The AEO Glossary: Every Term You Actually Need to Know

An AEO glossary is a reference for the vocabulary used in answer engine optimization, covering the crawling, indexing, retrieval and citation terms that decide whether AI systems can find and quote a company. The useful version of that reference does something most glossaries skip, which is telling you where each term came from, because roughly a third of the words in circulation are written down in a public specification and the rest were invented by people selling services.

That distinction is the whole point of this page. If a term appears in an internet standard, you can hold a vendor to it. If it came out of a research paper, it has a fixed meaning you can check. If a company coined it for its own product, it means whatever that company says it means. And if the marketing industry made it up, two agencies will define it differently and neither is wrong.

So every entry below carries a provenance label. We use four.

Label	What it means	How much weight to give it
Standard	Written down in a published specification that implementers follow	Treat as fact. You can cite the spec in an argument
Research	Comes from published academic or industry research with a stable definition	Reliable, though implementations vary in the details
Vendor	One company’s product, feature or crawler name	Accurate only for that company, and it can change without notice
Coined	Invented by the marketing industry, no authoritative definition exists	Ask whoever used it what they meant. They may not know

The terms are grouped by where they sit in the pipeline rather than alphabetically, because alphabetical order scatters related ideas and this subject is hard enough to hold in your head already. If you want a single term, use the contents list. If you are trying to build a mental model, read straight through, because the groups run in the order the machine does the work.

Where the vocabulary sits, in the order the machine works.

The acronyms themselves

Four competing names for roughly the same work. Since these are the terms people actually search for, they belong near the top rather than buried at the back.

AEO, answer engine optimization

Coined. Making a company retrievable and quotable by systems that answer questions directly instead of returning links. It won the popularity contest without ever acquiring a fixed scope, so one agency’s AEO is content rewriting and another’s is technical crawl work. Ask which before you sign anything.

GEO, generative engine optimization

Research. The one term here with a real origin. It was introduced in a paper presented at KDD 2024 by researchers from Princeton and IIT Delhi, along with a benchmark for measuring it, and the paper reported visibility gains of up to 40 percent from content changes alone. That figure is an upper bound rather than a typical result, and the same paper notes effectiveness varies by domain.

LLMO, large language model optimization

Coined. Points at the wrong component. The model is not the part you can influence. The retrieval system in front of it is, and that is where essentially all of the work happens. Accurate only if you genuinely mean influencing training data, which operates on a timescale of model releases rather than weeks.

AI SEO

Coined. Ambiguous in a way that causes real confusion, because it is used for two unrelated things. Sometimes it means optimising to appear in AI answers. Sometimes it means using AI tools to do ordinary SEO faster. Those are different budgets and different outcomes, so it is worth asking which one is on the table.

The crawl layer

Everything here decides whether a machine is allowed to fetch your page at all. It is the least glamorous part of the subject and the part where we find the most broken.

Robots Exclusion Protocol

Standard. The rules governing how automated clients ask permission before fetching content. It spent 28 years as an informal convention invented by Martijn Koster in 1994 before being published as RFC 9309 in 2022, which is now an IETF standards-track document. The RFC is short and worth reading once. It specifies that crawlers should not use a cached copy for more than 24 hours, that parsers must handle at least 500 kibibytes of the file, and that the protocol is explicitly not a security measure.

robots.txt

Standard. The plain text file at the root of a domain that implements the protocol above. It grants or withholds permission per user agent and per path. It cannot enforce anything, so a crawler that ignores it faces no technical obstacle, only a reputational one. Most AI visibility problems we are hired to solve turn out to begin in this file.

User agent

Standard. The name a client announces itself under, and the thing robots.txt rules attach to. The practical consequence for AEO is that one company runs several agents doing different jobs, so a single blanket rule usually produces an outcome nobody intended. Matching is by substring, which means a rule naming one agent does not automatically catch a similarly named sibling.

X-Robots-Tag

Standard. An HTTP response header carrying indexing directives such as noindex, documented in the MDN reference. It matters here because it is invisible in the page source. A page can look perfectly fine in a browser and carry a header telling every compliant system to leave it out. Check headers, not just HTML.

Content-Signal

Coined. A robots.txt extension expressing what a publisher permits per use case, separating search indexing from model training from real-time grounding. It is not part of RFC 9309 and no standards body has adopted it. It carries real legal weight anyway, because publishers use it to state an express reservation of rights, and text-and-data-mining law in some jurisdictions gives that reservation force.

The index layer

Once a page can be fetched, something has to store it in a form that can be searched later. This is where the biggest gap sits between how people imagine AI search works and how it actually works.

Corpus

Research. The body of documents a system can draw on. The critical thing to understand is that there is no single corpus. Each engine maintains its own, they overlap unevenly, and being present in one tells you close to nothing about the others. This is why a company can be quoted confidently by one assistant and be entirely absent from another on the same question.

Sitemap

Standard. An XML file listing the URLs you want discovered, defined by the Sitemaps protocol. It is a hint about discovery rather than a promise of inclusion. Still worth maintaining accurately, and worth checking that the one your robots.txt advertises is the one your site actually serves, because those drift apart more often than you would expect.

IndexNow

Vendor. A push protocol letting a site tell participating engines that a URL changed, rather than waiting to be re-crawled. Documented at indexnow.org. Adoption is partial, so it speeds up some engines and does nothing for others. Cheap to implement and not a lever that will fix a visibility problem on its own.

Embedding

Research. A numerical representation of text as a list of numbers, positioned so that passages with similar meaning sit near each other. Hugging Face has a readable primer if you want the mechanics. The consequence for writing is that retrieval matches on meaning rather than exact wording, so hitting a keyword exactly matters far less than being unambiguous about what you are describing.

Vector search

Research. Finding content by comparing embeddings rather than matching words, explained plainly in Elastic’s overview. Most production systems combine it with keyword matching rather than replacing one with the other, which is why classic on-page clarity has not stopped mattering.

Chunk, or passage

Research. The unit a document gets split into before being indexed and retrieved. This single idea explains more AEO outcomes than anything else in this glossary. Your page does not compete as one object. It competes as dozens of fragments, and one good fragment can carry a page that ranks nowhere while thirty-nine others contribute nothing.

The retrieval layer

This is the part that runs at the moment somebody asks a question, and it is where the difference between AI search and classic search actually lives.

Retrieval-augmented generation

Research. Usually shortened to RAG. The architecture where a model consults a searchable index at question time instead of relying only on what it absorbed during training. The framing that helps most is that a model has two kinds of memory, one baked in during training and one looked up live. Answer engine optimization is almost entirely about the second.

Query reformulation

Research. The step where the system rewrites a person’s question into one or more machine-friendly search queries before retrieving anything. Sometimes called query fan-out when it generates several at once. This matters because you are not competing for the phrasing the buyer typed. You are competing for whatever the system rewrote it into, and you do not get to see that.

Reranking

Research. A second scoring pass over the retrieved candidates, usually with a slower and more accurate model than the one used for the initial sweep. Practical consequence: getting retrieved is not the same as getting used. A passage can clear the first filter and lose at the second, which is one reason visibility can move without anything on your site changing.

Context window

Vendor. How much text a model can hold at once, counted in tokens, and set per model by whoever built it. Retrieved passages compete for room inside it alongside the question and the system instructions. Windows have grown enormously, which has made the old advice about writing very short passages much less important than it was two years ago.

Grounding

Research. Tying a generated statement to a specific retrieved source rather than letting the model assert it unsupported. A grounded answer is one where each claim traces back to a document. Grounding is the mechanism that makes citation possible at all, so anything that improves how cleanly your content can be tied to a claim improves your odds of being named.

Hallucination

Research. A confident statement that is not supported by any source. Relevant to AEO in a way people underrate, because models hallucinate about companies. They invent pricing, attribute features to the wrong vendor, and describe products that were discontinued. The NIST AI Risk Management Framework is the useful reference for the vocabulary around this, and it is worth reading if you need to explain the risk to a board rather than to a marketing team.

The answer layer

Answer engine

Coined. Any system that responds to a question with a composed answer instead of a list of links. Nobody owns the definition, and the boundary is genuinely blurry, since a traditional search engine showing a summary at the top of a page of links is doing both things at once.

AI Overview and AI Mode

Vendor. Google’s two answer surfaces. An AI Overview is the summary block above conventional results. AI Mode is the separate conversational experience. Both draw on Google’s ordinary search index, which is why Google’s own documentation says no special markup or files are needed to appear in them.

Citation

Vendor. A named, usually clickable reference to a source inside a generated answer. Worth being precise here, because the word gets used for two different things. There is the visible link a reader can follow, and there is the internal fact that a passage was used to construct the sentence. The first is measurable from outside. The second is not, and the gap between them is where most reporting on this goes wrong.

Mention

Coined. Your company being named in an answer without necessarily being linked. For most B2B buying questions this is the thing that actually matters, because a buyer who reads your name in a shortlist has been influenced whether or not they clicked.

Zero-click

Coined. A search that ends without the person visiting any website. It predates AI by years, and AI answers have made it much more common. The term is a description of a measurement problem rather than a strategy, so treat anyone offering a zero-click solution with care.

The entity layer

This layer runs underneath the other four. It is about whether a machine knows which thing in the world your brand name refers to, and it is the most neglected part of the whole subject.

Entity

Research. A specific thing that can be distinguished from other things with the same name. Your company is an entity. So is the common noun your company is named after, and so is the larger firm in another sector using the same word. If a retrieval system cannot tell which one a question refers to, being quotable does not help you, because the quote may be attached to somebody else.

Knowledge graph

Vendor. A structured store of entities and the relationships between them. Every large search company runs one and none of them publish the contents, so you cannot inspect your own record directly. What you can do is make the facts on your own site unambiguous and consistent with what third parties say about you.

JSON-LD

Standard. A way of embedding machine-readable statements in a page using JSON, published as a W3C recommendation. It is the format almost everyone uses for structured data now. Worth separating the format from the vocabulary, since JSON-LD is the syntax and schema.org is the dictionary of terms you express in it.

sameAs and disambiguation

Standard. A structured-data property that points from your entity to an authoritative record of the same entity elsewhere. It is the cheapest disambiguation signal available and almost nobody uses it. The HTTP Archive Web Almanac found JSON-LD on 41 percent of pages, up from 34 percent two years earlier, while pages pointing sameAs at Wikidata sat at 0.17 percent and Wikipedia at 0.13 percent.

Two in five pages carry structured data. Fewer than two in a thousand use it to say which entity they are. Source: HTTP Archive Web Almanac 2024

Wikidata

Standard. A free, structured, machine-readable knowledge base with a stable identifier for each entity, introduced here. It is read widely by systems that need to resolve a name to a thing. Having a record is not something you can simply create for yourself, since inclusion has its own notability rules, but knowing whether one exists for your company is a five-minute check most companies have never run.

The crawler names, in one place

All Vendor, all subject to change, and all worth knowing by name because they are the things your robots.txt is talking to. The pattern to notice is that each company separates training from retrieval, and the user-initiated agents behave differently again.

Agent	Company	Job	Blocking it costs you
GPTBot	OpenAI	Training data collection	Nothing in today’s answers
OAI-SearchBot	OpenAI	Search and linking	Presence in ChatGPT search
ChatGPT-User	OpenAI	User-initiated fetches	Not a normal crawler, so rules may not apply
ClaudeBot	Anthropic	Training data collection	Nothing in today’s answers
Claude-SearchBot	Anthropic	Search indexing	Presence and accuracy in Claude
Claude-User	Anthropic	User-initiated fetches	Visibility for user-directed search
PerplexityBot	Perplexity	Search and linking, explicitly not training	Presence in Perplexity results
Perplexity-User	Perplexity	User-initiated fetches	Little, since it generally ignores robots.txt
Google-Extended	Google	Gemini training opt-out only	Nothing in Search or AI Overviews

Proposals that are not standards yet

llms.txt

Coined. A markdown file at the root of a site offering a clean, hand-written summary of your content plus links to plain versions of your pages, proposed by Jeremy Howard in September 2024 with a second version in August 2026. The status is worth being exact about, because it gets misrepresented in both directions.

It is a community proposal, not an RFC and not backed by a standards body. Google’s documentation states you do not need to create AI text files to appear in its AI features. And yet thousands of sites publish one, OpenAI and Anthropic and Google all publish llms.txt files for their own developer documentation, and Chrome’s Lighthouse audits for its presence. So the honest summary is that it is a widely adopted convention with no guarantee attached, which is cheap to add and should not be sold to you as a ranking factor.

The measurement layer

These are the terms you will argue about in reporting meetings. Most are Coined, which is exactly why you should agree on definitions before anybody presents a number.

Prompt set

Coined. The fixed list of questions you test against, written down once and then left alone. It is the closest thing this field has to a keyword list, and the discipline is the same. If the prompt set changes between reports, the reports are not comparable, and a surprising number of vendor dashboards quietly rotate theirs.

Baseline

Coined. Your measured position before you change anything, taken across repeated runs rather than once. Without one you cannot separate a real improvement from ordinary variance, and variance in this field is large enough that a single before-and-after comparison is close to worthless.

Citation share

Coined, but countable. The proportion of runs across a fixed prompt set in which you are named. Expressed as a fraction with the denominator stated, so 34 of 120 runs rather than a score out of a hundred. It is the metric we use, because anybody can recount it and disagree with us using the same data.

Crawl-to-refer ratio

Coined. How many pages an AI platform fetches from you for every visitor it sends back. Published as infrastructure data by network providers, and the figures run into the thousands to one for some platforms. Treat published ratios as directional rather than precise, because traffic arriving through mobile apps carries no referrer and therefore undercounts the referral side.

Log file analysis

Standard practice. Reading your own server access logs to see which agents requested what, and what status code they received. It is the only measurement in this entire glossary that reports ground truth rather than inference, since it records what actually happened on your server. It is also the one almost nobody does.

Variance

Research. The spread of different answers to the same question, driven by the index changing underneath you and by the model sampling probabilistically as it writes. It is the reason one check is a sample rather than a measurement, and the reason any AI visibility figure quoted without a run count should be treated as an anecdote.

Terms we will not use, and why

All Coined, all in circulation, and all vague enough that we avoid them in client reporting.

Share of model. Presented as a percentage of some total, without anybody defining the total. Share of which prompts, on which engine, sampled how often? Without those three answers the number is decoration.

AI visibility score. A single composite number, usually out of a hundred, produced by a proprietary formula. It moves, which feels like information, and you cannot decompose it into anything you could act on.

LLM SEO. Mixes up the model with the retrieval system in front of it. Nearly all the work is aimed at the retrieval system, and the name points at the wrong component.

We do use citation share, because it is countable. Pick your prompt set, run it repeatedly across engines, and count the proportion of runs in which you are named. That number is boring, reproducible and arguable, which is what you want from a metric.

Working that number out for your own domain is a couple of hours of manual runs, and if you would rather see it done once before you build the habit, our free visibility check reports it with the denominators attached.

Frequently asked questions

Is AEO an official term?

No. It is a coined marketing term with no authoritative definition, which is why two agencies will scope it differently. The nearest thing to a formal term is generative engine optimization, which at least originated in a peer-reviewed paper rather than a pitch deck. We use AEO because clients use it.

What is the difference between a citation and a mention?

A citation names you and usually links to you. A mention names you without a link. Both influence a buyer, only one sends traffic, and standard analytics records neither. For most B2B categories the mention is the more valuable of the two and the harder one to measure.

Do I need to add an llms.txt file?

Not to appear in Google’s AI features, since its documentation says so directly. It is inexpensive, widely adopted as a convention, and carries no guarantee from any provider. Add it if you want to, and do not let it be sold to you as the reason your visibility will improve.

Why does robots.txt matter more than schema for AI visibility?

Because one is a gate and the other is a hint. A blocked crawler produces total absence that no amount of markup can repair, whereas structured data improves clarity for systems that already reached you. Sequence follows from that, and it is the reason we look at the file before the markup.

What is the single most useful term here?

Chunk. Once you accept that retrieval operates on fragments rather than pages, most of the confusing behaviour in this field becomes predictable, including why a page that ranks nowhere gets quoted constantly and why your best page can be invisible.

---

If the vocabulary is new to you, the complete guide to AEO covers how these pieces fit together and which of them actually decides whether you get cited.

We update this page as terms settle or fall out of use, and we mark the date when a definition changes rather than editing silently. The newsletter is where those changes go out.