AEO glossary: the terms used in AI search visibility work, defined plainly
Last updated October 14, 2026. Written by the Third Coast AI team.
This glossary defines the terms you will meet in answer engine optimization (AEO) and generative engine optimization (GEO) work, from the engines themselves (ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity, Copilot) through the way they choose sources (retrieval, grounding, citation, entity, selection and absorption) to the way the work is measured (prompt bank, share of voice, decision factor, verification run) and the technical plumbing that decides whether an engine can read you at all (GPTBot, OAI-SearchBot, robots.txt, llms.txt). Each entry is one to three sentences, written so a buyer can use it in a meeting and an engine can quote it without context.
The definitions reflect how we use the words at Third Coast AI and how the published research uses them; where practitioners disagree we say so. For the longer explanation of the discipline, start with what AEO and GEO are.
Which terms describe the engines and the answers they give?
| Term | Definition |
|---|---|
| AI engine, answer engine | A system that answers a question with written prose instead of a list of links. In this glossary it means ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity and Microsoft Copilot. |
| ChatGPT | OpenAI's assistant. With web search on it retrieves live pages through its search partners and OAI-SearchBot; with search off it answers from training data alone. |
| Google AI Overviews | The AI-written summary that appears above Google results for many queries. BrightEdge's tracking puts it on nearly half of Google queries in 2026. |
| Google AI Mode | Google's conversational search tab, which issues many related searches for one question and writes a single answer with citations. |
| Gemini | Google's assistant. With grounding on it retrieves from Google Search before answering, which makes its citations behave like AI Mode's. |
| Perplexity | An answer engine that cites many sources per answer and favors recent content. |
| Copilot | Microsoft's assistant, fed by the Bing index and by structured data on pages, which is why schema matters more there than elsewhere. |
| Grounding | Giving a language model live search results or documents to answer from, instead of relying on what it memorized in training. "Retrieval on" means the same thing. |
| Retrieval | The step where an engine goes and fetches candidate pages for a question before writing. Retrieval-augmented generation (RAG) is the technical name for the pattern. |
| Hidden sub-query | One of the related searches an engine runs behind a single question. Our captures show 10 to 20 behind a typical buyer question, consistent with what Google has described about AI Mode. Never confuse it with the question the buyer typed. |
| Zero-click | A search that ends without a visit to any site because the answer was on the results page. Pew measured click-through at 8% with an AI summary against 15% without. |
| AI referral traffic | Visits that arrive from a link inside an AI answer. Small in volume for most sites, and Semrush measured those visitors converting at about 4.4 times the rate of classic organic traffic. |
How does an engine pick its sources, in these terms?
These are the terms that matter most in the proof audit, and the two defined in the boxes below are the ones we use every day.
Selection and absorption
The two-stage model proposed by Zhang, He and Yao in 2026. Selection is whether the engine retrieves and considers your page at all, predicted by crawler access, index presence, brand priors and topical coverage. Absorption is how much of the page it uses, predicted by where the answer sits, statistic density, quotations, sources and structure; top-quartile list density was cited at 8.94 times the bottom quartile in their 21,000-citation sample.
Decision factor
A criterion an engine uses to choose between companies in a category, in the engine's own words, such as response time, licensing, years in business or price transparency. We extract them with decision-factor prompts, count how often each appears, and audit the client's pages for proof of each one.
| Term | Definition |
|---|---|
| Citation | A link or named reference inside an AI answer pointing to the page the engine used. Citation rate is the share of runs in which your page appears as a source. |
| Mention | Your company named in an answer, with or without a link. Being mentioned is the AEO outcome; being cited is the GEO outcome. |
| Entity | A thing the engine can identify with confidence across sources: one company with one consistent name, location, services and set of identifiers. Entity consistency is how well those sources agree. |
| Knowledge Graph | Google's database of entities and the relationships between them. Appearing in it as a distinct organization helps the engines resolve you. |
| Wikidata | The open, structured database that feeds many knowledge graphs. Any real organization can have an item, which gives the engines a stable identifier. |
| sameAs | A schema.org property on your Organization markup that lists your other profiles (LinkedIn, Wikidata, Google Business Profile), closing the chain between them. |
| Schema markup | Structured data in a page's code describing what the page is about. Hygiene everywhere, a real lever for Copilot and Bing, and weak evidence of driving citations elsewhere. |
| Brand prior | What an engine already believes about a company before it retrieves anything, shaped by training data and by how often the company is mentioned elsewhere. |
| Third-party corroboration | Other sites saying the same thing your site says. Ahrefs' 75,000-brand study found unlinked mentions correlate with AI citations at 0.664, against 0.218 for backlinks. |
| Earned media | Coverage you did not pay for: press, trade publications, podcasts, community threads. The 5W citation index attributed about 85% of the AI citations it analyzed to earned media. |
| Topical coverage | How much of a subject cluster your site answers, not just the head term. Pillar-and-cluster architecture is the usual way to build it. |
| Passage | A self-contained chunk of a page that answers one sub-query on its own. Engines cite passages, which is why answer placement matters more than page length. |
| Freshness | How recently a page was published or substantively updated. It matters for time-sensitive queries and matters little for evergreen ones; date changes without content changes are detected. |
| Clarifying question | A question an engine asks the user, or asks itself, before answering. In mining runs they reveal which attributes the engine thinks matter. |
| Sentiment | How an engine describes you, as distinct from whether it names you. Tracked separately because an accurate but lukewarm description is its own problem. |
Is there a separate vocabulary for the measurement and the method?
| Term | Definition |
|---|---|
| AEO (answer engine optimization) | Getting your company named, described accurately and cited when an AI engine answers a buyer's question. The business-goal side of the work. |
| GEO (generative engine optimization) | Making your pages the ones a model retrieves and quotes. The term comes from Aggarwal and colleagues at Princeton (KDD 2024), who measured quotations, statistics and cited sources lifting visibility in generated answers by about 41%, 33% and 28%. |
| LLMO (large language model optimization) | An older synonym for GEO that some practitioners still use. Same work, different label. |
| Share of voice | The share of runs for a prompt set in which your company is named, by engine, compared with the companies named instead. The headline AEO metric. |
| Visibility rate | Share of voice for one prompt or one topic, reported as a rate over repeated runs rather than a single yes or no. |
| Prompt bank | The fixed set of 60 to 120 prompts, written the way buyers ask, that is run and re-run to measure change. It is the measurement instrument, and the client should own it. |
| Visibility prompt | A prompt that asks the engine for a recommendation ("best [category] in [place]"). Measures whether you are named. |
| Decision-factor prompt | A prompt that asks what to look for in a category. Mines the criteria the engine uses. |
| Competitor prompt | A prompt that asks the engine to compare you with a named rival. Shows which sources the engine leans on and where its verdict comes from. |
| Fact-check prompt | A prompt that asks the engine a factual question about your company (location, price, services). Finds what the engine gets wrong. |
| Proof audit | A page-by-page check of your site against the decision factors: stated, proven with a number or source, or missing. |
| Citation read | Reading the pages an engine already cites for a topic to learn their shape: answer position, statistic count, tables, heading phrasing. The brief for your page comes from here. |
| Publish gate | A fixed checklist a page must pass before it goes live: answer first, sourced statistics, a quote, a list, a table, nothing invented, all of it in the raw HTML. |
| Verification run | Re-running the prompt bank after new pages are indexed and logging the change per prompt, by engine, with misses included. |
| Snapshot versus rate | A single run of a prompt is a snapshot and is unreliable, because answers vary by phrasing and session. A rate over repeated runs is the only honest number. |
| Correction campaign | A page, or set of pages, written to fix something the fact-check prompts showed the engines getting wrong about you. Often the fastest visible result. |
| Tracking tool | Software that re-runs prompts on a schedule and reports mentions and citations. We use PromptWatch and Profound; none of the tools capture hidden sub-queries, which is why we also run our own prompt miner. |
Does the technical plumbing need its own vocabulary?
It does, because a surprising share of AEO problems are a page the engines cannot read. Cyrus Shepard's May 2026 synthesis of 54 studies scores URL accessibility 9.5 out of 10, the highest of the 23 factors he reviewed, and the terms below are the ones that come up when we check it.
| Term | Definition |
|---|---|
| Crawler, bot | Software an engine uses to fetch pages. Each engine runs its own, and each can be allowed or blocked separately. |
| GPTBot | OpenAI's crawler for collecting training data. Blocking it affects what future models know about you. |
| OAI-SearchBot | OpenAI's crawler for live search results in ChatGPT. Blocking it removes you from ChatGPT's retrieval. |
| ClaudeBot, PerplexityBot, Bingbot | The equivalent crawlers for Anthropic, Perplexity and Microsoft (Bing feeds Copilot and part of ChatGPT's retrieval). |
| robots.txt | The file on your site that tells crawlers what they may fetch. A CDN's bot protection can block AI crawlers even when robots.txt allows them, so test with raw requests. |
| llms.txt | A proposed plain-text file describing a site for language models. Read by some developer tools, scored 2.0 out of 10 as a citation factor in Shepard's synthesis, and cheap to publish. |
| Raw HTML | The page as delivered by the server before any JavaScript runs. ChatGPT's retrieval does not execute JavaScript, so text rendered by scripts is invisible to it. |
| Organization schema | The schema.org markup that states who you are: name, URL, logo, address, sameAs. The one piece of structured data we treat as mandatory. |
| NAP consistency | Name, address and phone number matching exactly across every profile. Helps the engines resolve a business as one entity. |
| Manual action | A penalty applied by a Google reviewer. Self-promotional listicles and scaled AI-generated pages have both drawn them. |
| Bing Webmaster Tools AI Performance report | A free first-party report showing how Bing's AI features use your site. Worth verifying for any company regardless of who does its AEO. |
Why do we avoid some common terms?
- We do not describe the hidden searches behind a question with the industry's shorthand; we say hidden sub-queries, because the plain phrase tells a buyer what is happening.
- We do not say "ranking" in an AI engine, because the engines do not rank; they name and cite, and a page can be cited from the 21st position or lower, as Semrush's AI search traffic study found.
- We do not promise a "score" as an outcome, because a single number from a tool hides the prompt-level change that is the real result.
- We do not use "AI SEO" as a service name, because it suggests the SEO playbook transfers whole, and Ahrefs measured the overlap between Google's top ten and AI Overview citations falling from 76% to 38% between July 2025 and March 2026.
Half the arguments about AI search are two people using one word for two things. Citation and mention are different outcomes, selection and absorption are different stages, and a rate is not a snapshot. Agree on the words first and the plan gets shorter.
Third Coast AI
If a term you have met is missing, the method guide uses most of these in context, the tools comparison covers the vocabulary the software vendors use, and the FAQ answers the questions about us.
Frequently asked questions
A mention is your company named in an AI answer, linked or not. A citation is your page listed as a source the engine used. Mentions are the AEO outcome and citations are the GEO outcome, and a company can have either without the other.
The related searches an engine runs behind one question before it writes an answer. Our captures show 10 to 20 behind a typical buyer question, and a page that answers the typed question but none of the hidden ones is rarely cited.
That the engine fetched live pages before answering, rather than answering from training data alone. ChatGPT with web search and Gemini with grounding are the retrieval-on modes we run, because they show which pages the engine would cite today.
They describe the same work from two sides. AEO is the goal of being the company an engine names; GEO, the term from the 2024 Princeton paper, is the page-level craft of being the source it quotes. Most practitioners use them interchangeably.
Only if you have decided you do not want to appear in AI answers. Blocking OAI-SearchBot removes you from ChatGPT's retrieval, and BuzzStream found publishers that blocked crawlers were still cited through cached data and third-party mentions, so blocking rarely protects anything.