Last updated October 2026.
Ask five LLM visibility vendors to define share of voice and you get five formulas back. One measures ad impressions. Another measures AI-generated answers only. A few just describe whatever their dashboard already calculates, then reach for a definition that fits it.
Most GEO and AI-visibility vendors sell tracking software without ever publishing a glossary for the words on their own dashboards. Fewer still frame those words around tracking methodology instead of a sales pitch.
This glossary fills that gap: 45 terms, alphabetized, each one answered in its first sentence. A handful are one vendor’s own coinage, and each of those says so. None of the notes below are a formal per-term score. They flag where the industry uses one word to mean several different things.
The 45 terms, A to Z
Agentic Search
Agentic search is a search flow where an AI system runs several steps on its own: reformulating a query, checking sources, and sometimes browsing before it answers. It complicates “being cited,” since the system can cite a page it visited mid-task rather than one from a static index.
AI Overviews
AI Overviews is Google’s generative answer block that sits above traditional results for many search queries. It draws on Google’s own index and current web content, and it is one of the five engines this site’s benchmark covers, alongside ChatGPT, Perplexity, Gemini, and Microsoft Copilot.
Answer Engine Optimization (AEO)
Answer Engine Optimization means shaping content so an AI system can lift a clean, direct answer out of it: a definition, a step list, or a specific number. Vendors often use AEO and GEO interchangeably, but AEO is narrower. It targets the extracted answer, not the whole response around it.
API Access
API access is a documented, self-serve way to pull a tool’s tracking data into another system, such as a warehouse or a client dashboard. A pricing page listing “API access” as a bullet point does not guarantee a public price or a self-serve signup flow exists.
Attribution
Attribution is the process of tracing a mention or a citation in an AI answer back to a specific brand, page, or domain. It sounds simple until an answer paraphrases a brand without linking to it, which is exactly the case an unlinked mention is built to catch.
Bot Traffic
Bot traffic is the flow of requests AI crawlers send to a website while gathering material for a model to draw on later. It shows up in server logs, not in an AI chat window, so tracking it needs log analytics, not a prompt-panel tool.
Brand Mention
A brand mention is any appearance of a company’s name inside an AI-generated answer, whether or not the answer links to the company’s site. It is the broadest of the counting terms here. Every citation is a mention, but not every mention is a citation.
Citation
A citation is a brand mention that comes with a visible source link or a named source domain inside the AI answer. It is a narrower, more valuable count than a mention, since it shows the model treated a specific page as evidence.
Citation Map
A citation map traces which source URLs fed which parts of an AI answer, page by page rather than domain by domain. Profound builds its reporting around this level of detail, which is more granular than a source-domain list alone.
Co-citation
Co-citation happens when an AI answer names a brand next to a competitor without linking either one, often in a comparison sentence like “brands such as X and Y.” It signals association, not endorsement, and most tools count it separately from a sourced citation.
Competitive Gap Analysis
Competitive gap analysis compares a brand’s mention or citation rate against named competitors across the same prompt panel, then flags where a rival shows up and the brand does not. It turns one blended score into a ranked list of specific prompts worth fixing first.
Crawler
A crawler is the automated program an AI company runs to fetch and read pages before a model can cite or learn from them. Each engine runs its own crawler under its own rules, and blocking one in robots.txt has no effect on the others.
Direct Answer
A direct answer is the specific fact, number, or instruction an AI system pulls out and states plainly, as opposed to a longer synthesized summary with hedged wording. It is the unit AEO work targets: one clean sentence a model can lift without editing it first.
Discoverability
Discoverability is how easily an AI crawler or a retrieval step can find and parse a page, before content quality ever enters the picture. A page can be well written and still invisible to an engine if it sits behind a login wall or a JavaScript-only render.
Engine
An engine, in this glossary, is one of the AI products a tracking tool measures: ChatGPT, Perplexity, Google AI Overviews, Gemini, or Microsoft Copilot. Count only the ones a tool actually tracks on your plan tier, not the ones listed on its homepage.
Entity
An entity is a specific, named thing (a company, a person, a product) that a model can recognize and connect to facts it already holds. Consistent naming, a stable domain, and structured data all help a model attach the right facts to the right brand.
Generative Engine Optimization (GEO)
Generative Engine Optimization is the practice of shaping a brand’s content and off-site presence so it gets surfaced, cited, or recommended inside AI-generated answers. It is the umbrella term this whole glossary sits under, and it covers AEO, technical crawlability, and reputation work together.
Grounding
Grounding is when a model backs its answer with retrieved, current information instead of generating purely from training. A grounded answer can cite a page from this morning. An ungrounded one can only repeat what the model already knew, which may be stale or wrong.
Hallucination
A hallucination is a confident, wrong statement a model generates because nothing it retrieved or learned actually supports the claim. In a visibility context, a hallucinated brand mention, naming a company that makes no such product, counts as noise, not as a citation worth chasing.
Knowledge Graph
A knowledge graph is a structured map of entities and the facts connecting them, used by Google and other engines to answer who a company is and what it does. A brand with a clean, consistent knowledge graph presence tends to get named more accurately in generated answers.
Large Language Model (LLM)
A large language model is the underlying system, trained on huge amounts of text, that generates the answers a chatbot or a search feature shows you. ChatGPT, Gemini, and Copilot are products built on top of an LLM. The model itself is the engine doing the writing.
llms.txt
llms.txt is a proposed plain-text file, placed at a site’s root, listing pages a site owner wants an AI system to read first. Adoption is uneven, and no major engine confirms it changes what gets cited, so treat it as a courtesy signal, not a ranking lever.
Margin of Error
Margin of error is how far a single measurement could sit from a brand’s true visibility rate, shrinking as sample size grows. Evertune’s own published methodology shows the swing: one prompt asked once carries roughly a plus or minus 44-point margin; asked 100 times, that narrows to about plus or minus 10 points.
Mention Rate
Mention rate is the share of prompt-panel runs, out of the total run, where a brand’s name shows up anywhere in the answer. It is the loosest of the rate metrics here, since it counts a passing, unlinked mention the same as a fully sourced citation.
Model Coverage
Model coverage is the list of AI engines a tool actually measures on a given plan, not the list on its marketing page. Two tools can both claim broad coverage while one tracks five engines on every tier and the other unlocks most of that list only at its top price.
Non-determinism
Non-determinism is the property that makes an AI model return a different answer to the identical prompt from one run to the next. It is the reason a single ChatGPT reply is an anecdote, and the reason prompt sampling exists at all.
Perception Tracking
Perception tracking measures how favorably or unfavorably an AI answer describes a brand, separate from whether the brand gets mentioned at all. A brand can show up in every answer and still lose on perception if the surrounding language leans negative or outdated.
Position
Position is where in an AI answer a brand’s mention or citation appears: first sentence versus a buried list item. Some tools weight an early position higher than a late one, on the theory that readers see less of a long answer as it goes on.
Prompt
A prompt is the exact question or instruction typed into an AI engine to generate a test answer. Wording changes the answer, so a tracking tool that quietly rewrites its prompts between runs breaks the comparison it is supposed to provide.
Prompt Panel
A prompt panel is the fixed set of prompts a tool runs on a schedule to measure a brand’s visibility over time. It has to stay stable to mean anything. Swap the prompts mid-quarter and a chart’s “improvement” may just be a different question, not a better answer.
Prompt Sampling
Prompt sampling is running the same prompt multiple times inside one measurement cycle, then reporting a mention rate across those runs instead of a single reply. Evertune publishes the clearest run-count claim in this category: up to 100 samples per prompt in a single cycle, per its own methodology.
Prompt Volumes
Prompt Volumes is Profound’s name for a demand-side dataset showing what real people actually type into AI assistants, separate from the answers those prompts generate. It helps a team build a prompt panel around real search behavior instead of guessing at plausible questions.
Re-baseline
To re-baseline is to reset a tracking program’s starting point after a major model update, since old scores are not comparable to new ones once the underlying model has changed. Skipping this step makes a routine model swap look like a visibility crash on a chart.
Refresh Cadence
Refresh cadence is how often a tool reruns its full prompt panel, daily, weekly, or on some slower cycle. It is a different claim from prompt sampling, since a tool can refresh daily while still asking each prompt only once per day inside that cycle.
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation is the technique behind grounding: a model searches a live index for relevant text, then writes its answer from what it finds instead of relying only on training data. Most AI Overviews and chat answers that cite fresh sources run on some form of RAG.
Sentiment Analysis
Sentiment analysis scores whether the language around a brand mention reads positive, negative, or neutral. It sits behind perception tracking, and vendors disagree on the scale. Some report a simple three-way label; others, a numeric score with no published range to compare it against.
Share of Model
Share of model is the percentage of AI-generated answers, across a defined prompt panel and engine set, that mention a specific brand instead of its named competitors. It answers a narrower question than overall visibility: not whether a brand ever gets mentioned, but how often, relative to the field.
Share of Voice
Share of voice is an older term borrowed from traditional media and paid search, measuring a brand’s presence against competitors across a channel. Some vendors use it interchangeably with share of model, which blurs a real distinction: ad impressions and backlinks versus generated AI answers.
Source Domain
A source domain is the website an AI answer credits when it cites a fact, listed at the domain level rather than the exact page. It is coarser than a citation map, and it is the easiest citation data point for most tools to report reliably.
Structured Data
Structured data is markup, such as schema.org tags, added to a page’s code to describe its content in a format a machine can parse directly. It does not guarantee a citation, but it removes ambiguity a model would otherwise have to infer from plain text.
Synthetic Prompt
A synthetic prompt is a test question a vendor writes, not one pulled from real search or chat data. Ahrefs Brand Radar takes the opposite approach for part of its panel: a database of 471 million-plus prompts it calls search-backed, tied to real Google query volume instead of invented from scratch.
Topic Cluster
A topic cluster is a group of prompts in a panel that share a theme, grouped so a report can show a rate for pricing questions or comparison questions separately instead of one blended average. Clustering also narrows the margin of error within each group.
Unlinked Mention
An unlinked mention is a brand mention with no source link or named domain attached, the AI equivalent of an unlinked brand mention in a news article. It still signals awareness, but it gives a marketing team nothing to trace back to a specific page.
Visibility Score
A visibility score is a single blended number summarizing a brand’s overall AI presence, usually built from mention rate, citation rate, and sometimes position. Because every vendor builds its own formula, the same brand can show a different score on two different tools in the same week.
Zero-click Answer
A zero-click answer is a generated response complete enough that the person reading it never visits a source page. It is the tension behind every metric in this glossary: a brand can win the citation and still get none of the traffic that citation used to send.
Use these definitions before you buy
A glossary only earns its keep if it changes how you read a pricing page. Before you trust a vendor’s chart, ask which definition of share of model, mention rate, or visibility score sits behind it, and whether that definition matches the one you would use.
See how these terms turn into an actual A to F grade on the LLM visibility benchmark, and read the measurement protocol behind every score on it.