The common understanding is that LLMs read web pages the way humans do — top to bottom, absorbing context progressively. That model is wrong for retrieval-augmented generation. RAG systems do not read pages. They chunk content into passage-level segments, score each segment against the query for relevance and attributability, retrieve the highest-scoring passages, and feed only those passages to the language model for answer synthesis. The extraction mechanism operates at the passage level, not the page level, and the scoring criteria prioritize claim specificity and verifiability over topical relevance alone.
Chunking Strategies Determine Which Text Boundaries the Retrieval System Treats as Extractable Units
RAG systems split web content into chunks using a combination of HTML structural markers, token-count windows, and semantic coherence scoring. The chunking strategy determines whether a nuanced argument split across two paragraphs is evaluated as one unit or broken into fragments that lose their meaning independently.
The dominant chunking approaches in production search RAG systems include structural chunking (splitting at HTML heading boundaries, treating each H2 or H3 section as a chunk), fixed-window chunking (splitting at predetermined token counts, typically 128-256 tokens per chunk with overlapping windows to preserve context at boundaries), and semantic chunking (splitting at natural semantic boundaries where topic shifts are detected through embedding similarity drops between adjacent sentences).
Structural chunking is the most common approach in search-focused RAG because it leverages the content creator’s own organizational decisions. When a page uses heading hierarchy consistently, each section becomes a natural extractable unit. This is why heading structure matters for AI citation beyond its traditional SEO role: headings define chunk boundaries, and chunks that align with complete semantic units score higher than chunks that split mid-argument.
Content structure influences chunk quality directly. A page that places its primary claim in one paragraph and its supporting evidence in the next paragraph may see these split into separate chunks, with the claim-only chunk scoring lower because it lacks evidence and the evidence-only chunk scoring lower because it lacks the claim it supports. Pages that contain self-contained argument units within each paragraph (claim plus evidence in the same text block) produce chunks that maintain their informational value regardless of where the chunking algorithm places boundaries. [Confirmed]
Relevance Scoring at the Passage Level Prioritizes Semantic Match Density Over Keyword Frequency
The retrieval system scores each chunk against the query using dense vector similarity, not traditional keyword matching. Passages with high semantic alignment to the query intent — even without exact keyword matches — score higher than keyword-optimized passages with lower conceptual precision.
The embedding-based scoring mechanism converts both the query and each content chunk into high-dimensional vector representations using transformer models trained on semantic understanding. The similarity between the query vector and each chunk vector determines the relevance score. This approach captures meaning rather than vocabulary: a passage discussing “server response latency” scores high for a query about “how fast websites load” because the semantic concepts overlap, even though no keywords match.
This differs fundamentally from BM25-style retrieval, where term frequency and inverse document frequency drive scoring. BM25 rewards exact and near-exact keyword matches, creating an optimization surface where keyword placement matters. Dense vector retrieval rewards conceptual precision, creating an optimization surface where the clarity and specificity of the semantic content matters. A passage that precisely describes a concept using domain-specific terminology scores higher than a passage that repeats the query’s keywords while vaguely addressing the concept.
The content characteristics that produce high semantic similarity scores include: specific assertions rather than general statements (specificity aligns more precisely with specific queries), domain-specific terminology used accurately (the embedding model recognizes domain concepts and matches them to domain-specific queries), and complete answer units that address the full scope of the query within a single passage rather than distributing the answer across multiple passages. [Confirmed]
Attributability Scoring Filters Passages Based on Whether Claims Can Be Verified Against the Source
After relevance scoring, the system applies an attributability filter that evaluates whether each claim in a passage is traceable to the source document. Google’s published research on attributed question answering (Bohnet et al., 2023) established the framework for evaluating whether a generated answer can be supported by cited sources.
The attributability evaluation asks: if a reader followed the citation to the source page, would they find evidence supporting the claim attributed to that source? Passages containing unsupported assertions (“everyone knows that…”), vague generalizations (“studies have shown…”), or claims that require external context to verify (“as mentioned earlier…”) score lower on attributability because a reader following the citation would not find specific supporting evidence at the source.
Passage-level writing patterns that pass the attributability threshold include: self-contained claims that include their own evidence within the paragraph (“According to Ahrefs’ 2025 study of 600,000 URLs, 86% of top-ranking pages contained AI-generated content”), assertions with explicit sourcing (“Google’s March 2024 spam update introduced scaled content abuse as a new policy category”), and factual statements with verifiable specifics (“the Web Rendering Service enforces a practical timeout of approximately five seconds”).
Writing patterns that fail attributability include: aggregated conclusions without specific sources (“research consistently shows”), opinions presented as facts without evidence (“this is the most effective approach”), and context-dependent claims that only make sense within the article’s broader argument (“given what we discussed above, this means…”). Each of these patterns produces passages that the retrieval system cannot confidently attribute to the source page. [Confirmed]
The Generation Model Selects and Recombines Passages, Creating Citation Gaps Where Source Content Is Paraphrased Beyond Attribution
When the language model synthesizes an answer from multiple retrieved passages, it may paraphrase content sufficiently that the original source is no longer cited. This citation gap occurs in the generation phase, after retrieval and attribution scoring are complete.
The generation model receives the retrieved passages as context and produces a synthesized answer. During synthesis, the model may combine information from multiple passages into a single sentence, rephrase a passage’s claim using different vocabulary, or extract a fact from one passage and present it within a framing derived from another. Each transformation step reduces the traceability of the final sentence back to a specific source passage.
The citation gap phenomenon means that some content producers see their information appear in AI answers without receiving citation credit. The information was retrieved from their page, used by the generation model to inform its answer, but paraphrased to a degree that the attribution system cannot confidently link the generated sentence back to the specific source passage. This is particularly common when the generation model synthesizes a general principle from multiple specific examples across different sources — no single source “owns” the synthesized claim.
The practical implication for content producers is that distinctive, specific claims are more citation-resistant to paraphrase than general observations. A passage stating “Google’s SpamBrain system reduced search spam by over 40% in 2024” is difficult to paraphrase without retaining attribution to the source, because the specific metric and named system are identifiable. A passage stating “spam detection improved significantly” is easy to paraphrase beyond attribution because the claim contains no distinctive elements. Content that embeds unique data points, proprietary metrics, or distinctive analytical framing maintains citation attribution through the generation model’s paraphrasing more effectively than generic content. [Observed]
Token Budget Constraints Force the Retrieval System to Prefer Concise, Claim-Dense Passages Over Verbose Explanations
RAG systems operate under strict token budgets for the context window passed to the generation model. The retrieval system must maximize information density per token, creating a structural preference for concise passages that deliver claims efficiently.
Production RAG systems typically allocate 2,000-8,000 tokens for retrieved context within a generation request. When 5-15 source passages must fit within this budget alongside the query and system instructions, each passage slot holds approximately 150-500 tokens. Passages that deliver a claim in 40-60 words (approximately 50-80 tokens) leave room for more source diversity within the budget. Passages that take 200 words (approximately 250 tokens) to make the same claim consume more of the budget while contributing the same informational value.
The structural preference for conciseness means that passages delivering a claim in fewer tokens consistently score higher in the final retrieval ranking than verbose passages making the same point. Two passages with identical semantic relevance and attributability scores compete on information density: the passage that delivers the claim in fewer tokens wins because it provides better token economics for the generation model’s context window.
This token budget constraint creates a writing optimization target that differs from traditional web content advice. Traditional SEO content strategy often encourages detailed explanations, context-setting introductions, and comprehensive coverage. The AI citation optimization target is claim-dense passages where every sentence contributes a verifiable assertion. Introductory sentences, transitional phrases, and contextual preambles consume tokens without contributing extractable claims, reducing the passage’s information density and its competitiveness against more concisely written alternatives. [Reasoned]
What content structure produces the highest-scoring chunks when RAG systems split pages at heading boundaries?
Pages where each H2 or H3 section contains a self-contained argument unit produce the highest-scoring chunks. Each paragraph should contain both the claim and its supporting evidence within the same text block. When a claim appears in one paragraph and the evidence in the next, structural chunking may split them into separate chunks, and each scores lower independently because one lacks evidence and the other lacks the assertion it supports.
Why do passages with specific data points get cited more than passages with general statements?
RAG systems apply attributability filtering that evaluates whether each claim can be verified against the source. Passages with specific named entities, quantified metrics, and dated events provide verification anchors the retrieval system can cross-reference. General statements like “research consistently shows” fail the attributability test because a reader following the citation would not find specific supporting evidence at the source page.
How does the generation model’s paraphrasing cause content to be used without receiving citation credit?
During answer synthesis, the language model may combine information from multiple passages, rephrase claims using different vocabulary, or extract facts from one source and frame them using another. Each transformation step reduces traceability back to the original source. Distinctive, specific claims resist paraphrase-based citation loss better than generic observations because specific metrics and named systems remain identifiable even after rephrasing.
