AI search visibility has roughly 60 working terms, most of which describe measurement rather than tactics, and the four that matter most are citation rate, mention rate, share of voice, and entity resolution.
The rest of this page defines the other 56, grouped by what they describe, with a primary source attached wherever the term comes from a specification or a vendor’s own documentation rather than from industry habit.
Vocabulary in a young field does two jobs. It lets people work together, and it lets people sound authoritative without saying anything. This glossary is written to serve the first job. Every definition is short enough to quote, and where a term has a formal definition somewhere, the definition here defers to it and links out.
One test runs through the whole list. If a term describes something you could measure, it earns its place. If it describes an attitude toward AI search, it is not here.
Which four terms actually matter?
If you take nothing else from this page, take these. They are the four numbers a board conversation about AI visibility eventually reduces to, and the four that most reporting confuses with each other.
- Citation rate
- The share of answers to a fixed prompt set that link to your domain. A rate, not a count, because a citation exists only for the answer that produced it.
- Mention rate
- The share of answers that name your brand at all, with or without a link. Always higher than citation rate, and the gap between them is diagnostic on its own.
- Entity resolution
- Whether an engine can confidently identify what your company is and distinguish it from anything with a similar name. The floor beneath the other three: unresolved entities are rarely named.
An engine that cannot tell what you are will not name you, so the other three sit at zero regardless of how good your content is.
Fix here and the rest become measurable.Once you are resolvable, you start being named. This is the first number that responds to third-party evidence.
Rising here with nothing else moving is normal and fine.Being named is easier than being linked. A wide gap means engines learned about you from someone else’s page.
The gap is the diagnostic, not either number alone.It moves when a competitor changes, not just when you do. Treat a drop as a question about the market before treating it as a failure.
The other three can all hold steady while this falls.Everything below supports those four. The groups are ordered by how often the terms come up in a working week, not alphabetically, because an alphabetical glossary is a lookup table and this is meant to be readable start to finish once.
What do the measurement terms mean?
Most of the vocabulary in this field describes measurement, which is a reasonable signal that the field is still working out what is worth measuring. These are the terms that decide whether a number means anything.
- Prompt set
- The fixed list of questions run against every engine. Fixing it is what makes a time series comparable, and changing it mid-quarter destroys the comparison you were building.
- Unbranded prompt
- A question phrased the way a buyer asks before they know the vendors, containing no brand name. The only kind that tests whether an engine surfaces you unprompted.
- Branded prompt
- A question naming your company directly. Useful for checking accuracy, useless for checking visibility, because you supplied the answer in the question.Not to be confused with: Unbranded prompt.
- Run
- One execution of one prompt against one engine at one moment. A single run is a sample of one and tells you nothing on its own.
- Requery frequency
- How often the same prompt is re-run per engine per day. The number that separates a measurement from a spot check, and the first thing to ask any vendor.
- Volatility band
- The range an untouched brand’s numbers move within, week to week, from nothing but engine variance. Any change smaller than the band is noise.
- Baseline
- The volatility-adjusted starting position recorded before any fix is attempted. Without one, no later result is falsifiable.
- Consideration set
- The brands an answer actually names when the question is a buying question. Appearing in it is the outcome most B2B teams are really buying.
- Sentiment
- What was said about you when you were named, scored rather than assumed. Occasionally reveals that a rise in mention rate is costing you.
- Name order
- Where in the answer your brand appears relative to others. Not a ranking, but consistently last in a list of five is a different position from consistently first.
- Answer archive
- The stored full text of every answer collected. Storing a yes or no on mention makes the archive worthless for any question you have not already thought of.
- Engine coverage
- How many distinct answer surfaces a tool queries. Coverage is not a feature list item; a tool that misses the engine your buyers use is measuring somebody else’s market.
- Prompt intent
- What stage of the buying process a prompt represents, from problem-aware through comparison to purchase. Different intents behave differently and should not be averaged together.
- Composite score
- A single number combining several measures. Useful for a trend, dangerous if the components are not published, because a score you cannot decompose is a score you cannot act on.
- Attribution gap
- The distance between being named in an answer and anything appearing in your analytics. Structural, not a tooling problem, and no vendor closes it by measuring.
- Hallucinated attribute
- A confident, specific and wrong claim an engine makes about you, most often pricing, integrations or compliance. Distinct from absence and fixed differently.
What do the engine and retrieval terms mean?
These describe how an answer gets built. You cannot change the architecture, but knowing it tells you which levers exist and which are imaginary.
- Answer engine
- Any system that responds to a question with generated prose rather than a list of links. The category includes assistants, search surfaces and in-product copilots.
- Retrieval-augmented generation
- The architecture behind most answer engines: fetch documents at query time, then generate an answer from them. It is the reason AEO can work at all. Source: Lewis et al., 2020
- Parametric memory
- What a model holds in its weights from training. Cannot be updated without retraining, and the RAG paper notes a model’s ability to access it precisely is limited. Source: Lewis et al., 2020
- Non-parametric memory
- The external corpus a system consults during a request. The half of the architecture you can actually influence. Source: Lewis et al., 2020
- Retriever
- The component that selects which documents reach the model. It runs before generation, so it is a filter no amount of good writing can bypass.
- Dense retrieval
- Matching on learned embeddings rather than shared words, so a passage can be retrieved without containing the query terms. Source: Karpukhin et al., 2020
- Sparse retrieval
- Classical term-matching such as BM25, which scores on word overlap. Dense retrieval outperformed a strong BM25 baseline by 9 to 19 points on top-20 accuracy. Source: Karpukhin et al., 2020
- Embedding
- A numeric vector representing meaning, positioned so that similar text sits close together. What lets a passage match a question that shares none of its words.
- Chunk
- The passage-sized unit a page is split into before indexing. Engines retrieve chunks, not pages, which is why a section that cannot stand alone rarely gets cited.
- Chunk-level retrieval
- Scoring individual passages rather than whole documents. The practical reason every H2 section should answer its own heading.
- Grounding
- Tying a generated statement to a retrieved source. When grounding works you get a citation; when it fails you get a confident sentence with nothing behind it.
- Citation
- A source attached to a generated answer. Anthropic documents citations as always enabled for its web search, with the cited text carried alongside the URL.
- Hallucination
- Generated content not supported by any retrieved source. In this category it usually surfaces as a wrong fact about a vendor rather than an invented vendor.
- Context window
- How much text a model can consider at once. Relevant because a crowded retrieved set means fewer sources make it in, and being retrieved is not the same as being used.
The buyer’s question, as asked.
Scores chunks against the query, dense or sparse, and returns the top few.
EmbeddingChunkCorpusProse written from what came back, grounded where it can be.
GroundingCitationBM25 and its relatives score on term overlap, so a passage has to contain something close to the query’s words to be found at all.
Learned embeddings put similar text close together, so a passage can be retrieved without containing the query terms. Karpukhin et al. report 9 to 19 point gains in top-20 accuracy over a strong BM25 baseline.
What do the crawler and access terms mean?
The least glamorous group and the one that most often explains a zero. Every term here is a setting somebody on your team can check this afternoon.
- Crawler
- An automated client that fetches pages. The category matters more than the word: training, search and user-initiated crawlers behave differently and are configured separately.
- Product token
- The identifier a crawler presents in its user-agent string, matched case-insensitively against groups in robots.txt. Source: RFC 9309
- robots.txt
- The file that tells crawlers which paths they may fetch. It manages crawl traffic and is explicitly not a mechanism for keeping a page out of search results. Source: Google
- Robots Exclusion Protocol
- The standard behind robots.txt, formalised in 2022. It defines that the most specific matching rule wins and that an allow beats an equivalent disallow. Source: RFC 9309
- Search crawler
- The crawler behind an assistant’s answer index, such as OAI-SearchBot. Blocking it removes you from that product’s answers. Source: OpenAI
- Training crawler
- A crawler collecting data for model training, such as GPTBot. Blocking it is a defensible choice and, per OpenAI, independent of the search setting. Source: OpenAI
- User-initiated fetcher
- A fetch triggered because a person named a page. Both OpenAI and Perplexity note robots.txt rules may not apply to this class of visit. Source: Perplexity
- noindex
- A directive that keeps a page out of results. The correct tool where robots.txt is the wrong one, because a disallowed page can still be indexed if others link to it. Source: Google
- nosnippet
- A directive preventing any snippet being shown for a page. A blunt instrument in an answer-engine world, since a snippet is the extractable part. Source: Google
- data-nosnippet
- An attribute excluding a specific part of a page from snippets while leaving the rest usable. Useful for boilerplate you would rather not see quoted back. Source: Google
Feeds the answer index. Blocking it removes you from that product’s answers.
Do we want to be findable there?Feeds model training. Independent of the search setting, per OpenAI’s own documentation.
Do we want to be training data?Someone asked for your page by name. Robots rules may not apply to this class at all.
Nothing to decide, it will fetch.User-agent: OAI-SearchBot # product token, matched case-insensitively
Allow: /docs/ # most specific match wins
Disallow: /docs/internal/ # longer path, so this wins inside /docs/internal/
Disallow: /*.pdf$ # * is any run of characters, $ anchors the endWhat do the entity and identity terms mean?
An engine has to know what you are before it can decide whether to recommend you. These terms describe that layer, and they are where structured data does its real and narrower job.
- Entity
- A thing an engine can identify and hold facts about, as opposed to a string it merely matches. Brands become entities; brand names alone do not.
- Knowledge graph
- A store of entities and the relationships between them. Being in one is what lets an engine answer about you without retrieving a page.
- sameAs
- The schema.org property asserting that a described entity is the same as one identified elsewhere, defined as a URL that unambiguously indicates the item’s identity. Source: schema.org
- Structured data
- Machine-readable markup describing a page. Google positions it as helping search understand content, and makes no claim that it improves ranking position. Source: Google
- JSON-LD
- The markup format Google recommends for structured data, embedded in a script tag rather than woven through the HTML. Source: Google
- DefinedTermSet
- The schema type for a glossary, holding DefinedTerm entries. The type used by this page, which is the only honest way to publish a vocabulary. Source: schema.org
- Disambiguation
- Making it impossible for an engine to confuse you with something similarly named. The real job structured data does, as distinct from the job it is sold as doing.
- Notability
- The bar an entity clears to earn an independent reference. Wikidata admits an item describable “using serious and publicly available references”, which is a fair summary of the whole discipline. Source: Wikidata
What do the evidence and content terms mean?
What the corpus says about you, and who wrote it. Ranked roughly by how much weight it carries and inversely by how fast you can change it.
- Third-party evidence
- Anything written about you on a domain you do not control. The slowest layer to move and the one that moves the numbers most.
- Owned content
- Material on your own domain. It amplifies evidence that already exists and does very little on its own, which is why it belongs last in the sequence.
- Corpus
- The full body of text an engine can draw on about your category. Your site is a small and rarely decisive part of it.
- Review platform signal
- Structured, dated, third-party statements about a product on sites built for the purpose. Dense, citable, and frequently retrieved for comparison questions.
- Digital PR
- Earning coverage on domains you do not own, judged here by whether it produces citable statements rather than by referral traffic.
- E-E-A-T
- Google’s framing of experience, expertise, authoritativeness and trust. Its guidance asks whether it is self-evident who authored your content and whether pages carry a byline. Source: Google
- Named author
- A real person credited on a page, with a title and a traceable record. One of the few attribution signals entirely within your control.
- Living page
- A reference page updated in place at a permanent URL rather than republished. URL churn resets accumulated citations, which is the entire argument for the pattern.
Which terms get used wrong most often?
Three pairs account for most of the confusion in this category, and in every case the two words describe genuinely different things that happen to move together some of the time.
The share of answers that name your brand at all, with or without a link. Always higher than citation rate, and the gap between them is diagnostic on its own.
The share of answers to a fixed prompt set that link to your domain. A rate, not a count, because a citation exists only for the answer that produced it.
Your mentions as a proportion of all brand mentions across the same prompt set. The only one of the four that moves when a competitor changes, rather than when you do.
The brands an answer actually names when the question is a buying question. Appearing in it is the outcome most B2B teams are really buying.
The file that tells crawlers which paths they may fetch. It manages crawl traffic and is explicitly not a mechanism for keeping a page out of search results.
A directive that keeps a page out of results. The correct tool where robots.txt is the wrong one, because a disallowed page can still be indexed if others link to it.
A fourth confusion is worth naming without a diagram. “AI visibility” and “AEO” get used interchangeably, and they are not the same kind of word. Visibility is the outcome you measure. AEO is the work you do to change it. A vendor selling you visibility is selling a number; a vendor selling you AEO owes you a method.
This page is a living reference and gets revised as the vocabulary settles. If a term here is defined in a way you can show is wrong, that is worth more to us than a term we are missing.