# Revised v2-correcties en retired claims Master v1.0 blijft de canonieke kennisbron. Revised v2 begrenst overclaims en bewaakt regressies; het overschrijft de Master niet automatisch. ## COR-RV2-001 — What is Semantic SEO? **Status:** UPDATED **Doeltermen:** TERM-001 **Evidencecontext:** Evidence: Industry / framework term; supported by search and retrieval concepts Semantic SEO is the practice of researching, structuring and publishing information around meaning, entities, relationships, context and user tasks rather than relying only on isolated keywords. 2026 correction. Google does not publish one ranking system called Semantic SEO. Google now documents generative Search features that use retrieval, grounding and model-generated query fan-out while still relying on foundational Search systems and SEO practices. This extends the environment in which semantic content operates, but it does not create a proven universal second scoring layer. [S01 S02] **Praktisch gebruik:** Practical use. Build a clear information architecture, answer real user tasks, use stable product and organization data, add original evidence, and keep pages technically accessible. Treat AI citations as an additional visibility outcome, not a replacement for rankings, clicks or conversions. [S01 S19] **Niet overclaimen:** Do not overclaim. Do not promise that a semantic content network automatically earns rankings or citations. Do not assume the same retrieval pipeline or weighting applies across Google, ChatGPT, Bing, Perplexity and Claude. ## COR-RV2-002 — What is an entity? **Status:** UPDATED **Doeltermen:** TERM-002 **Evidencecontext:** Evidence: NLP, knowledge representation and search terminology An entity is a distinct thing or concept that can be identified or distinguished, such as a person, organization, place, product, event or abstract concept. [S04] 2026 correction. The text that names an entity is an entity mention, not the entity itself. Modern systems may also perform entity typing, linking, resolution and disambiguation. A product model should ideally have a stable name and identifier so that facts from different variants are not mixed. [S06] **Praktisch gebruik:** Practical use. Name the entity explicitly when a sentence would otherwise be ambiguous or when a passage may be reused outside its immediate context. Natural pronouns and shorter references remain acceptable when the reference is clear. **Niet overclaimen:** Do not overclaim. Do not repeat the full entity name in every sentence. Do not assume that a name alone causes recognition in a knowledge graph or guarantees an AI citation. ## COR-RV2-003 — What is a Triple? **Status:** REWRITTEN **Doeltermen:** TERM-003A, TERM-003B **Evidencecontext:** Evidence: W3C open standard; patent background An RDF triple is a statement with three components in this order: subject, predicate and object. Example: (For Motion On Four, has motor power, 1000 watts). A set of RDF triples forms an RDF graph. [S04] 2026 correction. The earlier examples "Tom Hanks played" and "Tom Hanks said" were incomplete because the object was missing. EAV - Entity, Attribute, Value - is a separate data-model pattern and should not be presented as the same standard. Google has patented graph-indexing techniques involving triples, but that does not prove every web page is stored as RDF or EAV. [S04] **Praktisch gebruik:** Practical use. Write important facts completely and unambiguously. Where relevant, include the entity, property, value, unit, variant, condition, source and date. This improves factual clarity and data reuse even without making an extraction claim. **Niet overclaimen:** Do not overclaim. A complete factual sentence is not automatically an RDF triple in technical syntax, and a triple-shaped sentence is not guaranteed to be extracted, ranked or cited by a search or AI system. ## COR-RV2-004 — What is Topical Authority? **Status:** REWRITTEN **Doeltermen:** TERM-004 **Evidencecontext:** Evidence: Industry term; Google news system; Semantic SEO framework term Topical Authority is an industry term for the degree to which a source is perceived as knowledgeable, useful and trustworthy within a defined topic. 2026 correction. Three meanings must be separated: the broad SEO term; Google's documented topic-authority system for certain news queries; and Koray Tugberk Gubur's broader Semantic SEO framework involving topical maps, source context, coverage and historical data. Google publishes no universal Topical Authority score or formula for every site. [S03 S23] **Praktisch gebruik:** Practical use. Define a realistic topic boundary, cover the questions and relationships required for real user tasks, publish accurate and original evidence, maintain content, and earn credible external recognition where appropriate. [S03 S19] **Niet overclaimen:** Do not overclaim. Do not present Topical Authority as "coverage plus information gain", a fixed score, a page-count target or a guarantee. A large number of pages can increase overlap and low-value content instead of authority. ## COR-RV2-005 — What is a Query Network? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-005 **Evidencecontext:** Evidence: Semantic SEO planning framework; related to query processing research A Query Network is a planning model that connects related queries, intents, entities, conditions, user tasks and follow-up questions. 2026 correction. Query Aspect, Query Definition and Query Theme can remain useful fields within the framework, but they are not established universal fields used by every search engine. Google now documents model-generated query fan-out in specific generative Search processes. That is a platform behavior, not proof that every internal query model is identical to the framework. [S01 S05 S23] **Praktisch gebruik:** Practical use. Use a Query Network to discover missing questions and relationships. Decide whether each item belongs on an existing page, in a section, table, video, tool or genuinely separate page. **Niet overclaimen:** Do not overclaim. Do not create one page for every query or every observed fan-out. Do not treat a tool-generated network as a copy of a search engine's internal representation. ## COR-RV2-006 — What is a Semantic Content Network? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-006 **Evidencecontext:** Evidence: Content architecture and internal information design A Semantic Content Network is an organized set of pages, passages, media and tools whose topics, entities and user tasks are meaningfully connected. 2026 correction. The concept remains useful as content architecture. Patent literature describes passage-level and theme-based search techniques, but it does not prove that every site is scored as one explicit semantic network or that every missing subtopic becomes a competitor-owned "theme bucket". [S08 S23] **Praktisch gebruik:** Practical use. Give each page a distinct role, connect related information with useful links, and prevent contradictory facts. Use videos, images, tables and tools where they solve the task better than another article. [S19] **Niet overclaimen:** Do not overclaim. Do not manufacture links or pages merely to make the network look dense. A network of overlapping, generic pages is weaker than a smaller set of clearly differentiated resources. ## COR-RV2-007 — What is Microsemantics? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-007 **Evidencecontext:** Evidence: Local meaning with linguistic foundations Microsemantics examines how wording, syntax, negation, modality, punctuation, units, argument roles and references affect meaning at sentence and passage level. 2026 correction. Small changes can alter meaning, but not every punctuation mark or word-order change produces a search effect. An answer-span patent shows one possible system for selecting consecutive words from a document. It does not prove that punctuation determines universal extractability or that every AI answer copies a verbatim span. [S06] **Praktisch gebruik:** Practical use. Check that a claim states who or what it concerns, what is asserted, the value and unit, and any necessary condition or exception. Make uncertainty and limitations explicit. **Niet overclaimen:** Do not overclaim. Do not optimize punctuation, sentence length or word order to an invented score. Clarity, accuracy and natural language remain the objective. ## COR-RV2-008 — What is Macrosemantics? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-008 **Evidencecontext:** Evidence: Document and site structure; discourse and information architecture Macrosemantics examines meaning and coherence at section, page and site level, including topic boundaries, information order, headings, internal links, navigation and recurring entity relationships. 2026 correction. Descriptive headings can help users and systems understand sections. Thematic-search patent literature describes passages that may be paragraphs or may be associated with headers. It does not establish a universal heading-vector score, a compulsory heading-query match score or one paragraph per heading. [S08] **Praktisch gebruik:** Practical use. Use headings that accurately describe the section, place information in a logical sequence and make internal links useful for the reader's next task. [S01] **Niet overclaimen:** Do not overclaim. Do not force every heading to repeat the primary query or entity. Do not use sitewide n-gram counts as a substitute for human review of meaning and overlap. ## COR-RV2-009 — What is Topical Coverage? **Status:** UPDATED - FRAMEWORK / AUDIT MODEL **Doeltermen:** TERM-009 **Evidencecontext:** Evidence: Content planning and quality assessment Topical Coverage is the extent to which a source responsibly addresses the questions, entities, relationships, conditions, evidence and tasks required within a defined topic boundary. 2026 correction. Coverage is not the number of words or URLs. It must be evaluated against the intended audience and task. Information Gain can be used as a separate practical question - what does this source add? - but the patent does not establish a universal formula in which coverage and novelty are the two halves of all ranking. [S10 S19] **Praktisch gebruik:** Practical use. Map the essential basics, comparisons, exceptions and decisions. Add original measurements, experience or tools where they genuinely improve the answer. Remove or merge pages that do not have a distinct role. **Niet overclaimen:** Do not overclaim. Do not assume complete consensus coverage is useless. Users often need a reliable explanation of established facts before unique information becomes meaningful. ## COR-RV2-010 — What is Historical Data for SEO? **Status:** REWRITTEN **Doeltermen:** TERM-010 **Evidencecontext:** Evidence: Court-record evidence; patent background; analytics distinction Historical Data is information accumulated over time about queries, documents, links, changes, search-result interactions and performance. 2026 correction. Public court records describe Google systems such as NavBoost and Glue using query, click and other SERP-interaction data. Older Google patent literature also describes possible uses of document and link history. These sources do not make Google Analytics bounce rate or engagement time direct ranking factors, and they do not support a special "position 94" rule. [S11 S12 S13] **Praktisch gebruik:** Practical use. Use historical data to monitor query demand, content decay, outdated facts, changing user needs and the effect of meaningful updates. Keep an experiment log so that multiple simultaneous changes are not incorrectly attributed to one cause. **Niet overclaimen:** Do not overclaim. Do not try to "optimize hovers" or manipulate clicks. Search-result interaction data is a platform dataset, not a direct writer control. Chrome-related evidence should be treated separately and cautiously; the broad Gmail authority claim should not be used. ## COR-RV2-011 — What is Relevance for Information Retrieval? **Status:** REWRITTEN **Doeltermen:** TERM-011A **Evidencecontext:** Evidence: Information Retrieval research Relevance is the degree to which information is useful or appropriate for a query, information need, task, user and context. 2026 correction. Modern systems can combine lexical or sparse retrieval, dense retrieval with embeddings, hybrid retrieval, candidate generation and reranking. BM25, TF-IDF and exact terms remain useful. Dense retrieval and RAG did not close the lexical chapter. [S14 S15 S16] **Praktisch gebruik:** Practical use. Write with accurate terminology, names, model codes, numbers and units where exactness matters, while also explaining meaning and relationships naturally. Make passages clear enough to answer the intended task. **Niet overclaimen:** Do not overclaim. Do not reduce relevance to one sequence of "retrievable, extractable, citable" or imply that ranking is merely an entry ticket. Retrieval, ranking, citation, mention and recommendation are related but distinct outcomes. ## COR-RV2-012 — What are Representative and Represented Queries? **Status:** REWRITTEN **Doeltermen:** TERM-012A, TERM-012B **Evidencecontext:** Evidence: Query clustering terminology; framework-specific counterpart A representative query is a query selected to stand for a group or cluster of related queries. Depending on the method, it can be the most frequent query, a central query, a medoid or a generated summary of the cluster. 2026 correction. It is not always the broadest query. "Represented Query" can remain as a framework term for a query represented by the selected cluster label, but it is not a universally standardized opposite of representative query. Search systems may also rewrite, expand or generate related queries at runtime. [S05 S09] **Praktisch gebruik:** Practical use. Cluster queries by user task, intent, entity and required answer. Use factual consistency across pages, but do not force every answer into identical question-and-answer phrasing. **Niet overclaimen:** Do not overclaim. Do not conclude that every query in a cluster deserves the same page or that every generated variant must be targeted separately. ## COR-RV2-013 — What is Semantic Distance? **Status:** UPDATED **Doeltermen:** TERM-013 **Evidencecontext:** Evidence: Graph, ontology and vector-space terminology Semantic Distance expresses how far apart meanings or concepts are under a chosen representation and distance measure. 2026 correction. Vector distance, graph distance, taxonomic or ontological distance and task-specific contextual distance remain different valid models. Embeddings have become important for dense retrieval, but they have not universally replaced knowledge graphs or ontologies. [S04 S14] **Praktisch gebruik:** Practical use. State which representation and measure are being used in an analysis. Treat SEO tool distances as model outputs, not objective facts about meaning. **Niet overclaimen:** Do not overclaim. PageRank, clicks and query volume may affect other ranking or evaluation processes, but they are not part of the general definition of semantic distance. ## COR-RV2-014 — What is Semantic Similarity? **Status:** UPDATED **Doeltermen:** TERM-014A **Evidencecontext:** Evidence: NLP and machine-learning terminology Semantic Similarity is the degree to which two words, passages or other representations have similar meanings. 2026 correction. Embeddings and cosine similarity are common methods, but similarity can be modeled in several ways. A generative-summary patent describes optional methods for checking whether a summary portion can be verified by document content, including source identifiers or embedding comparisons. It does not prove that all citations are assigned by one nearest-sentence embedding match. [S07 S14] **Praktisch gebruik:** Practical use. Write naturally and precisely. Use terminology that users and sources actually use, but do not imitate a guessed AI sentence in the hope of becoming its closest match. **Niet overclaimen:** Do not overclaim. Similarity is not the same as relatedness or relevance. "Battery" and "range" are strongly related, but they do not mean the same thing. ## COR-RV2-015 — What is Semantic Relevance? **Status:** UPDATED **Doeltermen:** TERM-015 **Evidencecontext:** Evidence: Information Retrieval and content-quality terminology Semantic Relevance is the degree to which information meaningfully helps answer a specific query or complete a specific task in context. [S14 S15] 2026 correction. A page can use similar words and still be irrelevant, or use different words and provide the best answer. Relevance may depend on user intent, product variant, location, date, legal jurisdiction and evidence quality. **Praktisch gebruik:** Practical use. Define the user task first. Then select the facts, examples, comparisons and next steps needed for that task. **Niet overclaimen:** Do not overclaim. Do not treat the presence of semantically related terms as proof that the page is useful or complete. ## COR-RV2-016 — What is Natural Language Processing? **Status:** UPDATED **Doeltermen:** TERM-016A **Evidencecontext:** Evidence: Established scientific field Natural Language Processing, or NLP, is the field concerned with computational analysis, representation, retrieval, extraction, transformation and generation of human language. 2026 correction. LLMs, transformers, embeddings, Named Entity Recognition, entity linking, relation extraction, classification, translation and summarization are all part of modern NLP. NLP was not superseded by LLMs. Older statistical and rule-based methods also remain useful for many tasks. [S14 S16] **Praktisch gebruik:** Practical use. Use NLP concepts to understand what systems can attempt and where errors can occur. Validate extracted entities and facts against sources and variants. **Niet overclaimen:** Do not overclaim. Do not say a model "understands" a fact merely because it produces fluent text, and do not treat a library list such as NLTK or spaCy as the definition of NLP. ## COR-RV2-017 — What is Sliding Window in NLP? **Status:** BACKGROUND **Doeltermen:** TERM-017A **Evidencecontext:** Evidence: Technical processing technique A sliding window processes overlapping or sequential segments of tokens or other units. Its behavior depends on window size, step size and overlap. 2026 correction. Sliding window is distinct from chunking and from a model's context window. Some systems use windowed attention or overlapping segments; others do not. This technique does not create a direct SEO writing rule. [S01] **Praktisch gebruik:** Practical use. Use clear section structure and avoid ambiguous references where important facts may be read independently, but preserve natural document flow. **Niet overclaimen:** Do not overclaim. Do not require every paragraph to repeat the entity name, ban phrases such as "as explained above", or assume a fixed chunk size used by all search and AI systems. ## COR-RV2-018 — What is Sequence Modeling in NLP? **Status:** BACKGROUND - UPDATED **Doeltermen:** TERM-018 **Evidencecontext:** Evidence: NLP and machine-learning terminology Sequence Modeling represents or predicts ordered data such as tokens, tags, actions or sentences. It supports generation, translation, classification, tagging, speech recognition and many other tasks. 2026 correction. Transformers are dominant in many modern language systems, but RNNs and LSTMs are not simply "history". The practical content lesson is narrower: word order, grammatical roles, negation and reference can change meaning. [S16] **Praktisch gebruik:** Practical use. Write grammatically correct sentences and check that subject, action and object are unambiguous. Use active or passive voice according to clarity, not an assumed ranking preference. **Niet overclaimen:** Do not overclaim. Do not treat internal model architecture as a page-level optimization target. ## COR-RV2-019 — What is a Central Entity? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-019 **Evidencecontext:** Evidence: Semantic SEO planning framework A Central Entity is the main entity used as an organizing anchor within a Semantic SEO content plan. A neutral page-level alternative is "primary entity". 2026 correction. Entity clarity can matter at page and passage level, especially when a passage is ambiguous outside its immediate context. This does not mean the entity must be repeated literally in every section or anchor text. [S23] **Praktisch gebruik:** Practical use. Record the primary entity, relevant variants, identifiers, attributes and relationships before writing. Mention it again only where a reader or retrieval system could otherwise confuse the reference. **Niet overclaimen:** Do not overclaim. Do not use fixed entity-density targets or force the same phrase into every heading and link. ## COR-RV2-020 — What is Source Context? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-020 **Evidencecontext:** Evidence: Semantic SEO framework; source and provenance concepts Source Context describes who the source is, what role it has, which audience it serves, what expertise and evidence it has, and what commercial or institutional interests may affect the content. 2026 correction. Source Context is not the same as provenance. Provenance tracks where a specific claim, measurement, image or conclusion came from, who created it, by which method and on what date. A generative-search patent may use document confidence measures, but it does not establish one universal source-trust formula. [S07 S24] **Praktisch gebruik:** Practical use. Show real authorship, review responsibility, methods, commercial relationships, update dates and primary sources. Keep product and organization information consistent across channels. **Niet overclaimen:** Do not overclaim. Do not rely on a brand description alone as proof that every claim is trustworthy. ## COR-RV2-021 — What is Central Search Intent? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-021 **Evidencecontext:** Evidence: Content strategy and search-intent analysis Central Search Intent is a framework term for the main user outcome or strategic search purpose around which a page or content network is organized. 2026 correction. A website can have a broad user mission, while each page has a primary task and may have secondary intents. Conversational systems can generate follow-up or drill-down queries, but one intent does not need to be literally repeated across navigation, footer and every page. [S09] **Praktisch gebruik:** Practical use. Define the page's primary user task, the secondary questions that belong on the same page and the next action the user should be able to take. **Niet overclaimen:** Do not overclaim. Do not use only the four traditional intent labels as a complete model of a complex decision journey. ## COR-RV2-022 — What is a Knowledge Domain? **Status:** REWRITTEN **Doeltermen:** TERM-022 **Evidencecontext:** Evidence: Knowledge organization and framework terminology A Knowledge Domain is a defined field with its own entities, terminology, relationships, evidence standards, user tasks, risks, update needs and typical information formats. 2026 correction. The earlier broad claim that search engines combine Gmail, Chrome activity and social-media data to determine a source's authority within a knowledge domain is not supported in that general form and is removed. Domain-specific search behavior can exist without such a universal authority score. [S12 S13] **Praktisch gebruik:** Practical use. Document the domain's terminology, legal or safety boundaries, primary sources, user groups and acceptable evidence. Adapt depth and format to the task. **Niet overclaimen:** Do not overclaim. Do not infer a hidden Knowledge Domain score from bounce rate, layout or private user data. ## COR-RV2-023 — What is a Contextual Domain? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-023 **Evidencecontext:** Evidence: Semantic SEO planning framework; related to disambiguation A Contextual Domain is the meaning and use environment in which a term, entity, query or task should be interpreted. 2026 correction. Context can help distinguish Apple the company from apple the fruit. This may involve lexical cues, entities, user state, query history, classifiers, knowledge resources or language models. It is not performed only by LLMs. [S09] **Praktisch gebruik:** Practical use. Make the relevant context explicit when it changes the answer: product variant, country, date, user type, use case, budget, risk or physical condition. **Niet overclaimen:** Do not overclaim. Do not claim that dwell time or bounce rate directly calculates a contextual-domain score. ## COR-RV2-024 — What is a Contextual Layer? **Status:** UPDATED - FRAMEWORK TERM **Doeltermen:** TERM-024 **Evidencecontext:** Evidence: Semantic SEO planning framework A Contextual Layer is a focused refinement inside a broader context, such as audience, situation, location, time, expertise, product variant, legal jurisdiction or risk level. 2026 correction. Contextual layers can inspire some fan-out queries, but they are not the same thing. A fan-out query is generated by a system for retrieval; a contextual layer is a planning lens. Multiple layers may be handled in one section, table, filter, tool or page. [S01 S09] **Praktisch gebruik:** Practical use. Create separate content only when the contextual qualification materially changes the task, facts, evidence, decision criteria or user action. **Niet overclaimen:** Do not overclaim. Do not build one page for every possible modifier or generated subquery. ## COR-RV2-025 — What is Query Fan-out? **Status:** ADDED **Doeltermen:** QRY-015 **Evidencecontext:** Evidence: Official Google documentation; patent background kept separate Query fan-out is the generation of several concurrent, related queries to retrieve additional information needed for a broader user question. [S01] 2026 correction. Google publicly documents model-generated fan-out queries for generative Search features. US12265560 describes older subquery generation from a compound query, with a 2017 priority date, and does not describe an LLM. US20240289407 separately describes LLM-generated synthetic and drill-down queries in a stateful chat patent application. [S05 S09] **Praktisch gebruik:** Practical use. Use likely follow-up questions to audit completeness. Answer them in the most appropriate existing resource instead of automatically creating a new page. **Niet overclaimen:** Do not overclaim. Do not treat every fan-out as visible, stable or individually targetable. The same prompt may lead to different retrieval activity over time or across platforms. ## COR-RV2-026 — What is Answer Span Extraction? **Status:** ADDED - PATENT-DESCRIBED MECHANISM **Doeltermen:** IR-017, IR-018 **Evidencecontext:** Evidence: Google patent, not confirmed as universal current behavior Answer Span Extraction selects a sequence of tokens from a document that answers a question. [S06] 2026 correction. US11481646 describes a neural system that selects consecutive words from one electronic document and can merge candidate spans that are near-duplicates or refer to the same entity. It does not say that repetition across an entire website is rewarded, nor that all generative answers copy text verbatim. [S06] **Praktisch gebruik:** Practical use. State key facts clearly and keep tables, body text and product data consistent. A concise answer can help users even when no extractive system is involved. **Niet overclaimen:** Do not overclaim. Do not repeat the same sentence across pages to chase an assumed span signal. ## COR-RV2-027 — What is a Generative Summary or AI Overview? **Status:** ADDED **Doeltermen:** AIS-012 **Evidencecontext:** Evidence: Official Google documentation; patent background A generative summary is a model-written response assembled from retrieved information. Google AI Overviews and AI Mode can show links to web pages that support information in the response. [S01] 2026 correction. Google documents retrieval, grounding and fan-out in current generative Search. US11769017 describes possible methods for generating summaries, selecting documents, linking portions to sources and checking verifiability. It includes several possible mechanisms; embedding comparison is optional, not the sole universal citation rule. [S07] **Praktisch gebruik:** Practical use. Publish accurate, accessible and current information that can support a user answer. Make important conditions and limitations visible and maintain the underlying facts. **Niet overclaimen:** Do not overclaim. Do not write "the sentence the AI will write" or assume a citation proves endorsement, ranking position or complete factual support. ## COR-RV2-028 — What is Thematic Search or Result Clustering? **Status:** ADDED - PATENT-DESCRIBED MECHANISM **Doeltermen:** TERM-005A, TERM-011E, TERM-011F **Evidencecontext:** Evidence: Google patent, not proof of universal deployment Thematic search organizes retrieved results or passages around identified themes or subthemes. [S08] 2026 correction. US12158907 describes generating passage summaries, clustering them into themes and presenting thematic search results. A passage may be a paragraph, section or content associated with a header. The patent does not establish one compulsory heading-plus-paragraph format or a universal heading-query score. **Praktisch gebruik:** Practical use. Use descriptive headings and cover genuinely distinct decision factors. Keep sections coherent and avoid mixing several unrelated answers under one vague heading. **Niet overclaimen:** Do not overclaim. Do not manufacture a fixed theme set from one patent or assume every missing heading is a lost ranking bucket. ## COR-RV2-029 — What is Stateful or Conversational Search? **Status:** ADDED **Doeltermen:** QRY-017 **Evidencecontext:** Evidence: Patent application and general conversational-search concepts Stateful search uses information from earlier turns or session context to interpret a current query and continue the task. 2026 correction. US20240289407 describes possible use of user state, rewritten queries, synthetic queries and suggested next steps. It is a patent application, not proof that every detail is deployed in Google AI Mode or other products. [S09] **Praktisch gebruik:** Practical use. Map common follow-up questions and make navigation between orientation, comparison, decision and action easy. Keep location and date explicit when they change the answer. **Niet overclaimen:** Do not overclaim. Do not assume every user sees the same follow-up chain or that a state embedding is the universal implementation. ## COR-RV2-030 — What is Information Gain? **Status:** ADDED - PATENT CONCEPT AND PRACTICAL QUALITY QUESTION **Doeltermen:** TERM-009 **Evidencecontext:** Evidence: Google patent; not a confirmed universal ranking factor Information Gain is the additional useful information a document provides relative to information already presented or consumed in a given context. [S10] 2026 correction. US12013887 describes contextual information-gain scores for new documents relative to documents already shown to a user, including possible reranking in search or assistant experiences. It does not confirm that every page receives a universal score or that consensus information automatically scores near zero. [S10] **Praktisch gebruik:** Practical use. Add real tests, measurements, expert observations, transparent limitations, tools or clearer synthesis where they improve the user's decision. Still explain the necessary basics accurately. **Niet overclaimen:** Do not overclaim. Do not invent novelty, publish unsupported first-party claims or treat uniqueness as a substitute for correctness. ## COR-RV2-031 — What is Embedding-based Retrieval? **Status:** ADDED **Doeltermen:** TERM-011C, TERM-016C, IR-016 **Evidencecontext:** Evidence: Information Retrieval research Embedding-based or dense retrieval represents queries and documents or passages as vectors and retrieves items based on learned similarity. [S14 S15] 2026 correction. Dense retrieval is important, but many systems also use lexical and hybrid methods. Exact names, numbers, product codes and phrases can remain essential. **Praktisch gebruik:** Practical use. Use accurate language, synonyms where natural and clear relationships. Preserve exact identifiers and technical terms where precision matters. **Niet overclaimen:** Do not overclaim. Do not replace keyword research with an assumed vector model or treat a tool's similarity score as ground truth. ## COR-RV2-032 — What is Hybrid Retrieval? **Status:** ADDED **Doeltermen:** TERM-011D **Evidencecontext:** Evidence: Information Retrieval research Hybrid Retrieval combines lexical or sparse methods with dense or embedding-based methods, often followed by reranking. [S15] 2026 correction. Research shows that sparse, dense and combined retrieval can each be useful depending on the data and task. Hybrid retrieval is a more accurate general description of modern retrieval than "keywords are over". **Praktisch gebruik:** Practical use. Keep both exact factual language and semantically complete explanation. Test search behavior with real queries rather than optimizing to one assumed retrieval type. **Niet overclaimen:** Do not overclaim. Do not claim every platform uses the same hybrid stack or weights. ## COR-RV2-033 — What is Retrieval-Augmented Generation (RAG)? **Status:** ADDED **Doeltermen:** AIS-010 **Evidencecontext:** Evidence: Scientific term; also used in official platform explanations RAG combines information retrieval with a generative model so that the generated response can use externally retrieved material. [S16 S01] 2026 correction. RAG quality depends on the source collection, query, retrieval, passage selection, context allocation and generation. Retrieved material can still be incomplete, misread or incorrectly attributed. **Praktisch gebruik:** Practical use. Maintain source quality, update dates, clear variants and accessible content. Evaluate retrieval and answer accuracy separately. **Niet overclaimen:** Do not overclaim. Do not assume every LLM answer uses RAG or that retrieval automatically prevents hallucinations. ## COR-RV2-034 — What is Grounding? **Status:** ADDED **Doeltermen:** AIS-011 **Evidencecontext:** Evidence: Official platform terminology and AI-system design Grounding means basing an AI response on retrieved, current or otherwise verifiable information rather than only on model parameters. [S01 S16] 2026 correction. Grounding can improve traceability, but the system may still select the wrong source, misunderstand a condition or combine incompatible facts. **Praktisch gebruik:** Practical use. Publish sourceable claims, clear conditions, dates and provenance. Check how platforms actually represent the information. **Niet overclaimen:** Do not overclaim. Do not treat a source link as proof that every sentence in the answer is fully supported. ## COR-RV2-035 — What are Citation and Attribution in AI Search? **Status:** ADDED **Doeltermen:** AIS-014, AIS-015 **Evidencecontext:** Evidence: Platform output and measurement terminology A citation links or points to a source used or presented with an AI response. Attribution associates a statement, idea, image or fact with a source or creator. 2026 correction. Citation, brand mention, recommendation, ranking and referral traffic are different outcomes. A cited URL may not be prominently shown, and a brand may be mentioned without its own page being cited. [S20 S21] **Praktisch gebruik:** Practical use. Measure citations, cited pages, brand mentions, factual accuracy, referral traffic and conversions separately. **Niet overclaimen:** Do not overclaim. Do not use citation count alone as a proxy for authority or commercial value. ## COR-RV2-036 — What is Source Eligibility? **Status:** ADDED - PRACTICAL TERM **Doeltermen:** AIS-013 **Evidencecontext:** Evidence: Technical access, indexing, policy and quality conditions Source Eligibility is the set of conditions that allows a page or other resource to be considered for retrieval, display or citation by a particular search or AI system. 2026 correction. Eligibility can depend on crawl access, indexability, supported formats, source quality, freshness, query relevance and product-specific controls. Eligibility is not the same as selection. [S01 S22] **Praktisch gebruik:** Practical use. Manage robots rules, noindex, canonicalization, rendering, sitemaps and platform-specific crawlers deliberately. Keep important information available in machine-readable text as well as media. **Niet overclaimen:** Do not overclaim. Do not assume allowing a bot guarantees retrieval or citation. ## COR-RV2-037 — What is Agentic or Reasoning Search? **Status:** ADDED **Doeltermen:** AIS-012, AIS-016 **Evidencecontext:** Evidence: Research and emerging industry term Agentic search describes systems that can iteratively plan, issue searches, inspect results, reformulate queries and continue until a task or stopping condition is reached. [S17 S18] 2026 correction. Search-R1 is a research system that trains an LLM to generate multiple searches during reasoning. ASPIRE is a separate selective-prediction framework. These studies show possible techniques, not a confirmed common production pipeline for all search products. **Praktisch gebruik:** Practical use. Make claims easy to verify and keep authoritative primary sources available. Design pages and tools so a human or agent can complete the task without hidden steps. **Niet overclaimen:** Do not overclaim. Do not claim that every reasoning engine cross-checks every statement or always refuses low-confidence sources. ## COR-RV2-038 — What are GEO and AEO? **Status:** ADDED - INDUSTRY TERMS **Doeltermen:** AIS-021, AIS-022 **Evidencecontext:** Evidence: Industry terminology; official Google position GEO, Generative Engine Optimization, and AEO, Answer Engine Optimization, are industry labels for work intended to improve visibility in generated or direct-answer experiences. 2026 correction. Google states that optimization for its generative Search features is still SEO and does not require a separate set of special AI rules. Other platforms have their own access, retrieval and citation behavior. [S01] **Praktisch gebruik:** Practical use. Use the terms when they help organize work, but keep foundational SEO, content quality, technical access, evidence and user experience at the center. **Niet overclaimen:** Do not overclaim. Do not sell one universal GEO formula, fixed chunk size, AI schema or citation score. ## COR-RV2-039 — What is AI Visibility? **Status:** ADDED **Doeltermen:** AIS-023 **Evidencecontext:** Evidence: Measurement term; platform reports AI Visibility is the observable presence of a page, source, brand or product in generative and answer-based experiences. 2026 correction. Visibility can include impressions, cited URLs, citations, mentions, recommendation share, grounding queries and referral traffic. Google and Bing now expose different subsets of these metrics, and they are not directly interchangeable. [S20 S21 S22] **Praktisch gebruik:** Practical use. Build a stable prompt and query set, test over time and record platform, language, location and date. Connect visibility to factual accuracy and business results. **Niet overclaimen:** Do not overclaim. Do not treat one prompt, one citation or one third-party tool score as a reliable overall measure. # Appendix: ingetrokken of begrensde draftclaims ## RET-RV2-001 **Oude claim:** US12265560 describes an LLM writing subqueries **Besluit:** Retired. The patent has a 2017 priority date and describes non-LLM subquery generation. Current model-generated fan-out is instead supported by Google's official 2026 guide and separately by a later stateful-chat patent application. **Runtimeactie:** block ## RET-RV2-002 **Oude claim:** A clean triple is precisely the shape a generative engine extracts **Besluit:** Narrowed. Complete facts are useful, but no source proves a universal triple-shaped extraction rule. **Runtimeactie:** replace_with_boundary ## RET-RV2-003 **Oude claim:** Information Gain is a granted universal ranking signal **Besluit:** Narrowed. The patent describes contextual scoring relative to information already shown to a user. **Runtimeactie:** replace_with_boundary ## RET-RV2-004 **Oude claim:** The Blind Librarian chapter is closed **Besluit:** Retired. Sparse, dense and hybrid retrieval remain active. **Runtimeactie:** block ## RET-RV2-005 **Oude claim:** Citation assignment is exactly embedding similarity **Besluit:** Retired. The patent describes several possible verification and source-linking mechanisms. **Runtimeactie:** block ## RET-RV2-006 **Oude claim:** Header-defined passages are the unit and heading-query match is scored **Besluit:** Narrowed. The patent allows several passage forms and does not establish a universal heading score. **Runtimeactie:** replace_with_boundary ## RET-RV2-007 **Oude claim:** NLP is superseded **Besluit:** Retired. LLMs and transformers are part of NLP. **Runtimeactie:** block ## RET-RV2-008 **Oude claim:** Every important passage must be completely self-sufficient **Besluit:** Narrowed. Important sections should be clear, but natural cross-reference and document flow remain valid. **Runtimeactie:** replace_with_boundary ## RET-RV2-009 **Oude claim:** RNN/LSTM are history **Besluit:** Retired. Transformers dominate many uses, but sequence models remain valid technical approaches. **Runtimeactie:** block ## RET-RV2-010 **Oude claim:** Knowledge Domain stands unchanged **Besluit:** Retired. The broad Gmail-Chrome-social authority claim was unsupported and has been removed. **Runtimeactie:** block ## RET-RV2-011 **Oude claim:** Contextual layers are executable fan-out queries **Besluit:** Narrowed. They can overlap, but are different concepts and do not justify one page per query. **Runtimeactie:** replace_with_boundary ## RET-RV2-012 **Oude claim:** Reasoning engines verify every claim and abstain when confidence is low **Besluit:** Narrowed. Research demonstrates possible methods; commercial implementation varies and is not fully public. **Runtimeactie:** replace_with_boundary