The GEO glossary for understanding how AI finds and cites you
Clear definitions covering generative search, RAG, retrieval, citations and measurement. It includes vocabulary used across our corpus of 369 GEO research studies, with a dedicated page for every concept.
Showing 92 terms
GEO fundamentals
Core concepts in generative search, answer engines and language models.
GEO
Generative Engine OptimizationGEO is the optimization of a brand and its sources to earn presence, mentions and citations in AI-generated answers. It covers content, authority, technical access, external distribution and measurement.
Read the full definitionAEO
Answer Engine OptimizationAEO prepares content so an answer engine can extract a clear, direct response. It overlaps with GEO, although AEO emphasizes answering questions while GEO focuses on visibility inside generative answers.
Read the full definitionLLM SEO
Search optimization for large language modelsLLM SEO is a practical label for work that improves how language models find, understand and represent a brand. In the market it is often used as a rough synonym for GEO.
Read the full definitionAI search
Search assisted or generated by artificial intelligenceAI search returns a synthesized answer, conversation or recommendation rather than only ranking links. It may use model memory, live web search or both.
Read the full definitionGenerative engine
A system that retrieves or uses knowledge and generates an answerA generative engine combines an AI model with search, retrieval or tools to write a new answer. Its output may include citations, links, products, maps or recommendations.
Read the full definitionAnswer engine
A system that provides a direct answer to a queryAn answer engine tries to resolve the query within its own interface. It may be generative or rely on snippets, knowledge bases and structured results.
Read the full definitionZero-click search
A query resolved without visiting an external resultA zero-click search ends inside the search or assistant interface because the user gets the needed information there. In AI, brand visibility can exist even when no website session follows.
Read the full definitionAI Overviews
AI-generated summaries in Google SearchAI Overviews are generated answers shown by Google for certain searches and may link to supporting sources. Their composition and sources can vary by query, location and time.
Read the full definitionGoogle AI Mode
Google's conversational search experienceGoogle AI Mode expands search into a conversational experience with generated answers, follow-up questions and source links. It should not be measured as if it were identical to AI Overviews.
Read the full definitionLarge language model
LLMA large language model learns patterns from very large text collections to predict and generate language. It can answer from its parameters or connect to search, RAG and external tools.
Read the full definitionPrompt
An instruction or query sent to a modelA prompt is the text and context that guide a model's answer. In GEO it represents a search need that can be phrased in many ways and produce different results.
Read the full definitionBranded query
A query that includes a brand or product nameA branded query explicitly asks about a company, product or person. It usually measures recognition and accuracy, not the ability to be discovered among alternatives.
Read the full definitionNon-branded query
A query that expresses a need without naming a brandA non-branded query asks for options, solutions or comparisons without specifying which company should appear. It is the main surface for measuring discovery and competition in AI answers.
Read the full definitionRetrieval and RAG
How an AI system finds, filters, ranks and uses information before answering.
RAG
Retrieval-Augmented GenerationRAG is an architecture in which a system retrieves relevant documents or passages and adds them to the context before generating an answer. It enables the use of external, recent or private information.
Read the full definitionRetriever
The component that finds candidate documents or passagesA retriever searches an index, the web or a vector database for information related to the query. Its output feeds later reranking and generation stages.
Read the full definitionGenerator
The component that writes the final answerThe generator is the model that turns instructions and retrieved context into an answer. It may omit, summarize or combine sources even when they were retrieved correctly.
Read the full definitionWeb grounding
The use of web results to support an answerWeb grounding connects generation to pages retrieved in real time. It reduces exclusive reliance on model memory and may produce visible links or citations.
Read the full definitionGrounding
Anchoring an answer in available evidenceGrounding is the connection between generated claims and a source, datum or context that supports them. A well-grounded system can better verify what it says and where it came from.
Read the full definitionInformation retrieval
The process of finding information relevant to a queryRetrieval selects documents or passages before an answer is generated. It may use lexical matching, ranking signals, embeddings or a combination.
Read the full definitionReranking
Reordering candidates after initial retrievalReranking scores a reduced set of documents or passages again with a more precise model. A source may be retrieved yet still fall out before reaching the generator.
Read the full definitionCandidate source pool
The group of documents considered before generationThe candidate pool contains sources that passed an initial retrieval stage. If a page does not enter that group, the generator cannot select or cite it.
Read the full definitionQuery fan-out
Decomposition of one query into several subqueriesQuery fan-out is the process by which an engine expands a complex question into smaller searches. Sources retrieved for those subqueries are later combined into an answer.
Read the full definitionSynthetic query
A query generated by a system to retrieve informationA synthetic query is not written literally by the user: the engine creates it to reformulate, expand or split the need. It may use wording different from the original prompt.
Read the full definitionSemantic search
Retrieval based on meaning rather than only exact wordsSemantic search represents queries and documents by meaning to find conceptually related content. It often relies on embeddings and similarity measures.
Read the full definitionEmbedding
A numerical representation of content meaningAn embedding converts text, images or other content into a vector. Semantically similar items are placed close together and can be retrieved by similarity.
Read the full definitionVector database
A system that stores and searches vectors by similarityA vector database indexes embeddings to retrieve items close to a query. It is common in enterprise RAG systems and semantic search.
Read the full definitionChunking
Splitting documents into retrievable passagesChunking splits a long document into units that can be indexed and retrieved. A passage that is too short loses context; one that is too long mixes topics and reduces precision.
Read the full definitionTop-k
The number of highest-ranked results a system keepsTop-k indicates how many documents, passages or candidates move to the next stage. A small k raises the ranking bar; a large one adds coverage but also noise.
Read the full definitionContext window
The amount of information a model can process in one interactionThe context window limits the text, instructions, history and retrieved passages a model considers when answering. Available content may be truncated or compete for attention.
Read the full definitionTraining cutoff
Approximate time boundary of training dataThe training cutoff marks how far knowledge learned during training extends. It does not prevent a search-connected product from using later information.
Read the full definitionReal-time retrieval
Searching for information while generating the answerReal-time retrieval allows current sources to be searched while the answer is generated. Latency, bot access and the provider's index affect which pages appear.
Read the full definitionKnowledge graph
A network of structured entities and relationshipsA knowledge graph represents people, brands, products, places and their relationships. It helps resolve ambiguity and connect facts from multiple sources.
Read the full definitionEntity
An identifiable object such as a brand, person, product or placeAn entity is something a system can distinguish and connect to properties and other entities. Correct disambiguation prevents a brand from being confused with similar terms or companies.
Read the full definitionEntity salience
The relative importance of an entity within contentEntity salience expresses whether a brand or concept is central or merely incidental in a document. It is not the same as counting name occurrences.
Read the full definitionMulti-source RAG
RAG that combines evidence from multiple sourcesMulti-source RAG retrieves and synthesizes information from different documents, domains or collections. It can improve coverage but also introduces conflicts and attribution issues.
Read the full definitionPassage retrieval
Selecting specific passages rather than full documentsPassage retrieval looks for the most relevant block within a page. This is why a well-ranked URL may still fail to provide the passage a generator needs.
Read the full definitionMachine readability
How easily a system can interpret content and dataMachine readability depends on semantic HTML, stable structure, clear data and explicit relationships. It does not mean writing for robots or sacrificing the human experience.
Read the full definitionCitations and sources
Terms for evaluating attribution, grounding, source selection and hallucinations.
Citation
A visible reference to a source within an answerA citation links, identifies or attributes part of an answer to a source. It should be measured separately from a simple brand mention.
Read the full definitionSource attribution
Explicitly assigning a claim to its originAttribution indicates which source supports a specific claim. An answer may show links yet attribute a fact incorrectly or connect it to the wrong source.
Read the full definitionCitation grounding
The degree to which a citation truly supports generated textCitation grounding evaluates whether linked evidence contains what is needed to support a claim. It is not enough for the source and answer to discuss the same topic.
Read the full definitionCitation correctness
The share or quality of citations that correctly support a claimCitation correctness measures whether each reference is relevant and faithful to its content. It detects decorative, out-of-context or incorrectly matched citations.
Read the full definitionCitation completeness
The degree to which verifiable claims have supportCitation completeness checks how many claims needing evidence are accompanied by a source. An answer may have correct citations yet leave many facts unsupported.
Read the full definitionSource selection
The process of choosing which sources an engine uses or displaysSource selection happens after candidate retrieval and may depend on relevance, authority, freshness, diversity and usefulness for the answer.
Read the full definitionSource diversity
The variety of domains, formats and viewpoints usedSource diversity measures whether an answer depends on a few domains or combines independent viewpoints. More sources do not guarantee better evidence if they copy one another.
Read the full definitionSource freshness
The age or recency of a source selected by an engineSource freshness compares a source's publication or update date with query time. Its importance changes with intent and topic stability.
Read the full definitionCitation probability
An estimate of the likelihood that a source will be citedCitation probability summarizes signals that make a URL more likely to move from candidate to visible evidence. It is an estimate, not a guarantee for any query.
Read the full definitionCitability
How easily a source can be used and citedCitability combines clarity, extractability, evidence, attribution and query fit. Content can be good for people yet difficult to turn into a supportable passage.
Read the full definitionAI hallucination
Generated content that appears plausible but is unsupportedA hallucination occurs when a model invents or distorts facts, sources or relationships. For a brand it can affect prices, services, locations, people or reputation.
Read the full definitionSource concentration
The share of citations accumulated by a small group of domainsSource concentration shows whether a few sites dominate citations for a topic or engine. High concentration makes it harder for a new domain to enter.
Read the full definitionContent and authority
Signals that make a source interpretable, retrievable and useful in an answer.
Structured data
Data in an explicit format describing entities and propertiesStructured data makes products, organizations, authors, dates and relationships easier to interpret. On the web it is commonly implemented with Schema.org vocabulary.
Read the full definitionSchema markup
Implementation of Schema.org vocabulary on a pageSchema markup describes in a structured way what type of entity or content appears at a URL. JSON-LD is common, but markup does not replace visible content.
Read the full definitionDefinedTerm
Schema.org type for representing a defined termDefinedTerm declares the name, description and membership of a concept in a glossary or terminology set. It helps express the semantic role of a page like this one.
Read the full definitionllms.txt
A proposed file for pointing AI systems to useful contentllms.txt is a proposed convention for summarizing and linking useful content for models and agents. It is not a universal standard and does not guarantee indexing, retrieval or citations.
Read the full definitionInformation gain
New value a source adds beyond what is already availableInformation gain expresses how much useful, distinctive knowledge a page adds. Original data, methodology, examples and synthesis increase its potential value as a source.
Read the full definitionContent freshness
The material recency of published informationContent freshness depends on whether facts remain current, not merely on changing a date. It matters more for pricing, software, news and comparisons than for stable concepts.
Read the full definitionTopical authority
Perceived sustained expertise on a topicTopical authority comes from covering a field with depth, consistency, experience and clear connections across content. It is not earned by publishing many shallow pages.
Read the full definitionEarned media
Coverage or mentions earned on third-party channelsEarned media includes news, reviews, analysis, communities and editorial references the brand does not directly control. It can provide corroboration and access to domains engines already cite.
Read the full definitionE-E-A-T
Experience, expertise, authoritativeness and trustE-E-A-T is a Google-associated quality framework covering experience, expertise, authoritativeness and trust. In GEO it is a useful review lens for credentials and reliability, not a direct LLM score.
Read the full definitionDigital PR
Public relations focused on online coverage and authorityDigital PR creates assets, studies and stories that can earn editorial mentions and links. In GEO it expands the external sources that describe and validate an entity.
Read the full definitionContent extractability
How easily a useful passage can be isolated without losing meaningContent extractability indicates whether a system can isolate an answer, datum or list while preserving subject, context and conditions. It depends on writing and structure, not only technical format.
Read the full definitionDirect answer format
A structure that answers early and expands afterwardDirect answer format places a clear definition or conclusion first, then develops nuance, evidence and exceptions. It aids extraction without reducing content to one sentence.
Read the full definitionMultimodal GEO
Optimization for systems combining text, image, audio or videoMultimodal GEO studies how systems understand and rank content made of several modalities. Consistency among image, text, metadata and landing page becomes more important.
Read the full definitionVision-language model
VLMA vision-language model processes images and text together to describe, compare, search or rank content. Some GEO studies examine how these models rerank products and results.
Read the full definitionInternal linking
The network of links among pages on one siteInternal linking helps discover content, distribute context and express topical relationships. In a glossary it connects each term to the set and to neighboring concepts.
Read the full definitionCrawlability
A bot's ability to access and navigate a siteCrawlability depends on HTTP responses, robots rules, links, rendering and technical stability. An inaccessible page cannot enter many retrieval processes.
Read the full definitionVisibility measurement
Metrics for comparing brands, engines, prompts, sources and changes over time.
AI visibility
The measurable presence of an entity in AI answersAI visibility combines whether a brand appears, how often, in what position, with what sentiment and which sources accompany it. It should be defined before comparing tools.
Read the full definitionBrand mention
An explicit or recognizable appearance of a brandA brand mention occurs when an answer names or identifies the company, product or variant. It may exist without a link and may be positive, neutral or negative.
Read the full definitionRecommendation rate
The percentage of answers in which a brand is recommendedRecommendation rate counts answers where the brand appears as an advised option, not every mention. It requires classifying the brand's semantic role in the answer.
Read the full definitionMention role
The semantic function a brand serves in an answerMention role distinguishes recommendation, alternative, comparison, warning, example or incidental reference. It provides more information than a binary count.
Read the full definitionMention rate
The percentage of answers in which an entity appearsMention rate divides answers containing the brand by total valid runs. Technical errors should be excluded and results segmented by prompt and engine.
Read the full definitionCitation rate
The percentage of answers that cite a source or domainCitation rate measures how often a URL, domain or brand appears as a visible source. Its denominator may be answers, prompts or runs and must be stated.
Read the full definitionAI share of voice
A brand's share of presence versus competitorsAI share of voice compares a brand's presence with total mentions or visibility across a competitive set. The formula may weight position, engine or role.
Read the full definitionBrand sentiment
The tone or evaluation used by AI to describe a brandSentiment classifies how an entity is treated as positive, neutral, mixed or negative. Context and role matter more than isolated words.
Read the full definitionVisibility score
A composite index summarizing several presence signalsA visibility score combines metrics such as mentions, position, recommendation or engine coverage. There is no universal formula, so two platforms may produce different values.
Read the full definitionAverage position in AI
The average order in which a brand appears within answersAverage position summarizes where a brand appears in lists or recommendations. It requires rules for non-list answers, repeated mentions and absent brands.
Read the full definitionPrompt monitoring
Repeated execution of queries to measure changePrompt monitoring repeats a stable set of questions across defined engines and contexts. It reveals mentions, citations, sentiment and volatility.
Read the full definitionResponse volatility
Variation in results across equivalent runsResponse volatility reflects that the same prompt can produce different brands, positions or sources. It depends on sampling, the index, personalization and engine changes.
Read the full definitionOrganic search overlap
Overlap between AI-cited sources and organic search resultsOrganic overlap compares domains or URLs in generative answers with traditional results for the same query. Low overlap indicates different source-selection ecosystems.
Read the full definitionDiscovery gap
The gap between name recognition and discoveryDiscovery gap describes brands an AI recognizes when named but does not recommend in generic queries. It separates brand knowledge from competitive visibility.
Read the full definitionExposure bias
Unequal distribution of visibility among entities or sourcesExposure bias occurs when already prominent brands, creators or domains systematically receive more visibility. It can reinforce existing hierarchies.
Read the full definitionResearch metrics
Measures used in research to quantify ranking, stability, overlap and risk.
NDCG
Normalized Discounted Cumulative GainNDCG evaluates a ranking by considering relevance and giving more weight to higher positions. It is normalized so queries with different candidate sets can be compared.
Read the full definitionJaccard similarity
A measure of overlap between two setsJaccard divides shared elements by the union of two sets. In GEO it can compare cited domains, recommended brands or results across runs.
Read the full definitionKendall tau
Rank correlation between two ordered listsKendall tau measures how many pairs preserve the same order across two rankings. In GEO research it helps study the stability of positions or influence.
Read the full definitionPromotion success rate
How often an optimization moves an item to a target rankPromotion success rate measures whether a document or product reaches a threshold such as top 5, top 3 or first place after an intervention.
Read the full definitionWorst-case performance
The lowest result observed across queries or scenariosWorst-case performance evaluates the negative tail rather than only the average. In multi-query GEO it reveals whether an edit improves some questions but seriously harms others.
Read the full definitionDownside risk
The magnitude or frequency of losses versus a baselineDownside risk focuses evaluation on degradation rather than positive variation. It shows how much harmed queries worsen after a change.
Read the full definitionWin-tie rate
The share of cases that improve or do not worsenWin-tie rate counts queries where an intervention wins or ties against a baseline. It complements average improvement with a consistency measure.
Read the full definitionRisk and manipulation
Practices and attacks that can degrade quality or manipulate a generative answer.
Prompt injection
Malicious instructions inserted into content or inputsPrompt injection tries to alter model behavior through instructions the system should not follow. It may appear in retrieved pages, documents or user messages.
Read the full definitionContent poisoning
Publishing or altering information to contaminate answersContent poisoning introduces false or manipulated information into sources a model may retrieve or learn from. It aims to alter facts, recommendations or reputation.
Read the full definitionAdversarial GEO
Techniques designed to manipulate ranking or generationAdversarial GEO studies deceptive changes or attacks that exploit retrievers, rerankers or multimodal models to gain unearned exposure.
Read the full definitionKeyword stuffing
Artificial repetition of words to influence retrieval or rankingKeyword stuffing inserts terms excessively or unnaturally to manipulate relevance. It can reduce passage clarity, trust and usefulness.
Read the full definitionSource manipulation
An attempt to alter which sources a system retrieves, selects or citesSource manipulation tries to force selection or attribution through deceptive signals, artificial networks or content designed to exploit the pipeline.
Read the full definitionNo terms match that search.