Glossary
AI Search Glossary
Plain definitions for 75 essential terms across AI visibility, prompt engines, LLMs, retrieval, measurement, and AI search. Each entry gives the definition first, then what it means in practice, why it matters to a buyer, and how to act on it. Sources are linked where a claim is not ours.
Agentic Search
Agentic search uses an AI system that plans multiple steps, runs searches or tools, evaluates intermediate findings and continues until it can complete a task. Unlike a single retrieval pass, it may revise queries and compare several sources before producing an answer or taking an action.
AI Crawler
An AI crawler is a bot operated by an AI provider that fetches web pages either to train models or to answer live user queries. Each is identified by a user agent and can be allowed or blocked in robots.txt; OpenAI, for example, documents separate agents for training and for user-triggered browsing.
AI Search Advertising
AI search advertising is paid placement designed for AI-assisted search and conversational answer environments. Formats and availability vary by platform and market, so campaigns should distinguish currently purchasable inventory from emerging products being monitored or tested.
AI Search Visibility
AI search visibility is how often, and how favourably, AI engines mention, cite and recommend a brand across a defined set of buyer prompts. It is measured against a fixed prompt list run repeatedly across engines, so movement over time reflects your position rather than a change in the questions asked.
AI Share of Voice
AI share of voice is your brand's proportion of all brand mentions returned across a fixed prompt set. If ten prompts return sixty brand mentions and twelve of them are yours, your share of voice is twenty per cent. It is a relative measure, so it moves when competitors move too.
AI Visibility Baseline
An AI visibility baseline is the first documented measurement of a brand’s mentions, citations and recommendations across a fixed prompt set and named engines. It establishes the comparison point for later work and must preserve the collection method, date, model and sample size.
Answer Engine Optimization (AEO)
Answer engine optimization is the practice of structuring content so an answer engine can extract a single direct response and attribute it to your brand. It targets surfaces that resolve a question in place — AI Overviews, featured snippets, People Also Ask and assistant replies — rather than a ranked list of links.
Answer Sentiment
Answer sentiment is the positive, neutral or negative framing an AI response applies to a brand. Reliable analysis requires reading the surrounding claim rather than scoring the brand name alone, because a mention can appear in a warning, limitation or unfavorable comparison.
Answer Variance
Answer variance is the observable difference among responses to the same or closely related prompts, including changes in wording, sources, brands and ordering. It can be measured within one engine or compared across engines to show how stable a visibility result really is.
Attention Mechanism
An attention mechanism lets a neural network weigh which tokens in its available context are most relevant to one another. In transformers, self-attention helps the model represent relationships across a sequence, but it does not itself search the web or verify whether a claim is true.
Brand Mention
A brand mention is any appearance of your name in an AI answer, linked or unlinked. It is the weakest of the three visibility outcomes: it confirms the engine associates you with the topic, but it says nothing about whether the engine trusts your content or would suggest you to a buyer.
Canonical Tag
A canonical tag is an HTML link element that identifies the preferred URL for substantially duplicate pages. Search engines treat it as a strong signal rather than an absolute command, and may choose a different canonical when redirects, internal links, sitemaps and page content conflict.
Chunking
Chunking is splitting a document into smaller passages so a retrieval system can fetch and quote the part that answers a question rather than the whole page. Engines retrieve chunks, not sites, which means the unit that competes for a citation is a section of your page.
Citation (AI Search)
A citation in AI search is the attributed link an engine attaches to a claim inside its answer, crediting the source the statement came from. Citation is stronger than a passing mention because it is verifiable and clickable, and weaker than a recommendation because it credits a fact rather than endorsing you.
Citation Acquisition
Citation acquisition is the deliberate work of earning accurate brand coverage and references on independent sources that AI engines may retrieve or cite. It includes research, digital public relations, expert contributions and useful listings, not purchasing fabricated mentions or adding unsupported claims to low-quality sites.
Citation Frequency
Citation frequency is the proportion of measured AI answers that link to a brand or its content as a source. The denominator must be stated clearly, such as all prompt runs or only answers containing citations, because different definitions produce different percentages.
Citation Share
Citation share is a brand’s proportion of all observed AI citations within a defined competitor set, prompt set and time period. Unlike citation frequency, which asks how often one brand is cited, citation share describes how the available citation attention is divided among competitors.
Competitive Visibility Gap
A competitive visibility gap is the measurable difference between a brand and named competitors across AI mentions, citations, recommendations or prompt coverage. It is useful only when the same prompts, engines, sampling method and time window are applied to every company.
Content Freshness
Content freshness is how current a page’s substantive information is, including dates, prices, product facts, statistics and recommendations. Updating a date without reviewing the underlying claims does not make content fresh; engines and readers need evidence that the information remains accurate.
Context Window
A context window is the amount of text and other tokens a language model can consider in one interaction, including instructions, conversation history and retrieved documents. Material outside the active window cannot directly influence that response, even if it appeared earlier in a long session.
Conversational Search
Conversational search lets users ask natural-language questions, add follow-ups and refine constraints across an interaction. The meaning of a later question can depend on earlier turns, so visibility may change as the user narrows industry, budget, geography or technical requirements.
Crawlability
Crawlability is a search or AI crawler’s ability to request a page and reach its meaningful content and links. Server errors, blocked robots rules, inaccessible navigation and script-dependent rendering can reduce crawlability even when human visitors can open the page normally.
E-E-A-T
E-E-A-T stands for experience, expertise, authoritativeness and trust. Google describes it in its search quality rater guidelines as a way of assessing content quality, with trust the most important element. It is not a direct ranking score, but it names the signals that credible content tends to carry.
Embedding
An embedding is a numeric vector that represents the meaning of a piece of text, so systems can compare passages by similarity rather than exact keyword match. Retrieval layers in AI search embed both the question and candidate content, then select passages whose vectors sit closest to the query.
Entity
An entity is a distinct thing that search and AI systems recognise and store attributes about: a company, product, person, place or concept. Engines answer brand questions from entity understanding, so if your attributes are inconsistent across the web the engine's picture of you is inconsistent too.
Entity Optimization
Entity optimization is the work of making a company, product or person consistently understandable as a distinct entity across owned pages, structured data and credible third-party sources. It clarifies attributes and relationships; it does not guarantee a knowledge panel, ranking or AI recommendation.
Featured Snippet
A featured snippet is the block at the top of Google's results that quotes an extract from a page in order to answer the query directly, with a link to the source. Google says these are selected programmatically from pages that appear in the results, not submitted by site owners.
Fine-Tuning
Fine-tuning is additional training that adapts a pretrained model to examples for a narrower task, style or behavior. It is different from prompting, retrieval and ordinary website optimization; publishing content does not fine-tune a public AI model.
Generative Engine Optimization (GEO)
Generative engine optimization is the practice of shaping content, sources and entity data so generative AI engines retrieve it, quote it and attribute it when answering a user. The term was introduced in a 2024 research paper that tested which content changes raise a source's visibility inside generated answers.
Google AI Overviews
AI Overviews are Google's AI-generated summaries that appear above the traditional results for some queries, with links to the pages the summary drew on. Google describes them as using a customised version of Gemini to surface information alongside the links it cites.
Grounding
Grounding is tying a model's output to verifiable external sources so each claim in the answer can be attributed and checked. A grounded answer names where its statements came from; an ungrounded one is produced from the model's own parameters and cannot be traced back to any document.
Hallucination
A hallucination is a confident but false statement produced by a language model. For brands this shows up as invented pricing, discontinued products, wrong leadership names or misattributed positioning — errors that read as authoritative and are repeated to buyers who have no way to check them.
Hybrid Search
Hybrid search combines lexical matching, which looks for words, with vector search, which compares semantic meaning. Many retrieval systems blend both approaches so exact names and technical terms remain findable while conceptually similar wording can also match.
Indexability
Indexability is a page’s eligibility to be stored in a search engine’s index after it has been crawled. A page may be crawlable but excluded because of a noindex directive, canonical choice, duplicate content, low-value signals or other indexing decisions.
Information Gain
Information gain is the useful new evidence a page adds beyond what other available sources already say. In AI search, this can include original data, first-hand examples, clear comparisons or a better-supported explanation, though no public universal score determines whether a page will be cited.
JavaScript Rendering
JavaScript rendering is the browser-like processing required to turn script-driven page code into visible content and links. Some crawlers render JavaScript fully, some partially and some not at all, so critical information delivered only after client-side execution can be harder or slower to discover.
Knowledge Graph
A knowledge graph is a structured network of entities and the relationships between them — company to founder, product to category, brand to competitor. Search systems use graphs to answer questions about things rather than text strings, and AI answers inherit those relationships when describing a brand.
Large Language Model (LLM)
A large language model is a neural network trained on very large text corpora to predict language, and it is what produces the prose in ChatGPT, Gemini, Perplexity and Copilot. Left alone it answers from training data; connected to search or a document store it answers from retrieved sources instead.
Lexical Search
Lexical search retrieves documents by matching words or word forms in the query with words in the index. Modern systems may add stemming, synonyms and other language processing, but exact terms, names and identifiers still matter in ways that purely semantic retrieval can miss.
LLM Inference
Inference is the process of running a trained language model to produce an output from a prompt and its available context. During inference, model parameters are generally not retrained; the response is generated from existing parameters plus any instructions, conversation history and retrieved evidence.
LLM SEO
LLM SEO is the practice of making a brand retrievable, quotable and recommendable inside large language model answers. LLM SEO, LLMO and generative engine optimization are different names for the same work; the vocabulary varies by vendor, the underlying activity — content, entity and citation engineering — does not.
llms.txt
llms.txt is a community proposal for a markdown file at a site's root that offers language models a curated, readable summary of the site's most important content. It is a proposal published at llmstxt.org, not a standard adopted by search or AI providers, and support should not be assumed.
Mention Rate
Mention rate is the percentage of measured AI responses in which a brand name appears, regardless of whether the answer links to or recommends it. It is a useful awareness baseline but a weaker commercial signal than citation frequency or recommendation rate.
Model Drift
Model drift is a change in an AI model’s behavior or outputs over time, caused by model updates, retrieval changes, safety tuning or other system changes. In visibility measurement, drift can alter answers even when a brand and its content have not changed.
Model Parameters
Model parameters are the learned numerical values inside a language model that shape how it predicts tokens. They are adjusted during training or fine-tuning, not by a normal web search or page visit, and should not be confused with the external documents supplied during retrieval.
Multimodal Model
A multimodal model can process or generate more than one kind of information, such as text, images, audio or video. In AI search, multimodal input can let users ask questions about screenshots, products or documents, while retrieval may include visual as well as textual evidence.
Passage Retrieval
Passage retrieval selects a relevant section of a document rather than treating the entire page as one unit. AI search systems can use the selected passage as context for an answer, making section-level clarity and self-contained claims important even on long pages.
Prompt Clustering
Prompt clustering groups buyer prompts by shared meaning, intent or expected answer rather than exact wording. It reduces duplicate measurement and reveals the themes where a brand is consistently visible or absent across many variations of the same underlying question.
Prompt Engine
A prompt engine is a conversational system that interprets a natural-language request and generates an answer, often using a large language model plus search, retrieval and other tools. The term describes products such as ChatGPT, Gemini, Perplexity and Copilot rather than one specific technical architecture.
Prompt Engineering
Prompt engineering is the practice of designing instructions and context that guide a language model toward a useful response. In AI search measurement, prompts should mirror real buyer questions rather than manipulate an answer, because the goal is to observe how an engine represents the market.
Prompt Expansion
Prompt expansion adds related terms, entities or constraints to a user’s prompt before retrieval. It helps an engine find relevant evidence when the original wording is short or ambiguous, but it can also steer results toward a narrower interpretation than the user stated.
Prompt Intent
Prompt intent is the underlying task a person wants an AI engine to complete, such as learning, comparing, troubleshooting or choosing a vendor. Two prompts can use similar words but have different intent, so they may retrieve different sources and produce different brand recommendations.
Prompt Set
A prompt set is the fixed, documented list of buyer questions run repeatedly across AI engines to measure visibility. Because AI answers vary between runs, the prompt list must stay constant for results to be comparable: change the questions and you have started a new baseline rather than continued the old one.
Prompt Volatility
Prompt volatility is the degree to which an AI engine’s response changes when the same prompt is repeated over time or across sessions. It reflects probabilistic generation, changing retrieval results and model updates, so one answer should never be treated as a stable measurement.
Query Fan-Out
Query fan-out is when an AI search system breaks one complex question into several related searches, gathers evidence for each part and combines the findings into one response. A brand may qualify for one subquery but remain absent from the final answer if other required evidence is missing.
Query Rewriting
Query rewriting is the process of translating a user’s natural-language prompt into one or more search queries that a retrieval system can execute. The rewritten wording may emphasize entities, constraints or synonyms that never appeared verbatim in the original request.
Recommendation (AI Search)
A recommendation is when an AI engine actively names your brand as an option a buyer should consider, rather than merely quoting you as a source. Mention, citation and recommendation form a ladder: mention is presence, citation is evidence, recommendation is endorsement, and only the last one shapes a shortlist.
Recommendation Rate
Recommendation rate is the percentage of measured AI responses that actively suggest a brand as an option for the user’s stated need. It should be calculated from a fixed prompt set and distinguished from mentions or citations, which do not necessarily indicate endorsement.
Reranking
Reranking is a second-stage retrieval step that reorders an initial set of candidate documents or passages using a more detailed relevance model. It aims to move the best evidence closer to the top before the language model receives context or the search interface displays results.
Response Sampling
Response sampling is the practice of collecting multiple outputs for the same prompt and conditions to estimate how often a result occurs. Because generated answers are probabilistic, repeated samples provide a more defensible measure than a single screenshot or manually selected response.
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation is an architecture that fetches relevant documents at query time and passes them to a language model, which then answers from that material. It was introduced in a 2020 research paper and is the mechanism behind AI answers that carry live citations to specific web pages.
robots.txt
robots.txt is a site-root file that tells compliant crawlers which URL paths they may request. It controls crawling rather than indexing by itself, and different search and AI providers publish distinct user agents that site owners can allow or disallow.
Search Grounding
Search grounding is a form of grounding in which an AI system retrieves search results and uses them to support its response. The term emphasizes the retrieval source: public search indexes rather than a private document collection, uploaded file or internal database.
Source Inclusion Rate
Source inclusion rate is how often a domain or page appears in the set of sources retrieved or cited for a defined group of AI answers. It measures evidence visibility rather than brand endorsement and should identify the engine, prompt set and collection period.
Source Selection
Source selection is the process an AI search system uses to choose which retrieved documents or passages will support a response. Relevance, authority, freshness, diversity and product-specific rules may influence selection, but exact weighting is rarely public and should not be presented as a known formula.
Structured Data (Schema Markup)
Structured data is standardised machine-readable markup — usually schema.org vocabulary in JSON-LD — that tells machines what a page's visible content means. Google's documentation is explicit that structured data must describe content that is actually visible to users on the page.
System Prompt
A system prompt is a high-priority instruction supplied by an AI application to define the assistant’s role, boundaries and behavior. Users typically do not control it, and it can affect tone, tool use, source handling and which requests the application will answer.
Token
A token is a unit of text processed by a language model, often a word fragment, punctuation mark or short word. Models read and generate sequences of tokens, and context limits and usage costs are commonly expressed as token counts rather than word counts.
Topical Authority
Topical authority is the perceived depth and reliability of a source across a subject, built through useful coverage, expertise, internal structure and external corroboration. It is not a publicly documented numeric Google score, so it should describe an observed content strategy rather than a guaranteed ranking factor.
Training Data Cutoff
A training data cutoff is the latest period represented in a model’s pretraining data, though providers may not disclose one precise date and retrieved information can be newer. A cutoff therefore does not tell you whether a specific AI answer used live search.
Transformer Model
A transformer is a neural-network architecture that processes relationships among tokens using attention mechanisms. Introduced in 2017, it underpins modern large language models and allows models to weigh relevant parts of a sequence when producing or interpreting text.
Vector Database
A vector database stores embeddings and retrieves items whose vectors are similar to a query vector. It is commonly used in retrieval-augmented generation systems to find relevant passages, although public AI search products may combine vector stores with conventional search indexes and other systems.
Vector Search
Vector search retrieves content by comparing embeddings, matching passages on meaning rather than exact wording. It lets an engine surface a relevant passage even when the buyer's phrasing shares no keywords with the page, which is why AI answers often quote content that never targeted the query as written.
Web Grounding
Web grounding is the use of live or recently indexed web results as evidence for an AI-generated response. It can make answers more current and citeable than model memory alone, although the quality still depends on query formulation, source selection and faithful synthesis.
Zero-Click Search
A zero-click search is a query resolved on the results page or inside an AI answer, so the user never visits a website. The information still comes from published sources, which means the commercial question shifts from how much traffic a page earns to how often it is the source being read.
Ranking is no longer enough
You need to be cited, mentioned, and recommended.
Being cited, mentioned, and recommended are three different outcomes, and most brands only ever achieve the first one. Ranking is no longer enough because AI engines answer buyers directly and name only a short list of vendors as the recommendation — everyone else is cited in passing, if at all. Get a free AI Visibility Report to see exactly where your brand appears today across ChatGPT, Google AI Overviews, Gemini, Perplexity, and Copilot, where competitors are winning the recommendation instead, and what's keeping you from moving up the shortlist.