Skip to content
FixAEO
All posts
AEOGEOGlossaryAI SearchEducation

The AI search glossary: 120+ AEO and GEO terms, defined

A skimmable, plain-English glossary of 120+ AI search, AEO and GEO terms — from citation rate and query fan-out to GPTBot, RAG, BM25, grounding metadata and llms.txt, with primary sources.

Nitish Kumar YadavBy Nitish Kumar Yadav··Updated ·27 min read
On this page

You're on a call and a colleague says your brand "isn't showing up in GEO." You nod along. Then you realize you're not sure whether they mean AI Overviews, ChatGPT, or all of it at once — or whether GEO and AEO are even different things. I have that exact moment at least once a week, and I've been building in this space for a while.

The field is young and the words are a mess. There's no standards body deciding that AEO means one thing and GEO means another; vendors use them interchangeably and then disagree about it. So this is the glossary I wish someone had handed me: 120+ terms that actually come up, grouped by what they're for, each defined in a sentence or two of plain English. Where a term is genuinely contested or technical, I've linked the primary source so you can check me. Where we have a deeper post on it, I've linked that too.

I keep this list current — I last went through it on 11 November 2026, when I roughly doubled the term count. The plan is to keep adding as the vocabulary settles, so bookmark it rather than copy it. Use Ctrl+F; the terms are deliberately grouped, not alphabetical, because the first thing most people need is to see which bucket a word belongs to.

The core ideas

  • Answer engine optimization (AEO). Getting your brand named and cited inside an AI's written answer, not just ranked in a list of blue links. See what AEO actually is.
  • Generative engine optimization (GEO). A broader, usually interchangeable label for the same work — showing up across generative AI surfaces. There's no authority that settles AEO vs GEO, so most people treat them as synonyms. Here's the clean breakdown.
  • LLMO (large language model optimization). A newer alias for the same practice. Same work, different three letters. More on LLMO.
  • AI visibility. How often, and how prominently, AI engines surface your brand when people ask about your category. Why it matters.
  • Answer engine. Any system that returns a direct, synthesized answer instead of a page of links — ChatGPT, Perplexity, AI Overviews.
  • Generative engine. An answer engine whose reply is generated by a large language model rather than assembled from a fixed template.
  • Conversational search. Searching by asking in natural language and following up, the way you'd talk to a person. More on conversational search engines.
  • Zero-click search. A search answered on the results surface itself, so the user never clicks through to a site. What zero-click means for GEO.
  • Citation. When an AI answer links to or names your page as a source. This is the outcome AEO is really after.
  • Mention. When an answer names your brand in the text without linking to you. A mention builds awareness; a citation sends a signal and sometimes a click. Track both, separately.
  • SEO (search engine optimization). The older discipline of ranking pages in traditional search. AEO builds on it but optimizes for being quoted, not ranked. How they differ.
  • AEO tool. Software that tracks your citations and mentions across AI engines and tells you where you stand. What to look for in one.
  • SERP (search engine results page). The page you get back from a search. The term predates AI, but it matters here because AI answers increasingly sit on top of the classic SERP rather than replacing it — so a page can lose clicks to an AI summary while still ranking.
  • Blue links. The old shorthand for the ten organic results on a traditional SERP. "Ten blue links" is the thing AEO is moving beyond: being ranked in that list is no longer the same as being quoted in the answer above it.
  • SGE (Search Generative Experience). The original name for Google's AI answers, announced in May 2023 as an opt-in Search Labs experiment. It was renamed and rolled out as AI Overviews in 2024, so if you see "SGE" in older guides, read it as AI Overviews.
  • AI answer. The synthesized block of text an engine writes in response to a question, as opposed to a list of links. It's the real estate AEO competes for — a brand is either named inside it or it isn't.
  • Synthesis. The step where an engine blends several retrieved sources into one coherent answer in its own words. It's why your exact sentence rarely appears verbatim, and why clear, quotable, self-contained claims survive synthesis better than hedged prose.
  • Answer engine results. The citations, source cards and linked sources an answer engine shows alongside its text. Think of them as the AI-era equivalent of ranking positions — the slots you're trying to occupy.

The engines and surfaces

One honest nuance worth conceding: these surfaces don't behave the same way, and a brand can be strong on one and invisible on another. A citation on Perplexity and a mention in AI Overviews are not the same event, which is why measuring each engine on its own beats a single blended number.

Google's AI surfaces

Google alone has a confusing stack of AI and semi-AI boxes, and people mix them up constantly. Here's each one, roughly from newest to oldest, because the older "answer" formats are the direct ancestors of the AI ones.

  • AI Overviews. The AI-generated summary that sits above Google's organic results on many queries. It's the single biggest AI surface by reach, which is why losing visibility here stings — here's how to recover AI Overviews traffic.
  • AI Mode. Google's full conversational search tab, where you can ask follow-ups. It leans hard on query fan-out, running many sub-searches behind one question, so being the best answer to a narrow sub-topic can get you pulled into an answer you never directly targeted.
  • Gemini grounding. When Gemini "grounds" an answer, it runs a live Google Search and ties its reply to those results. The API returns the queries it ran and links the text back to sources, so grounded answers are the ones where a brand can actually earn a citation — more on Gemini and the web.
  • Featured snippet. The older, pre-AI answer box that quotes one page directly at the top of results. Google calls these "special boxes where the format of a regular search result is reversed" — pages that win snippets are often the same ones AI Overviews pulls from, so snippet-friendly structure is reusable.
  • Answer box. A loose umbrella term for any of Google's direct-answer boxes — featured snippets, dictionary definitions, calculators, sports scores. Worth knowing because "answer box" and "featured snippet" get used interchangeably even though the snippet is just one kind.
  • People Also Ask (PAA). The expandable list of related questions under many results. It's a free map of how real users phrase follow-ups, which makes it a cheap source of prompts to track and headings to write — AI answers and PAA draw from a similar pool of question intent.
  • Passage ranking. Google's ability to rank a single relevant passage from a page, not just the page as a whole, introduced in 2020 and expected to touch about 7% of queries. It's the classic-search echo of how AI retrieval works at the chunk level, and it rewards pages where each section answers its own question cleanly.

How engines find and use content

  • Large language model (LLM). A neural network trained on huge amounts of text to predict plausible language — the engine under every chatbot here. How LLMs actually work.
  • Next-token prediction. The core mechanism: the model generates one token at a time, each time guessing the most likely next piece given everything so far. Google's explainer describes models generating "probability distributions for next tokens given previous context."
  • Training data. The text a model learned from. Anthropic puts it plainly: models "aren't programmed directly by humans—instead, they're trained on large amounts of data." If your category is thin or outdated on the web, that's the pattern it absorbed.
  • Knowledge cutoff. The date after which a model's training data stops. Anything newer, the model only knows if it searches the web live.
  • Retrieval-augmented generation (RAG). A design where the model fetches relevant documents first, then writes its answer from them — introduced in a 2020 paper by Lewis et al. It's why a brand can be cited today without waiting for the next training run.
  • Retrieval. The fetch step inside RAG: pulling candidate passages that might answer the query.
  • Reranking. Scoring those passages so the most relevant ones get fed to the model, and cited, first.
  • Grounding. Tying an answer to real sources so it's checkable. Google says grounding helps "reduce model hallucinations by basing responses on real-world information."
  • Query fan-out. Splitting one user question into several background searches across subtopics, then merging the results. It's Google's term for how AI Mode works, and it means one brand can win on sub-questions you never typed.
  • Hallucination. When a model states something false or invented with full confidence. Grounding and retrieval reduce it; they don't eliminate it.
  • Token. The chunk of text a model actually reads — roughly three-quarters of a word in English, per OpenAI.
  • Tokenization. The step that splits text into tokens. Rare or invented brand names fracture into several fragments, which can make them harder for a model to surface cleanly — you can check yours in OpenAI's tokenizer.
  • Embedding. A list of numbers that represents a piece of text's meaning, so the model can judge relatedness by distance. OpenAI calls it "a vector (list) of floating point numbers."
  • Vector database. A store for embeddings that lets a retrieval system find the closest matches to a query quickly.
  • Semantic search. Search by meaning, using embeddings, rather than exact keyword matching.
  • Context window. How much text a model can hold at once — the prompt, everything retrieved, and its own answer, all counted in tokens.
  • System prompt. Hidden instructions that set a model's behavior before your question even arrives.
  • Fine-tuning. Extra training on narrower data to specialize a base model. Brand visibility work almost never touches this — you influence the training data and what's retrievable, not the model's weights.
  • Temperature. A setting that controls randomness in generation. Higher values make answers more varied — part of why the same prompt isn't perfectly repeatable.
  • Chunk / chunking. Before a page can be retrieved, a system splits it into smaller pieces — chunks — usually a few sentences or a section each. Retrieval then happens at the chunk level, not the page level, which is why one tight, self-contained section can get cited while the rest of the page is ignored.
  • Passage. A single chunk of text considered on its own merits. In practice, "write in passages" means each section should stand alone and answer one question — the opposite of a long, winding argument that only makes sense top to bottom.
  • Vector. The list of numbers an embedding produces. Two pieces of text with nearby vectors are treated as related in meaning, which is the whole trick behind semantic retrieval.
  • Cosine similarity. The most common way to score how close two vectors are — roughly, the angle between them, where 1 means identical direction and 0 means unrelated. It's the math under "find me the passages most similar to this question."
  • Semantic similarity. Closeness in meaning rather than wording. It's why a page that never uses the exact query phrase can still be retrieved if it covers the same idea — and why keyword stuffing matters far less than it used to.
  • BM25. A long-standing keyword-ranking formula that scores a document by term frequency, rare-word weighting and length. Elastic's plain-English walkthrough is a good primer. It still powers the keyword half of most retrieval stacks, so exact-match terms haven't stopped mattering.
  • Hybrid search. Running keyword search (like BM25) and vector search together, then merging the results. Microsoft describes it as combining "the strengths of vector search and keyword search" — it's now the default in most serious retrieval systems, which is why you want to win on both exact terms and meaning.
  • Grounding metadata. The structured record an engine can return alongside a grounded answer — the searches it ran and the links tied to each span of text. Google's grounding docs expose this, and it's the raw material tools use to tell whether you were actually cited.
  • Inline citation. A source link attached to a specific sentence or claim in an answer, rather than a pile of links at the bottom. Inline citations are the highest-value outcome in AEO because they tie your brand to a particular statement the user is reading.
  • Source card. The little clickable panel — logo, title, snippet — that engines like Perplexity and Copilot show for each source. It's a citation with branding attached, so a clear title tag and favicon do real work here.
  • Function calling / tool use. The mechanism that lets a model call developer-defined tools — a search API, a database, your MCP server — instead of answering from memory alone. OpenAI's function-calling guide covers the pattern; it's how an assistant goes and fetches live facts about your brand mid-answer.
  • AI agent / agentic. A setup where the model drives its own multi-step process — deciding what to search, read or do next — rather than following a fixed script. Anthropic draws the line between agents that "direct their own processes" and simpler workflows; agentic search means the engine may run several rounds of retrieval before it writes a word.

Crawling and access

  • AI crawler. An automated bot that fetches pages for an AI company — to train models, build a search index, or grab a page live when a user asks. How to check they can reach your site.
  • GPTBot. OpenAI's training crawler, which OpenAI says is "used to make our generative AI foundation models more useful and safe."
  • OAI-SearchBot. OpenAI's search crawler — the one that decides whether you can appear in ChatGPT's search answers. Block it and you stay out of those answers.
  • ClaudeBot. Anthropic's crawler. Site owners block it with a robots.txt rule that names ClaudeBot, repeated per subdomain.
  • PerplexityBot. Perplexity's crawler, which it says is "designed to surface and link websites in search results on Perplexity."
  • Google-Extended. A robots.txt token that controls whether Google may use content it already crawled to train future Gemini models. It does not affect your Search ranking.
  • robots.txt. The file at your site root that, in Google's words, "tells search engine crawlers which URLs the crawler can access on your site."
  • llms.txt. A proposed standard file that gives AI agents a clean, linked overview of your site. How to add one.
  • User agent. The identifier a crawler sends with each request, like "GPTBot" — it's what you match in robots.txt to allow or block a specific bot.
  • Crawl budget. How much of your site a crawler will fetch in a given window. Thin, slow, or bloated sites can get only partially crawled.
  • Structured data (schema). Machine-readable markup from the shared schema.org vocabulary that labels what's on a page — product, FAQ, author — so engines parse it reliably.
  • Entity. A distinct thing an engine recognizes — a brand, person, or product — and connects to known facts. Consistent, widespread mentions sharpen how clearly an engine "knows" your entity.
  • Applebot-Extended. Apple's control token for AI training. Apple says publishers "can opt-out from having their content used to train generative foundation models by disallowing Applebot-Extended in robots.txt" — blocking it doesn't remove you from Apple search, only from training.
  • Bytespider. ByteDance's crawler, associated with TikTok's parent company and widely flagged as aggressive. Many sites block it in robots.txt because it fetches heavily and its purpose is opaque — a reminder that "AI crawler" isn't one friendly category.
  • CCBot / Common Crawl. CCBot is the crawler of Common Crawl, a nonprofit that publishes a free, open snapshot of the web. It matters out of proportion to its name because Common Crawl's archive has been a major training-data source for many models — so being in it (or not) shaped a lot of what models already "know."
  • noindex. A <meta> tag or HTTP header that tells search engines not to index a page. Google says it "will drop that page entirely from Google Search results" — a blunt tool worth double-checking, since a stray noindex on a key page quietly removes it from the pool AI answers draw on.
  • Client-side rendering. When a page ships mostly empty HTML and builds its content with JavaScript in the browser. Some crawlers don't execute that JavaScript, so they see a blank page — if your core content only appears after JS runs, an AI crawler may never read it.
  • TDM / text-and-data-mining opt-out. A legal-flavoured opt-out, rooted in EU copyright rules, letting rights-holders reserve their content from automated mining and AI training. In practice it's expressed through the same robots tokens and metadata above — the point is that blocking training and blocking retrieval are different decisions with different consequences.

Measuring it

  • Citation rate. How often your brand is cited across a set of tracked prompts. This is the surface we built FixAEO to measure — it tracks your prompts and their fan-out across nine engines, shows which AI crawlers reach your pages, and scores visibility as citation rate. How to tie it to ROI.
  • Visibility score. A single rolled-up number for how present your brand is across engines and prompts. Definitions vary by tool, so always check what's actually being counted.
  • Share of voice. Your slice of all brand mentions in a category — how often AI names you versus your competitors. Monitoring competitor mentions.
  • Sentiment. Whether an AI describes your brand positively, neutrally, or negatively. A mention isn't automatically a good one.
  • Prompt. The question or instruction given to an engine. In tracking, a prompt is one test query you measure your visibility against.
  • Prompt volume (demand). How often real people actually ask a given question. High-demand prompts are the ones worth winning first.
  • Prompt set. The collection of prompts you track for a brand, ideally weighted by real demand rather than guesses.
  • Citation decay. The way a brand's citations fade over time as content goes stale or competitors publish fresher answers — a reason to monitor continuously, not once.
  • Citation share. Of all the citations an engine hands out across your tracked prompts, the slice that goes to you. It's more honest than a raw citation count because it adjusts for how competitive each prompt is — being the only brand cited on a niche question isn't the same as being one of ten.
  • Mention rate. How often your brand name appears in answers, cited or not, across a prompt set. Pair it with citation rate: a high mention rate with a low citation rate means the model knows you but isn't linking you, which is usually a content-structure problem.
  • Answer coverage. The share of your tracked prompts where you show up at all. It answers "in how many of the questions I care about am I present?" — a breadth metric, where citation rate is more about depth on each one.
  • Prominence / position. Where in an answer you land — named in the first sentence, buried in a list, or relegated to a footnote citation. Being mentioned first or framed as the recommendation is worth far more than a trailing link, so good tracking records position, not just presence.
  • Benchmark. A fixed reference point you measure against — your score last month, a competitor's, or a category average. Without one, a citation rate is just a number; the benchmark is what tells you whether it's good.
  • Prompt sampling. Because answers vary run to run, tools ask the same prompt several times and average the result. It's how you separate a real trend from the model's natural randomness — a single reading can mislead.
  • Brand entity recognition. Whether an engine correctly identifies your brand as a distinct entity rather than confusing it with a similarly named company or a common word. Weak recognition is a silent killer: the model can't recommend what it can't cleanly tell apart.

Tactics and signals

  • E-E-A-T. Google's shorthand for content that "demonstrates aspects of experience, expertise, authoritativeness, and trustworthiness." The same signals that help in Search tend to help in AI answers.
  • Topical authority. Depth and breadth of coverage on a subject, which makes an engine more likely to treat you as a go-to source.
  • Content freshness. How recently content was published or updated. Retrieval-based answers lean toward current pages.
  • UGC / Reddit citations. User-generated content — forum posts, reviews, Reddit threads — that AI engines cite heavily because it reads as real, unpolished experience. Why AI cites Reddit.
  • Comparison and listicle content. "Best X for Y" and "A vs B" pages, which AI answers pull from constantly when users ask for a recommendation. Getting products recommended by AI.
  • MCP (Model Context Protocol). An open standard that lets AI assistants connect directly to live tools and data sources. What our MCP does.
  • Digital PR. Earning mentions and links on reputable sites — still one of the strongest ways to build the web footprint an engine learns from.
  • Agent analytics. Measuring which AI crawlers hit your pages and how often — server-log visibility into bots, kept separate from human traffic. More on agent analytics.
  • sameAs. A schema.org property that points from your page to your canonical profiles elsewhere — your Wikipedia page, Wikidata entry, LinkedIn, Crunchbase. Schema.org says it links to a URL that "unambiguously indicates the item's identity", which is one of the cleaner ways to tell engines "this brand is that brand."
  • Knowledge panel. The box Google shows for a recognized entity. Google defines these as "information boxes that appear on Google when you search for entities" — having one is a strong signal that Google treats you as a real, disambiguated entity, which tends to help across AI surfaces too.
  • Wikidata. A free, collaborative, machine-readable knowledge base that a lot of systems read structured facts from. A correct, complete Wikidata item is one of the few entity signals you can influence directly, and it feeds far more than Google.
  • Wikipedia presence. Whether your brand has a Wikipedia article. It's hard to earn and governed by strict notability rules, but it's disproportionately influential — models trained heavily on Wikipedia inherit its framing of what your category contains, and who's in it.
  • First-hand experience. Content that shows you actually did or used the thing — real screenshots, numbers, mistakes — not a rewrite of what everyone else published. It's the "experience" in E-E-A-T, and it's a big part of why AI engines lean on forums; here's how to earn those first-hand citations.
  • Original research / statistics pages. Pages where you publish your own data — a survey, a benchmark, an analysis. These get cited far above their weight because AI answers need numbers and there are only so many sources for any given stat; our own AI search statistics roundup is built to be exactly this kind of linkable asset.
  • Internal linking. How your own pages link to each other. Beyond passing ranking signals, a clear internal structure helps crawlers find and understand your content, and it keeps related passages close enough that an engine can assemble a fuller picture of your expertise on a topic.

Risks and governance

AEO isn't only about getting cited — it's also about how your content and brand are used, and sometimes misused, by these systems.

  • Prompt injection. An attack where hidden or crafted text steers a model into doing something its developers didn't intend. OWASP ranks it the top LLM security risk. It matters for publishers because instructions buried in a page an agent reads can, in theory, manipulate the agent — so never treat page content an AI fetches as automatically safe.
  • Model card. A short document describing a model's intended use, training and limitations — a concept introduced in a 2018 paper by Mitchell and colleagues. Reading the model card for an engine you care about tells you what it was and wasn't built to do, which beats guessing from its answers.
  • Data provenance. The trail of where an engine's facts came from — which sources, how fresh, how trustworthy. Grounding and inline citations exist partly to make provenance visible; weak provenance is how confident-sounding wrong answers happen.
  • Opt-out. The umbrella term for telling AI systems not to use your content — via robots tokens like Google-Extended and Applebot-Extended, metadata, or legal reservations. The key nuance, worth repeating: opting out of training is a separate switch from opting out of live retrieval, and blocking the wrong one can quietly make you invisible in answers.

Shopping is moving into the answer box, and it has its own small vocabulary that product brands need.

  • ChatGPT Shopping. Product results inside ChatGPT search, with images, prices and review summaries. OpenAI says these are "selected independently" and "not ads", chosen from structured product metadata — which means clean, complete product data is how you earn placement, not a media buy.
  • Product feed. A structured file of your catalog — titles, prices, availability, images — that engines and marketplaces can ingest. Getting your feed into the right places is increasingly how products become eligible to appear in AI shopping answers; it's the commerce cousin of structured data. How to get products recommended by AI.
  • Merchant program. A direct relationship where a retailer submits its feed to an AI platform for first-party product data. These are early and evolving, but they point to a world where product visibility in AI answers is partly a feed-management discipline, the way it already is on Google Shopping.
  • Buying-intent prompt. A question asked by someone ready to buy — "best noise-cancelling headphones under $200," not "how do headphones work." These are the highest-value prompts for a commerce brand to win, and they behave differently from informational ones, which is a core reason ecommerce AEO is its own playbook.

Frequently asked questions

What's the difference between AEO, GEO and LLMO?

Honestly, not much — they're three labels for the same work: getting your brand surfaced and cited by AI answer engines. AEO (answer engine optimization) is the most common, GEO (generative engine optimization) is a broader framing, and LLMO (large language model optimization) is the newest alias. There's no standards body settling it, so most practitioners treat them as synonyms and move on; here's the fuller breakdown.

No, but its payoff is shifting. The same foundations — crawlable pages, clear structure, real expertise, a solid link profile — still decide whether an engine can find and trust you. What changes is the goal: you're optimizing to be quoted inside an answer, not just ranked in a list, and a page can now rank well yet lose the click to an AI summary above it. How AEO and SEO actually differ.

What does a citation actually get me?

Two things: a referral click when the user taps the source, and a trust signal when the AI names you as the basis for its answer. The second often matters more — being framed as the recommendation shapes a buyer's shortlist before they ever reach your site. A mention without a link still builds awareness, which is why it's worth tracking mentions and citations separately.

How is AI visibility measured?

You pick a set of prompts real people ask in your category, run them across the engines repeatedly, and record how often and how prominently you show up — as a citation rate, mention rate or share of voice against competitors. Because answers vary run to run, good measurement samples each prompt several times and averages the result. You can run a free check to see where you stand, or read how to tie it to ROI.

Do I need a separate strategy for each engine?

The foundations are shared, but the surfaces aren't. A citation on Perplexity, an appearance in AI Overviews, and a recommendation in ChatGPT are different events, and a brand can be strong on one while invisible on another. So the sensible approach is one solid content and entity base, then engine-specific tracking so you can see where you're weak and tune — not five disconnected strategies.

No. Semantic search is about matching meaning rather than exact keywords, using embeddings. Query fan-out is a step above that: the engine splits your one question into several related sub-queries, runs them (often with semantic search underneath), and merges the results. Query fan-out explained.

What's the difference between a mention and a citation?

A mention is when an answer names your brand in its text; a citation is when it links to your page as a source. A citation sends a signal and sometimes a click, while a mention mostly builds awareness. You can have plenty of mentions with few citations — usually a sign the model knows you but your content isn't structured to be quoted — so track both.

Keep this page open the next time one of these words gets thrown around on a call. And if you want to stop guessing where your brand actually stands, you can run a free AI visibility check and see your citation rate across the engines above.

Found this useful? Share it

X LinkedIn Reddit

Summarize with AI

Open this post in an AI engine.

ChatGPTClaudePerplexity

Related reading

Free AEO tools

Put this into practice with free FixAEO tools — no signup required.