>Fahlout_

The Search Engineering Dictionary

The vocabulary of SEO has evolved from "keywords and links" to embeddings, retrieval, and agents.These are the concepts that matter for how search and AI discovery actually work now.

Ranking Systems & Signals

Information Gain Score
A Google scoring system (patent US11354342B2) that measures how much novel content a document contains relative to what already exists in the result set. Google can demote or exclude a document that says the same thing as the other ranking pages, because it scores near zero.
Site Quality Score (Panda)
A site-level quality signal (patent US9031929B1) calculated from the ratio of navigational queries directed at the site vs. informational queries the site answers. A low score suppresses every page on the domain. Thin or repetitive pages in a programmatic build lower the ranking ceiling of the whole domain. The 2024 API leak revealed pandaDemotion as a pre-computed site-wide demotion in CompressedQualitySignals, plus babyPanda variants (lighter/faster iterations). Panda operates as 'algorithmic debt', a site-level quality tax. No page-level optimization can raise the ceiling that it sets.
Pairwise Quality Scoring
Google ranks a page in a head-to-head comparison against the other pages that compete for the same query. No absolute quality score sets the position. A strong page can rank #7 while a weaker page ranks #1. The weaker page wins the pairwise comparison on specific signals (links, entity coverage, click data).
Topical Authority
Not a single signal but the combined output of multiple overlapping systems: site-level topic embeddings (QualityAuthorityTopicEmbeddings), siteFocusScore, siteRadius, NsrChunks (per-section topic evaluation), and ClusterUplift. 50 thin articles do not move the topic embedding vector. 20 comprehensive, entity-rich articles do.
Content Freshness Scoring
A document scoring system (patent US8549014B2) that tracks the age distribution of the content on a page. The 2024 API leak revealed three related attributes. lastSignificantUpdate tracks a substantive revision and ignores a cosmetic edit. freshByDocFp fingerprints the document to show whether the content changed or only the timestamp. bylineDateConfidence scores the accuracy of the byline date, and a contradictory date degrades the freshness signal. Google stores only the last 20 versions of a document, and a date change with no substantive edit does not improve the freshness score.
N-gram Quality Prediction
A quality detection system (patent US9767157B2) that builds a phrase model from sites of known quality and scores new content against it. Creates a linguistic fingerprint of what quality writing looks like. The patent-level mechanism behind detecting thin content, keyword stuffing, and machine-generated text with unnatural phrase distributions.
Entity-Based Ranking
A ranking system (patent US10235423B2) that identifies entities in search results through a knowledge graph. The system weights each entity by type (person, place, organization, product) and computes a composite ranking score. Recognition as a distinct entity is a direct ranking input.
Core Web Vitals (LCP, INP, CLS)
Google's three field-measured performance metrics, Largest Contentful Paint, Interaction to Next Paint, and Cumulative Layout Shift, form the page experience ranking signal. Assessed using real Chrome User Experience Report (CrUX) data at the 75th percentile (p75), meaning 75% of visits must meet the threshold, not the average. CWV is a tiebreaker signal, and not a primary ranking factor. A repair to poor CWV does not move a page from position 30 to position 5. The repair can decide position 4 against position 8 among otherwise comparable pages. All three metrics must pass for a page to earn 'good' status.
Largest Contentful Paint (LCP)
Measures the time until the largest visible element in the viewport finishes rendering. Thresholds: good ≤ 2.5s, needs improvement ≤ 4.0s, poor > 4.0s. LCP fails more often than the other two Core Web Vitals, and an image is the LCP element on most mobile pages. The measurement decomposes into TTFB, resource load delay (discovery latency), resource load duration (download), and element render delay (main-thread blocking). Resource load delay usually contributes most, and it can take half the 2.5s budget before the download starts. Keep the LCP image visible to the preload scanner, set fetchpriority='high', serve AVIF, and cut TTFB.
Interaction to Next Paint (INP)
Measures the latency between a user interaction (click, tap, keypress) and the next visual frame, at the near-worst interaction of the visit. INP replaced First Input Delay (FID) as a Core Web Vital in March 2024, and FID measured only the first interaction. Thresholds: good ≤ 200ms, needs improvement ≤ 500ms, poor > 500ms. Each interaction decomposes into input delay (the main thread is busy), processing time (the handler runs), and presentation delay (style, layout, paint). Input delay usually contributes most, because a third-party script or a React hydration pass holds the main thread before the handler can start. A mobile device carries far less processing headroom, so a script that passes on desktop can fail there.
Cumulative Layout Shift (CLS)
Measures unexpected movement of page elements, scored as the largest 'session window' of shifts: a burst inside 1 second, with a 5-second gap between windows. Thresholds: good ≤ 0.1, needs improvement ≤ 0.25, poor > 0.25. CLS is unitless, and it multiplies the impact fraction (the viewport area affected) by the distance fraction (how far the elements moved). An above-fold shift scores worse because it displaces more content below it, and Google excludes any shift that follows a user interaction. Common causes: images without explicit dimensions, non-composited CSS animations, injected banners and chat widgets, web fonts without metric overrides, and ad slots without reserved space. The bfcache (back/forward cache) restores a page with a CLS of 0, and an ad-heavy layout remains the most persistent failure.
Engagement Signal
Behavioral metrics like dwell time, scroll depth, and repeat visits. NavBoost uses these per-topic to re-rank results. Can also feed back into RLHF loops for AI systems.
Passage-Level Ranking
Google evaluates and ranks individual passages within documents independently (patent US20090055389A1). Sections compete on their own merits: author popularity, word choice, passage diffusion, and the unique information that the passage adds. A single well-written section can surface even when the broader document isn't the best overall match.
Time on Task
Measures how long a user needs to complete a goal, or the percentage of users who succeed. The measure is finer than generic dwell time.
Topic Centroid
The average embedding vector computed from a site's keyword portfolio, representing its semantic center of gravity. Used to measure how well individual pages align with the site's core topics and to identify off-topic pages dragging down site-level quality scores.
contentEffort
An attribute from the 2024 Google API leak described as an LLM-based effort estimation for article pages. May quantify the human labor, the originality, and the resources behind a page, including unique images, original data, embedded tools, and linguistic complexity. On a correct reading of the leak, contentEffort is the closest algorithmic proxy for the Experience dimension of E-E-A-T. Nobody outside Google knows the current weight, or whether the attribute stays active.
Q* (Quality Star)
Google's aggregate site/document quality metric, revealed through the DOJ antitrust trial and 2024 API leak. Combines content quality, authority signals, and an evolved PageRank that measures 'distance from a known good source.' Q* is largely static and query-independent. A high Q* score applies across every topic on a domain. Described as 'deliberately engineered rather than machine-learned.' The relationship: E-E-A-T is the goal, Q* is the system, Site_Quality is the score.
CompressedQualitySignals
The pre-computed quality gatekeeper module in Google's ranking pipeline. Contains per-document signals, among them siteAuthority, pandaDemotion, navDemotion, anchorMismatchDemotion, and exactMatchDomainDemotion. The module can disqualify a page before query-time ranking begins, and no on-page optimization compensates for a poor score. The module is the mechanism that makes site-level quality a ceiling.
Twiddler
Specialized re-ranking functions in Google's SuperRoot framework that adjust results after the primary ranking algorithm (Ascorer) runs. Named Twiddlers include NavBoost (click-based), QualityBoost, FreshnessBoost, RealTimeBoost, and WebImageBoost. 'Lazy Twiddlers' only process the top 20-30 results with granular adjustments. When Google says that a system is 'not part of the core algorithm', the system often runs as a Twiddler. A Twiddler is a post-ranking overlay, and not a primary scoring input.
Index Tiering (Base / Zeppelins / Landfills)
Google's three-tier document storage system. Base (flash memory) holds the most important, frequently updated content. Zeppelins (SSDs) hold mid-priority pages. Landfills (hard drives) hold low-importance, irregularly updated content. Tier placement determines crawl frequency and serving speed. It also sets outgoing link value, because a link from a Base-tier page carries more signal than a link from a Landfill-tier page.
Copia / Firefly
Google's scaled content abuse detection system, revealed in the 2024 API leak. Copia (Latin for 'abundance') monitors content velocity, which is the ratio of generated URLs to substantive articles. Firefly aggregates inputs from Copia, page quality scores, and NavBoost user dissatisfaction signals to make site-wide demotion decisions. The pair is Google's primary mechanism against AI-generated content farms. No single 'AI content detector' exists. Rapid content velocity, low engagement, and low quality together trigger Firefly.
hostAge (Sandbox)
A PerDocData attribute described as used 'to sandbox fresh spam in serving time.' The value is the earliest date Google saw any page there. Google has publicly denied a sandbox for new sites, and the leak contradicts the denial. A new domain faces a trust-building period whatever its content quality. The disadvantage is structural against a site with 13 months of accumulated NavBoost click signals.

Retrieval & Discovery

Canonical Tag
An HTML link that identifies the master version of a page, preventing duplicate URLs from diluting authority.
Robots.txt / Robots Meta
Rules that allow or block crawling and indexing, managing crawl budget and LLM ingestion cost.
Sitemap / RSS / Atom Feed
XML files listing URLs and timestamps so crawlers and RAG sync jobs know what's new or updated.
Query Fan-Out
Google's AI systems decompose a single query into sub-queries across eight variant types: equivalent, follow-up, generalization, specification, canonicalization, translation, entailment, and clarification. The systems run the sub-queries in parallel, then stitch the results together. Gemini 3 averages 10.7 sub-queries per prompt (78% increase over Gemini 2.5), and 95% of fan-out queries have zero traditional search volume. Fan-out coverage has a 0.77 Spearman correlation with AI citation likelihood.
Sub-Query
A derivative search created during fan-out (e.g., "cost of living in Brooklyn" from "Is NYC affordable?"). Each sub-query feeds fresh documents into the retrieval pipeline independently.
Cosine Similarity
The mathematical distance between a query's embedding and your content's embedding in vector space. Embedding distance drives content selection for AI grounding, and keyword matching does not. Content that lands far from the query embedding loses the selection, whatever it ranks in traditional search.
Embedding Feed
A scheduled job pushing new content as embeddings to a vector store, ensuring freshness for RAG assistants.
Zero-Shot Retrieval
The ability of a retriever to surface your page for queries the model never saw during training.
Retrieval-to-Citation Drop-off
The gap between the pages that an AI system retrieves and the pages that it cites in the final answer. An AI platform retrieves far more pages than it cites, and most retrieved pages never appear in the response. Discovery is a prerequisite and not a guarantee. Citation selection depends on title-query alignment, content clarity, readability, and query-type dynamics. The drop-off varies by intent: product-discovery and how-to queries convert at higher rates than comparison or validation queries.
Persistent ID (GUID / GTIN / ISBN)
A globally unique identifier letting different datasets resolve the same entity without confusion.
Data Contract
An explicit schema (often versioned) that defines the fields an API or feed will always supply.
Content Delivery Network (CDN)
A globally distributed cache that serves APIs and structured data closer to users and crawlers, boosting speed and uptime.
API Versioning
A label on each API release (v1, v2…) that lets a consumer, an LLM included, rely on stable fields while you evolve the contract.
Schema Markup Validator
Tools (e.g., Google Rich Results Test) that check JSON-LD syntax and completeness, lowering Schema Error Rate.
Schema Error Rate
The percentage of URLs whose JSON-LD fails validation. High error rates hide facts from crawlers and embedding pipelines.
Accessibility & Semantic HTML (WCAG)
Proper headings, ARIA roles, alt text, and WCAG compliance so humans and multimodal agents can parse content.
Schema.org Markup
JSON-LD tags that declare entities (products, authors, FAQs, etc.) in a format search engines and LLM scrapers can parse.
API Uptime
The percentage of successful (200) responses from your data endpoints. Downtime means missing citations from AI systems and crawlers.
API Latency
The round-trip time of an API call. High latency degrades real-time chat or agent experiences.
Embedding Sync Lag
The time between publishing content and its appearance in your vector store, measured in minutes or hours.
Interactive Tool / Calculator
An on-page widget or a callable API that gives a user an instant calculation. Both raise engagement and task completion.
Semantic Unit
A 50-150 word content block capturing a single concept with explicit subject-predicate-object structure. The atomic unit of passage-level optimization. Splitting combined paragraphs into focused semantic units measurably improves cosine similarity, and adding proper headers compounds the gain.
Model Context Protocol (MCP)
An open standard (Anthropic, November 2024) for connecting AI systems to external tools, data sources, and systems. OpenAI adopted it across every product, ChatGPT desktop included. MCP is the de facto agent-to-tool layer. An AI agent uses it to call an API, query a database, and run an action through one surface.
NLWeb
A Microsoft protocol (co-developed with Schema.org co-founder R.V. Guha) that turns a website into an AI-queryable surface with its existing Schema.org, RSS, and structured data. Every NLWeb instance is also an MCP server. TripAdvisor and O'Reilly Media both run it. Yoast built NLWeb into its Schema Aggregation feature, which creates site-wide endpoints, so an AI agent reads a whole site without a page-by-page crawl.
llms.txt
A proposed plain-text file (like robots.txt) intended to help LLMs understand a site's structure and content. Widespread adoption but no major AI platform has confirmed using it in retrieval. Google's John Mueller confirmed it does not influence search rankings or AI Overview citations. Currently aspirational infrastructure, not a retrieval signal.

Authority & Trust

Entity Stacking
Building 30-50 unique trust signals across trusted third-party sources (citations, social profiles, reference sites, press) before investing in content or links. Google needs these signals before it considers a brand a real entity. The sources carry trust, so a team can build the whole stack quickly and trigger no spam signal.
ClusterUplift
Google groups sites with similar sites and applies collective quality boosts or demotions to the cluster. When the cluster has a quality problem, Google demotes every site in it, including the clean ones. The mechanism explains why a core update hits a whole niche while it leaves sites in other niches alone.
siteFocusScore / siteRadius
Two signals from the Google API leak measuring topical coherence. siteFocusScore quantifies how dedicated a site is to a single topic (specialist vs. generalist). siteRadius measures how much an individual page deviates from the site's central theme. High focus with low radius = strong topical authority signal.
Brand Mention
Your name in trusted publications without a link. Modern ranking systems treat these as implied links. Studies suggest brand search volume is a stronger predictor of AI citations than backlinks.
Author Markup / Bylines
Schema fields or HTML blocks that tie content to a real person, allowing knowledge graphs to attribute expertise. Author topic authority (patent US8458196B1) accumulates per topic. An author who writes repeatedly on a focused topic compounds the score.
Citation Velocity
The rate at which new referring domains cite your content. Spikes often precede ranking gains.
Knowledge Graph
A search-engine database of entities and their relationships. Accurate representation fast-tracks authority in both traditional rankings and LLM answers.
Entity Linking / Disambiguation
Associating a text mention with the correct entity ID (e.g., Apple-fruit vs. Apple-Inc.), preventing authority leaks to competitors.
Third-Party Review
User or expert ratings hosted off-site (e.g., Trustpilot) that act as external validation of quality.
NAP Consistency
Exact match of Name, Address, and Phone across directories. Local SEO and entity disambiguation both depend on it.
User-Generated Content (UGC)
Reviews, comments, or forum posts that add fresh, authentic signals of experience. Reddit has become one of the most visible sites in Google US results, appearing in the vast majority of product review queries.
Knowledge-Graph Triple
A fact stored as (subject → predicate → object), e.g., "Citibank → offersRate → 3.30% APY." The atomic unit of structured knowledge.
Digital Signature / Provenance (C2PA)
Cryptographic metadata that proves who created a file, and whether anyone altered it. The metadata raises trust in an LLM pipeline.
E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness)
Not a single ranking signal but a label for dozens of independent algorithmic features evaluated at document, domain, and entity levels. E-E-A-T enters rankings indirectly. Quality Raters evaluate pages against the E-E-A-T criteria, and those evaluations become training data for Google's RankEmbed models. The models then predict rater-like quality scores at scale. Trust is the most important dimension, and an untrustworthy page scores low whatever its expertise or authority. For AI citation, E-E-A-T acts as a gate rather than a weight, because most AI Overview citations come from sources with strong E-E-A-T signals.
YMYL (Your Money or Your Life)
Google's classification for a query where a low-quality result could harm the user: health, finance, legal, and civic information. YMYL queries receive differential E-E-A-T weighting: Google gives more weight to authoritativeness, expertise, and trustworthiness signals. The quality thresholds ratchet upward over time, and recent core updates cut traffic to major health publishers. YMYL verticals trigger AI Overviews at significantly different rates, with health and legal far more likely to show AIOs than commercial or political queries.

LLM Fundamentals

Large Language Model (LLM)
A neural-network model (e.g., GPT-5, Gemini) trained on massive text corpora and capable of predicting tokens, answering questions, and following instructions.
Token
The smallest unit an LLM processes (≈ one word or punctuation mark). Costs, context limits, and output length are all measured in tokens.
Embedding / Vector Embedding
A fixed-length list of numbers that captures a text's meaning so similar texts sit close together in multi-dimensional space.
Contextual Embedding
An embedding generated with awareness of the surrounding document, and not of the target passage alone. Resolves ambiguity when the meaning of a paragraph depends on its page context. The phrase 'their pricing model' carries no meaning until the reader knows which company the page discusses. Perplexity's pplx-embed-context-v1 is the first major open-source implementation, outperforming prior contextual models by 2-10 percentage points on ConTEB benchmarks.
Bidirectional Encoder
A text model that processes tokens with attention to both preceding and following context, unlike causal (left-to-right) language models. BERT pioneered the approach. Modern retrieval embeddings (pplx-embed, gte-Qwen) convert a causal decoder model into a bidirectional encoder, which produces a richer text representation for search.
Vector Store / Vector DB
A database (e.g., Pinecone, Supabase, Elasticsearch KNN) optimized for storing embeddings and running "nearest-vector" queries.
Context Window
The maximum token count an LLM can "remember" per interaction (prompt + response). Governs chunk sizing in RAG.
Chunking / Text Splitting
The process of splitting longer content into smaller, coherent segments before creating embeddings. Effective chunking follows the natural content boundaries, such as a heading or a paragraph. Each vector then holds one complete idea that an AI system can retrieve as context.
Prompt / System Message
Instructions prepended to the user prompt that set tone, policy, or formatting rules for an LLM conversation.
Temperature
A generation parameter (0-2) controlling randomness: lower = deterministic, higher = creative.
Hallucination
An LLM answer that sounds plausible but is factually wrong because the model filled gaps with guesswork.
Reinforcement Learning from Human Feedback (RLHF)
A training loop that uses human ratings of model outputs as rewards, aligning the LLM with helpful, harmless responses.
Fine-tuning / Supervised Fine-tuning (SFT)
Further training a pre-trained LLM on a smaller, task-specific dataset to adopt domain language, style, or private knowledge.
Multimodal LLM
A model that both consumes and produces text plus other media (images, audio, video), enabling richer search and UX.
Function Calling
When an LLM outputs a JSON payload instructing software to run a tool (e.g., calculateSavingsRate) mid-conversation.
Prompt Engineering
The craft of writing clear, constrained prompts (plus examples) to steer LLM outputs toward desired style and accuracy.
Zero-/Few-/Multi-Shot Prompting
Supplying zero, a few, or many examples in the prompt to guide model reasoning and relevance.
Cost per Token
What you pay an LLM provider for each input/output token. Crucial for budgeting large-scale RAG or generation.
Hallucination Rate
The share of AI answers where the LLM asserts unverified or false information. Monitor the rate whenever AI-generated content or an AI-assisted tool sits inside the product.

Diagnostic Framework

These concepts map to the Clinical Retrieval & Ranking Framework

Binding Constraint
The first layer a site cannot pass in the diagnostic framework. Nothing downstream matters until the team clears the constraint. The most common form of SEO capital destruction: optimizing Layers 4-7 while the site has an unresolved Layer 2 access problem.
Evidence-Builder Loop
Win achievable queries first to build authority priors, then use those priors to compete for harder queries. Topical authority measurably accelerates traffic acquisition. Sequencing matters: authority compounds on earlier wins.
Ceiling vs. Weight
Layer 1 (Eligibility) sets the ranking ceiling, which is the highest position that the domain authority, the penalties, and the YMYL risk allow. Layer 7 (Competition) sets the weight needed to reach that ceiling: the authority gap, the SERP feature concentration, and the differentiation. Misdiagnosing a ceiling problem as a weight problem wastes capital on content and links that can never rank.
Investment Screen
A three-level strategic filter that runs before the diagnostic. Level 1 is channel qualification: is organic search the right channel? Level 2 is page category allocation: is this the right page type for this business model? Level 3 is query-level expected value: does the payoff justify the investment? Each screen must clear before the next one runs.

Strategy & Architecture

Aggregator vs. Integrator
The strategic archetype that determines which SEO levers exist. Aggregator SEO is product-led and inventory-driven, and it uses SEO as the primary growth channel. The aggregator wins on data scale (cost leadership). Integrator SEO is marketing-led, it runs on company-created content, and it uses SEO as a supporting channel. The integrator wins on content quality (differentiation). Programmatic SEO architecture is fundamentally an aggregator play: the data asset is the product.
Product-Led SEO
SEO treated as a product experience rather than a traffic channel, coined by Eli Schwartz. The core principle: build the product in the way search algorithms optimize for. The rendered data is the product that the team optimizes for search. The idea applies directly to a programmatic architecture, where programmatic pages surface structured data at scale.
Search TAM
Total Addressable Market sizing applied to organic search. TAM is all the search volume in the category. SAM is the set of queries you could realistically target, and SOM is the set you can capture with current resources. Standard keyword-volume forecasting is unreliable, and scenario planning with stage gates outperforms a point estimate.
Data Moat
A competitive advantage built from proprietary data that creates self-reinforcing cycles: more data produces better products, which attract more users, which generate more data. Requires 2-3 years of consistent investment before delivering significant advantages. Zillow, TripAdvisor, and NerdWallet all built organic moats through proprietary data plus template infrastructure, not editorial volume.
Content Half-Life
The time it takes for a piece of content to lose half its organic visibility. Has compressed significantly for competitive topics. Refreshing legacy content consistently outperforms exclusive focus on new production. Strategy must allocate a meaningful share of content budget to maintenance.
Content Pruning
Systematic removal or consolidation of pages that drag down site-level quality signals. Embedding-based methodology: generate topic centroids, score all pages against them via cosine similarity, layer in performance data and freshness, then apply kill/keep/review thresholds. A programmatic build needs it most, because a programmatic page can drift off-topic.
Keyword Cannibalization
Internal competition where a search engine swaps rankings between several pages of one site for the same query. Near-duplicate pages also split the ranking signals. Most raw overlap is benign: successful sites average 4.7 ranking URLs per top keyword, and high-authority sites often hold two top positions at once. The harm concentrates on lower-authority sites and commercial-intent queries. There, a consolidation of truly duplicate pages recovers traffic. Diagnose from URL swapping and split clicks in Search Console, not from overlap alone.

Want to see how this applies in practice?

The glossary covers the vocabulary. The patterns go deeper. Real architectural problems from real audits, with the diagnosis and fix.

See Tech SEO Patterns