather than a validated general law, is that as organizations move toward smaller, cheaper models, the value of Stage A and Stage B investments may grow. Broader replication across more model sizes and more document types is needed to confirm or qualify this.

What is proposition-level retrieval and why does it matter for content writers?

Proposition-level retrieval, as tested by Chen et al. (EMNLP 2024) , indexes documents at the granularity of atomic self-contained claims rather than at the granularity of 100-word passages. In their Wikipedia benchmark, switching from passage-level to proposition-level indexing improved retrieval Recall@5 by up to 12 absolute points and downstream answer quality by up to 7.5 Exact Match points. For content writers, the implication is that content written in self-contained, fact-dense units (the DAE AI-Priority-Zone principle) survives proposition-level indexing better than content that embeds claims in multi-sentence reasoning. Both LangChain and LlamaIndex now support proposition-level chunking as standard options.

Does HTML markup really matter or is that just a theoretical finding?

Tan et al. (WWW 2025) tested it directly on 2,400 queries across six real-world QA benchmarks using Bing search results. HTML-preserving retrieval produced statistically significant improvements of up to 4.5 Exact Match points over plain-text baselines, with the largest gains on multi-hop reasoning tasks. The effect sizes are modest compared to the parsing ceiling or proposition retrieval, but they are real, measured, and directional. The practical implication is that HTML markup is machine-readable semantic information that flows through to the generation step when the pipeline preserves it. Whether any particular AI system preserves markup is a platform-operator decision, but the author’s responsibility is to make sure the markup is there to preserve in the first place.

Is semantic chunking worth implementing for my content?

Qu et al. (NAACL 2025 Findings) found that semantic chunking does not consistently outperform simple fixed-size chunking on natural, topically-coherent documents — and it costs roughly 100× more computation. On stitched multi-topic documents, semantic chunking produced dramatic improvements (up to 20 F1 points). The practical answer depends on what you write: if your content is focused and topically-coherent, semantic chunking is not worth the cost. If your content is sprawling “ultimate guide” material, semantic chunking helps a lot — but the cost may not be justifiable for the operator. The DAE recommendation flips the question: write topically-focused content so that your content performs well under any chunking strategy.

Sources & Methodology

Evidence Classification: [A] Peer-reviewed academic research / [B] Large-scale industry dataset (>100K samples) / [C] Industry study with documented methodology / [D] Vendor study / [DAE] Original DAE contribution

This article builds on the sources already synthesized in “Six Types of Authority AI Systems Actually Measure” (Article 4). The sources below are the five peer-reviewed benchmarks that directly support the four-stage cascade argument, plus the DAE framework references.

Peer-Reviewed Academic Research:

  • Junyuan Zhang, Qintong Zhang, Bin Wang, Linke Ouyang, Zichen Wen, Ying Li, Ka-Ho Chow, Conghui He, Wentao Zhang (2025). “OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation.” ICCV 2025 (pp. 17443–17453). arxiv.org/abs/2412.02592. Code and data: github.com/opendatalab/OHR-Bench (Accessed: April 12, 2026) — 8,498 Q&A pairs across 1,261 documents in seven domains; documents a ≥14% F1 gap between the best available parsing and structured ground-truth data. Stage A empirical foundation.
  • Yulong Hui, Yao Lu, Huanchen Zhang (2024). “UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis.” NeurIPS 2024 Datasets and Benchmarks Track. arxiv.org/abs/2406.15187. Code: github.com/qinchuanhui/UDA-Benchmark (Accessed: April 12, 2026) — 29,590 expert-annotated Q&A pairs across 2,965 documents in six subsets. The parsing-strategy comparison (Table 5) specifically tests two subsets (FinHybrid, PaperTab) and two models (GPT-4-Turbo, Llama-3-8B), measuring ~5.7% accuracy gain for GPT-4-Turbo and ~15% for Llama-3-8B on FinHybrid arithmetic Q&A over financial tables. On PaperTab, the effect weakens or partially inverts. Stage B empirical foundation (parsing robustness); model-size sensitivity is a directional observation from a narrow evidence base, generalized in this article as DAE synthesis.
  • Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, Dong Yu (2024). “Dense X Retrieval: What Retrieval Granularity Should We Use?” EMNLP 2024 Main Conference (pp. 15159–15177). aclanthology.org/2024.emnlp-main.845. DOI: 10.18653/v1/2024.emnlp-main.845. Code: github.com/chentong0/factoid-wiki (Accessed: April 12, 2026) — 256.9M atomic propositions vs 41.4M passages across English Wikipedia, tested on five open-domain QA benchmarks; +12.0 Recall@5 points with unsupervised retrievers, +2.1 to +7.5 downstream EM points. Stage C empirical foundation.
  • Jiejun Tan, Zhicheng Dou, Wen Wang, Mang Wang, Weipeng Chen, Ji-Rong Wen (2025). “HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems.” Proceedings of the ACM Web Conference 2025 (WWW ’25). arxiv.org/abs/2411.02959. DOI: 10.1145/3696410.3714546. Code: github.com/plageon/HtmlRAG (Accessed: April 12, 2026) — 2,400 queries across six QA benchmarks; +4.50 EM on HotpotQA, +3.25 Hit@1 on ASQA, statistically significant at p<0.05. Stage D empirical foundation.
  • Renyi Qu, Ruixuan Tu, Forrest Sheng Bao (2025). “Is Semantic Chunking Worth the Computational Cost?” Findings of the Association for Computational Linguistics: NAACL 2025 (pp. 2155–2177). aclanthology.org/2025.findings-naacl.114. DOI: 10.18653/v1/2025.findings-naacl.114 (Accessed: April 12, 2026) — 15 datasets across three chunking strategies and three embedding models; semantic chunking does not consistently outperform fixed-size on natural documents but produces 12–20 F1-point gains on stitched multi-topic documents. Stage C nuance source. Authors Qu and Bao are affiliated with Vectara, Inc.; methodology and benchmarks are transparent and publicly reproducible.

DAE Framework References:

Methodology: This article is authored by Manuel Hürlimann and follows the DAE Journalistic Source Principle. The four-stage cascade framing — parsing quality, parsing robustness, retrieval granularity, and markup preservation as four distinct decisions with multiplicative effects on Structural Authority — is an original synthesis built from five peer-reviewed benchmarks. Each stage is backed by a primary peer-reviewed source, and Stage C is additionally supported by Qu et al. as a nuance source. The triangulation rule from Production Guide V1.5 (§45) requires three independent sources for core claims; this article exceeds that requirement by using five independent peer-reviewed sources, all [A]-classified. The effect sizes reported (≥14% F1 on PDF OCR; ~5.7% to ~15% on UDA FinHybrid arithmetic Q&A; +12 Recall@5 on Wikipedia QA; +4.5 EM on HotpotQA multi-hop reasoning; 12–20 F1 on stitched multi-topic documents) are cited in their original measurement contexts and are not transferred as direct predictions to HTML web content outside the papers’ experimental conditions. What transfers is the structural direction of each finding — that each stage of the cascade has a measurable, non-trivial effect on RAG output quality — and the operational conclusion that all four stages have to be managed by content architecture, not just one. The multiplicative characterization of the cascade is a conceptual derivation from the sequential gating logic, not a directly measured interaction effect.

Contact: manuel@octyl.io

Update Log

  • [Future updates will be documented here.]

About the Author

Manuel Hürlimann is the creator of Digital Authority Engineering (DAE) — the systematic discipline of building machine-verifiable expertise that AI systems recognize, cite, and recommend. Based in Switzerland, he works as a consultant and lecturer at the intersection of AI search behavior, citation analysis, and brand authority.

Through the Authority Intelligence Lab at GaryOwl.com, he publishes original research on how AI systems select, evaluate, and cite sources — applying every principle to GaryOwl.com itself as a living lab. This article extends the Article 4 taxonomy with the first peer-reviewed-backed treatment of Structural Authority as a four-stage cascade, synthesizing five independent benchmarks from ICCV, NeurIPS, EMNLP, WWW, and NAACL.

Connect: GaryOwl.com · LinkedIn · manuel@octyl.io

Framework Disclosure: DAE is developed by GaryOwl.com and applied to GaryOwl.com itself as a living lab — every framework principle is simultaneously tested on this site. The framework is open for use with attribution. Validation is ongoing and published transparently; no guarantees implied. AI behavior varies by model and platform.

Article Navigation: ← Six Types of Authority AI Systems Actually Measure | Next: Mapping the Six Authority Types to the AI Pipeline →

GaryOwl.com – Authority Intelligence Lab

“Structure is not one thing. It is four things pretending to be one — and the pretending is why brands optimize for one decision and lose on the other three.” — Manuel Hürlimann

Scroll to Top