Adjusts each candidate answer passage's score by reading the heading hierarchy above it, so passages under irrelevant sections lose, and passages under matching sections win.
Patent Overview
- Inventor
- Steven D. Baker
- Assignee
- Google LLC
- Filed
- 2014-06-04
- Granted
- 2018-05-01
- Application Number
- US 14/295,902
The Challenge
A Passage Out Of Context Is Misleading
A passage might contain the right words and the right answer shape but live under a heading that signals a different topic. "Side effects of ibuprofen" buried under "Drug Interactions" is not the same answer as the same passage under "Adverse Reactions". Local context decides whether the passage actually answers the question, and that local context is encoded in the heading hierarchy above the passage.
- Local Topic Matters — A passage's enclosing section topic can flip its meaning. The same sentence under different headings can answer different questions. The heading is what disambiguates.
- Bare Passage Scoring Misses This — Scoring a passage on its own text while ignoring the heading hierarchy treats off-topic passages as on-topic. The bare score has no awareness of the section the passage lives in.
- Need A Hierarchical Context Score — The system needs to read the chain of headings above each candidate passage and let that chain modulate the passage's answer score. The full heading path is what reveals the contextual topic.
- Section Drift Within Long Documents — Long documents cover many topics. Without context awareness, a section about an unrelated topic that happens to share query words wins over a properly scoped answer passage elsewhere in the document.
- FAQ-Like Structures Need Heading Awareness — FAQ pages have headings that often are the questions themselves. The context score must recognize when the enclosing heading matches the query directly, because that is the strongest possible context signal.
Innovation
Heading Vectors Adjust Answer Scores
For each candidate answer passage, the system builds a heading vector that describes the path through the document's heading hierarchy from root to the passage. A context score is derived from how well that vector aligns with the query. The base answer score is then adjusted by the context score so that passages under matching sections rise and passages under mismatched sections fall.
- Score The Candidate Passage — Compute the base answer score using the methods established by the upstream passage-scoring pipeline (query-term match plus answer-term match).
- Build The Heading Vector — Traverse the document's heading hierarchy from the root H1 down through every nested heading to the heading directly above the candidate passage. The result is an ordered vector of heading texts.
- Derive Context Score — Compare the heading vector against the query (or expected answer-type vocabulary) to produce a context score that measures topical fit of the enclosing section.
- Adjust The Answer Score — Combine the base answer score with the context score to produce an adjusted answer score. Strong context boosts the score; weak context penalizes it.
- Select Across Candidates — Choose the best answer from the candidates by ranking on adjusted answer scores. Passages under matching sections beat passages under off-topic ones even when their bare scores are similar.
- Apply Reward For Direct Heading Match — When the enclosing heading exactly or nearly matches the query, the context score reaches its maximum. This is what makes FAQ headings ("What is X?") strong answer-passage signals.
Headings Are Scoring Features
Most SEO writing treats headings as navigation cues. The patent treats them as first-class scoring features that adjust the answer score for every passage they enclose. The implications are direct: outline structure is retrieval structure.
Hierarchy Is Signal
The chain of headings above any passage is a topical context vector that the system reads and uses to adjust scoring. Ignoring the chain throws away signal the engine is actively consuming.
- Heading Path As Context — The path from H1 down through nested headings to the passage's enclosing heading is the context vector. Each heading in the chain contributes to the context score.
- Question-Match Bonus — Headings that match the query text directly (FAQ-style) produce the strongest context scores. The system rewards this match because it is unambiguous evidence of topical fit.
Outlines are not just for readers. They are signal that the engine consumes directly.
<\/section>Technical Foundation
Heading Hierarchy As A Feature
The patent treats the document's outline as a first-class input to passage scoring. The outline is read structurally, with each level contributing to the context vector.
- Heading Vector — An ordered representation of the heading chain from the root heading to the heading enclosing the candidate passage. Each entry in the vector is the text of one heading at one level.
- Context Score — A score derived from the heading vector that measures how well the passage's section aligns with the query intent. Higher values indicate stronger topical fit.
- Adjusted Answer Score — The base answer score modulated by the context score. Used to rank candidate passages globally across all responsive documents.
- Query-Heading Alignment — The specific comparison between the heading vector and the query that produces the context score. Can be implemented as term overlap, semantic similarity, or a learned model.
Quality Metrics
- Context Score — The alignment function compares the heading chain against the query. Higher values when the section topic matches; lower values when it does not.
ctx(P, Q) = align(heading_vector(P), Q) - Adjusted Answer Score — Weighted addition of the context score onto the base answer score. The weight w controls how much heading context matters relative to passage content.
adj_score(P) = base_score(P) + w * ctx(P, Q)
Key Insight: The heading hierarchy lets the engine resolve the disambiguation that passage text alone cannot. Two passages with identical text can have very different intents depending on what section they live in. The heading vector is the structural data that exposes that difference.
<\/section>The Process
Where Context Scoring Fits
Context scoring runs as an additional layer on top of the base passage scoring. It does not replace base scoring; it modulates it.
- Base Score Per Passage — The upstream passage scoring computes query-term-match and answer-term-match scores per candidate passage.
- Heading Vector Extraction — For each candidate, walk up the document's outline from the passage to the root, collecting heading texts at each level.
- Context Score Computation — Compare the heading vector against the query to produce the context score. Stronger alignment produces higher context scores.
- Score Adjustment — Combine base score and context score per passage. The combination uses configured weights.
- Final Ranking — Rank candidates by adjusted score across all responsive documents. The top-scoring candidate wins the answer slot.
What This Means for SEO
What This Means for SEO
Headings are not just navigation. They are scoring features. The heading chain above any passage on your page can boost or sink that passage's chance of being a featured snippet, and this is one of the most actionable findings in the Baker patent set.
- Headings Decide Passage Eligibility — An on-topic answer paragraph buried under an off-topic heading is at a scoring disadvantage. The heading hierarchy is read literally. Treat your outline as the system's primary signal for what each section is about.
- Use Specific, Question-Aligned Headings — If a section is supposed to win a featured snippet for "side effects of ibuprofen", the heading should say "Side Effects". A vague heading like "What To Know" loses the context score even with great content underneath.
- Honor The Hierarchy — Nested H2/H3/H4 structure should be a coherent chain. H3 under H2 under H1 should each refine the topic of the one above. Random heading levels (H3 with no H2 parent) muddy the heading vector.
- FAQ Sections Should Use Question Headings — FAQ-style content benefits from this rule directly. Use the actual question as the H3 above each answer passage. The heading vector then aligns perfectly with the query, which produces the maximum context-score boost.
- Avoid Generic Container Headings — Headings like "Overview", "More Information", "Details" carry almost no context signal. Replace them with topic-specific labels that describe what is actually under the heading.
- Section Length Should Match Heading Specificity — A specific heading should lead to a section that delivers exactly on that specificity. A heading like "Side Effects of Ibuprofen" followed by 3,000 words on unrelated drug interactions wastes the context signal.
- Heading Word Order Matters For Match — The query-heading alignment looks at the heading text. When you write the heading, prefer the word order that matches the most common query phrasing for the intent. Small reorderings can shift alignment scores.
- Use Multiple Heading Levels For Coverage — An H2 establishes the broad topic; H3 sub-sections handle specific intents within that topic. The hierarchy gives you a way to cover multiple related questions on one page, each with its own scoring-friendly context.