Glossary | OppAlerts

Every term this discipline uses, one clean definition each, in plain language. Where the industry uses a term loosely, the entry gives the precise definition and notes how usage varies. Each entry links the chapter that treats the concept in depth. The book ends here on purpose: the last thing a reference guide owes its reader is a shared vocabulary.

A to Z

AEO (answer engine optimization): Optimizing to be the direct answer in systems that respond with an answer rather than a list of links. The framing predates LLMs (featured snippets and voice assistants were already answer engines) and emphasizes citation. One of several overlapping names for parts of AI search optimization; the full name debate is settled in The Name Mess: SEO, GEO, AEO, LLMO, and What This Book Calls Things.
Agent (agentic AI): An AI system that takes actions on the user's behalf, such as browsing pages, filling forms, or completing purchases, rather than only writing an answer. Commercial implications are covered in Commercial Visibility: Products, Comparisons, and Agentic Buying and The Road Ahead: Agents and What Comes Next.
AI Mode: Google's conversational search experience: a chat-style interface inside Google Search that answers with generated responses backed by Google's index. Distinct from AI Overviews, which sit on top of a standard results page. See The AI Search Ecosystem.
AI Overview: The generated summary Google places above organic results on some queries, with citations to source pages. Because it is built on Google's index and ranking, it behaves more like classic search than chatbots do. See The AI Search Ecosystem.
AI search: Any product experience where an AI-generated answer, rather than a list of links, is the primary response to a query: chat assistants, AI Overviews and AI Mode, and answer engines like Perplexity. The shift it causes is the subject of From Ten Blue Links to One Answer.
AI search optimization: This book's umbrella term: the work of improving how AI search systems represent and recommend a brand. It encompasses classic SEO, GEO, AEO, LLMO, and paid placement. Optimization is what you do; visibility is what you get, and the two terms are never interchangeable in this guide.
AI search visibility: The outcome AI search optimization produces and measurement tracks: how often, how prominently, and how favorably AI answers mention, recommend, and cite a brand. Measured as a distribution over repeated samples, per Measuring AI Search Visibility.
CCBot: Common Crawl's crawler. Common Crawl publishes a free web archive that many LLM training pipelines draw from, so CCBot access affects training-data presence across multiple providers at once. See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
ChatGPT-User: The OpenAI user agent that fetches a page when a ChatGPT conversation needs it live. It acts on user requests, and OpenAI documents it separately from GPTBot (training) and OAI-SearchBot (search). See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Chunk: The passage-sized unit a page gets split into for retrieval. Retrieval systems fetch and rank chunks, not whole pages, so a page's structure decides whether its facts survive extraction intact. The design consequences are the subject of Content Architecture: Building for the Chunk.
Citation: A source an AI answer names or links as backing. Citation rate is the share of sampled answers in which your site is credited. What models credit differs sharply from what they read; the evidence is in What Actually Correlates with AI Search Visibility.
ClaudeBot, Claude-SearchBot, Claude-User: Anthropic's three crawlers, for training data, search indexing, and live user-requested fetches respectively, each controllable separately in robots.txt. See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Context window: The bounded amount of text a model can consider at once: the prompt, any retrieved pages, and the answer so far. Retrieved content competes for this limited space, which is why extraction-friendly writing matters.
Crawler: A program that fetches web pages automatically, identified by a user-agent string. In AI search, different crawlers feed different systems (training corpora, search indexes, live answers), and the distinctions carry real strategic weight. Full directory in Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Embedding: A list of numbers representing a piece of text so that similar meanings sit near each other, which lets retrieval match by meaning rather than exact words. Explained properly in Embeddings: How Machines Represent Meaning.
GEO (generative engine optimization): Optimizing content for visibility in generated answers. Coined by a 2023 academic paper; in practice now used for the whole discipline. Its collision with "geo" as geography is one reason this book prefers AI search optimization. See The Name Mess: SEO, GEO, AEO, LLMO, and What This Book Calls Things.
Google-Extended: Google's robots.txt product token for controlling whether your content trains Gemini models. It is a training control, separate from Googlebot: blocking it does not affect Search rankings or AI Overviews. See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
GPTBot: OpenAI's training-data crawler. Disallowing it signals that your content should be excluded from future OpenAI training; it does not affect ChatGPT search results, which run through OAI-SearchBot. The common confusion here is debunked in Myths and Misconceptions.
Grounding: Giving a model retrieved documents to write from, so the answer rests on fetched text rather than memory alone. A grounded answer can cite its sources; an ungrounded one cannot. Mechanics in Retrieval and Grounding: How an AI Answer Gets Assembled.
Hallucination: A model stating something false with the same fluency as something true, because generation produces plausible text, not verified facts. Brand-accuracy monitoring exists because hallucinations about your company are a visibility problem too.
Knowledge cutoff: The date a model's training data ends. Everything after it is invisible to the model's memory and reachable only through retrieval. Providers publish cutoffs in their model documentation; the consequences run through Freshness, Cutoffs, and Temporal Accuracy.
Knowledge graph: A structured database of entities and their relationships, such as Google's Knowledge Graph or Wikidata. Entity presence work targets these; what it is and is not worth is covered in Entities: Becoming a Thing the Machine Knows.
LLM (large language model): A model trained on large text corpora to predict the next token, which at scale produces fluent answers to arbitrary prompts. The one term this book assumes; the mechanics are still worth reading in How an LLM Turns a Prompt into a Response.
LLMO (large language model optimization): Optimizing how LLMs represent a brand, with emphasis on the model and its training rather than the search layer. Another partial synonym in the acronym stack; see The Name Mess: SEO, GEO, AEO, LLMO, and What This Book Calls Things.
Mention rate: The share of sampled answers in which a brand appears at all, in any role. The broadest of the metric family defined in Measuring AI Search Visibility; distinct from recommendation rate, which requires the answer to actually endorse you.
OAI-SearchBot: OpenAI's search crawler, which powers appearance in ChatGPT search results. Documented as independent of GPTBot: a site can allow search while opting out of training. See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Parametric memory (model memory): What a model knows from training, stored in its weights: statistical associations rather than a lookup database. Frozen at the knowledge cutoff, and biased toward brands with years of accumulated coverage. The half of the book's core split explained in Model Memory and Live Retrieval.
Persona: In measurement, a buyer framing prefixed to a prompt ("as a senior on a fixed budget..."). Answers shift by persona, and in taste-driven categories the aggregate leaderboard is an average of several different competitions. Usage in Measuring AI Search Visibility and Competitive Analysis.
PerplexityBot, Perplexity-User: Perplexity's crawlers: PerplexityBot builds its search index, and Perplexity-User fetches pages during user sessions. See Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Prominence: Where a brand appears when it appears: first pick, list position, or trailing aside. A metric, not a ranking; it only means something averaged over repeated runs.
Prompt: The text a user sends to an AI system. The unit of research in this discipline, the way the keyword was the unit in SEO: inventoried, classified by intent, and sampled repeatedly. See Prompt and Topic Research.
Query fan-out: The step where an AI search system turns one prompt into several search queries, runs them, and pools the results before answering. Your content competes for retrieval on the fanned-out queries, not just the literal prompt. See Retrieval and Grounding: How an AI Answer Gets Assembled.
RAG (retrieval-augmented generation): The architecture behind grounded answers: fetch relevant documents, put them in the model's context, then generate. Named by a 2020 research paper and now the default pattern in AI search products. Walked through in Retrieval and Grounding: How an AI Answer Gets Assembled.
Ranking factors: The signals that correlate with visibility in a system's outputs. Some argue the term should not apply to AI search because there is no ranked index; this book uses it deliberately, because the practical question, which signals predict appearing in answers, is the same question it always named. The evidence is in What Actually Correlates with AI Search Visibility.
Recommendation rate: The share of sampled answers in which the answer actually recommends the brand, rather than mentioning it in passing or unfavorably. The metric most directly tied to commercial outcomes.
Retrieval (live retrieval): The search step: the product turns the prompt into web searches, runs them on an engine like Google, Bing, or Brave, and puts the fetched results into the model's context before it writes. The faster-moving half of the split in Model Memory and Live Retrieval.
robots.txt: The file at a site's root that tells crawlers what they may fetch, standardized as the Robots Exclusion Protocol. Compliance is voluntary but the major AI crawlers document honoring it. Directives for each AI bot are in Technical Accessibility for AI Crawlers.
Share of voice: A brand's mentions as a fraction of all brand mentions across sampled answers: the competitive version of mention rate, and the closest thing this channel has to a market-share number.
Temperature: The sampling setting that adds deliberate randomness to each token a model generates. It is why the same prompt produces different answers on different runs, and why single-query measurements mean almost nothing.
Token: The unit a model reads and writes: a word fragment, roughly four characters of English on average. Models predict one token at a time, and context windows and API pricing are both denominated in tokens.
Training data: The text corpora a model learns from: crawled web pages, licensed datasets, and other collected text. What lands in training data becomes model memory at the next training run; the pipelines are mapped in Where the Machines Get Their Data: Crawlers, Indexes, and Training Pipelines.
Volatility: How much answers vary across runs, prompts, models, and time. High volatility means larger samples before any number is trustworthy, and a sudden volatility rise on prompts you own is an early warning that the slot is loosening.