Reference

Glossary

Precise definitions for the SEO, GEO, AI crawler and LLM visibility terms used across SiteRank AI, including what each term does not mean.

These are the definitions SiteRank AI uses in its product, its reports and its writing. Several of these words are routinely overstated in the field, so where a term invites an over-claim, the prohibition is part of the definition rather than a footnote to it. Terms are grouped loosely by subject and each one links onward where a longer treatment exists.

SEO

Search engine optimisation: the practice of making a site easier for conventional search engines to crawl, index, understand and rank. It covers technical work (status codes, canonicals, sitemaps, robots directives, rendering, speed), content work (matching intent, covering a subject properly), and architecture (internal linking, site structure).

SEO is the discipline in this glossary with the strongest evidence base. The mechanisms are documented by the engines themselves, the outputs are observable in tools those engines provide, and decades of published measurement exist. When something in SiteRank AI is labelled a standards or provider-guidance finding, it usually sits here.

Related: SEO audit · GEO

GEO

Generative Engine Optimization: making a site easier for AI-mediated discovery systems to retrieve, interpret and attribute. In practice it means machine-readable structure, clear entity identification, self-contained passages, factual consistency and visible source transparency.

GEO is treated here as an emerging discipline, not a settled one. Very little of it is confirmed by any provider, so every GEO recommendation in SiteRank AI carries an evidence tier, and the two lowest tiers never affect a status band. Google's stated position on its own AI features is that no additional requirements or special optimisations are needed beyond normal SEO, and nothing in this product contradicts that without a cited source.

Related: GEO · What is GEO?

AEO

Answer Engine Optimization: used in the field interchangeably with GEO. Both describe optimising for systems that answer a question directly rather than returning a list of links.

We prefer GEO for consistency and do not present AEO as a separate discipline with a separate evidence base. Where a vendor draws a sharp line between the two, that line is usually a positioning choice rather than a methodological one. Treat any claim that AEO techniques differ materially from GEO techniques as unsupported unless it comes with evidence.

Related: SEO vs GEO

LLM Visibility

The measured observation of whether an AI system mentions or cites a site or entity, for a controlled prompt set, under stated conditions. It is a measurement we made — not a property the site possesses.

That framing carries real constraints. A visibility figure is only meaningful alongside the prompt set it came from, the provider and model that produced it, whether web-search grounding was on, the locale, the date, and the number of samples. Change any of those and you have measured something else. Because most generative systems expose no deterministic ordered result list, LLM visibility is never expressed as a rank or a position.

Mention

An occurrence of a tracked entity, brand or domain in an AI-generated answer, with or without an accompanying source link.

A mention is not a citation. Being named in an answer means the model produced your name in its output; it says nothing about whether your page was retrieved, whether the provider attributed you as a source, or whether the statement about you was accurate. Mentions and citations move independently, and a report that blends them into one number hides the more actionable of the two. SiteRank AI stores each mention against the observation that produced it, with the matched surface form recorded.

Related: Citation · Mention rate

Citation

An explicit source attribution or link returned by the provider alongside an answer — a citations array, a message annotation, or an inline link, depending on the API.

Citations are stored with the URL as given, the resolved domain after unwrapping any redirect wrapper, the title, whether it matched the tracked site, and which response field it came from. A citation means a source was attributed. It does not mean the answer endorsed that source, that a person saw the link, or that the same question in a consumer chat product would produce the same sources. Cited is not recommended.

The proportion of observations in which one entity appears, relative to the observations mentioning any entity in a competitor set you defined, within a prompt set you defined.

Both denominators are always shown, because the number is meaningless without them. Share of voice is not market share, not brand awareness and not an industry statistic. It changes when you add a prompt, remove a prompt, add a competitor or remove one — which makes it useful for tracking a set you hold constant, and worthless for comparing against anyone else's figure.

Citation Rate

The proportion of valid observations in which the provider returned at least one citation to any source. Reported with its numerator, its denominator, the sample size and a 95% Wilson score interval.

Distinguish it from the domain citation rate, which counts only observations citing your domain. A high citation rate with a low domain citation rate is a common and informative pattern: the model is citing sources for these questions, and none of them are you. Failed calls — timeouts, errors, blocked responses — are excluded from both numerator and denominator, and the excluded count is reported, because a failed request is not evidence the model omitted you.

Mention Rate

The proportion of valid observations in which the tracked entity was mentioned in the answer text, with or without a citation. Reported with numerator, denominator, sample size and a 95% Wilson interval, never as a bare percentage.

At small sample sizes that interval is wide — at a single observation it spans almost the whole range — and showing it is the honest answer rather than a presentational problem. Mention rate should be read alongside volatility: if repeated samples of the same prompt under identical conditions disagree with each other frequently, a change in mention rate between two runs may be noise.

Related: Mention · AI visibility

Prompt Monitoring

Running a fixed, versioned set of prompts against a provider API on a schedule, recording each response as an observation, and computing metrics across them.

The prompt set is the instrument. It should reflect questions real people plausibly ask about your category, it should be approved by a human before it costs money, and it should change rarely — every edit bumps the prompt version and breaks comparability with earlier observations, which is why the version is stored on every row. Monitoring an API is not monitoring a chat product: the orchestration, retrieval stack and personalisation differ, and we never present one as evidence of the other.

AI Crawler

An automated agent operated by an AI company that fetches web pages. The category is too broad to be actionable on its own, because the agents inside it do at least three different jobs and every major provider controls them independently.

The useful subdivisions are the search crawler that builds an index an answer product queries, the training crawler that collects material for model training, and the user-triggered agent that fetches one URL because a person asked. "Block AI bots" is a single instruction against three separate decisions, which is why it so often blocks the agent that would have cited you.

Training Crawler

A crawler that collects web content which may be used to train foundation models. OpenAI's GPTBot, Anthropic's ClaudeBot and Google's Google-Extended are the documented examples, and each respects robots.txt.

Disallowing one is a licensing and business decision about future models. It is not a visibility decision: it does not remove you from that provider's current answer product, and Google states explicitly that Google-Extended does not affect Search inclusion and is not a ranking signal. Blocking a training crawler is defensible and so is allowing one, so SiteRank AI never reports either choice as a defect and never scores it.

Search Crawler

A crawler that builds the index a search or answer product queries at question time. Googlebot and bingbot are the conventional cases; OAI-SearchBot, Claude-SearchBot and PerplexityBot are the AI-answer equivalents.

Disallowing one removes you from that product's results — OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. This is the category where a robots.txt rule has a direct, documented visibility consequence, and it is the one most often blocked by accident when a site writes a blanket rule intended to decline training use.

robots.txt

A text file at a host's root that tells compliant crawlers which paths they may fetch. Standardised as RFC 9309, the Robots Exclusion Protocol.

Three properties are routinely forgotten. It is advisory: compliance is voluntary, so it is a policy, not a control — enforcement is authentication, a paywall or an edge block. It is per host, so a rule at example.com does not govern docs.example.com. And its absence is not the same as its permissiveness: a robots.txt that answers 404 means no rules were retrieved, and behaviour in that state is operator-specific. WordPress serves a virtual file when none exists on disk, which a server rewrite can break.

llms.txt

A proposed markdown file, published at /llms.txt, giving a language model a curated map of a site: an H1 naming the site, a blockquote summary, optional general content, and H2 sections of markdown links. An ## Optional section marks links an agent may skip when a shorter context is needed.

The format is a published, stable specification. The effect of publishing one is not. No provider has documented that it reads the file, prefers sites that publish one, or uses it as a retrieval input. Presence is not ingestion, ingestion is not retrieval, and retrieval is not citation. SiteRank AI treats it as an emerging practice and never scores a missing llms.txt as a deficiency.

Related: llms.txt · What is llms.txt?

Structured Data

Machine-readable statements embedded in a page describing what the page is about — its type, author, publisher, date, and subject entities. On the modern web this almost always means JSON-LD, a W3C Recommendation, using the Schema.org vocabulary.

Structured data is the most reliably machine-readable claim a page makes about itself, and Google documents which types are eligible for which rich results. What it is not is a citation mechanism: no provider documents schema as a retrieval or attribution input for generative answers. Markup that misrepresents the page, describes content not visible to users, or invents authors, ratings or reviews is a policy violation regardless of eligibility.

Schema.org

The shared vocabulary that structured data uses: a published set of types (Organization, Person, Article, WebSite, WebPage, BreadcrumbList, FAQPage and hundreds more) and the properties each type accepts. It is maintained collaboratively and is an established standard.

Knowing the vocabulary is the easy part; the common failure on a WordPress site is contradiction. Two plugins each emitting an Organization block with a different name, or two Article blocks with different authors, leaves the page with no single answer about who wrote and published it. Removing a duplicate is usually a larger improvement than adding a new type.

Entity

A named thing — an organisation, person, place, product, programme or concept — with a set of surface forms, a canonical label, and optionally an external identifier that disambiguates it.

Entity clarity means a site says who it is consistently: the same organisation name in the schema, in the page copy, in the about page and in the author bylines, ideally tied to an identifier that distinguishes it from every other business with a similar name. Inconsistency here is cheap to find and cheap to fix, and it is one of the few GEO-adjacent practices with a mechanism that does not require speculation about how retrieval works.

Topical Authority

The strength, completeness and consistency of a site's information about a subject area, as computed and explained by the tool reporting it.

It is not a Google metric, not an industry score, and not a number any third party can validate. When SiteRank AI reports it, the report decomposes into documents, clusters, inbound internal links and coverage counts you can click through to. There is no composite 0–100 figure anywhere, because a percentage you cannot decompose is not explainable. Treat any vendor's topical authority score as that vendor's own calculation, not as an observation about search engines.

Topic Cluster

A group of documents on a site that cover facets of one subject, usually organised around a pillar page with supporting articles linking to and from it.

In SiteRank AI a cluster is derived, not declared: documents are vectorised locally with TF-IDF and grouped by cosine similarity above a threshold, with no fixed number of clusters imposed, because the number of subjects a site covers is not known in advance. Documents resembling nothing else stay unclustered, which is itself informative. Generated cluster labels can be wrong, and the interface says so rather than presenting a derived string as a fact.

Pillar Page

The document that best represents a topic cluster comprehensively — the page you would send someone to first on that subject, and the page the supporting articles link back to.

SiteRank AI reports pillar candidates rather than naming one, ranked on inbound internal links, relevance to the cluster centroid, and word count — weighted in that order. Length is deliberately the smallest term: a long page nothing links to is a long page, not a pillar. Which document should carry the role is an editorial and commercial decision, so the tool supplies the evidence and leaves the choice to you.

Internal Linking

Links between pages on the same site. They let crawlers discover pages, distribute importance across a site, describe a destination through anchor text, and give a reader a route through a subject.

Their effect on crawl discovery and on how a destination is described is well established in conventional SEO. Their effect on whether an AI answer engine cites you is not documented by any provider and should be treated as emerging practice at best. A page with no inbound internal links is an orphan — usually by accident rather than intent — and finding orphans is one of the highest-yield outputs of a site-wide link graph.

Canonical

The URL a page declares as its preferred address, usually through a rel="canonical" link element. It tells search engines which of several URLs serving similar content should be treated as the one to index.

Canonicals are a hint, not a directive: an engine may choose a different canonical if the signals disagree. On WordPress they are a common source of plugin conflict, because several SEO plugins can emit one, and two contradictory canonicals on a page is worse than none. SiteRank AI detects which plugin owns canonical output and reports where a fix belongs rather than emitting a competing value.

Related: WordPress SEO · Indexability

Indexability

Whether a page is eligible to be included in a search engine's index. It depends on several independent things agreeing: the page is reachable, it returns a success status, it is not disallowed from being fetched, it does not carry a noindex directive, and it canonicalises to itself.

Crawlability and indexability are different. A page blocked in robots.txt is not crawled, which means its noindex tag is never read — a classic way to keep a page in an index you were trying to remove it from. Because it reads the database rather than inferring from HTML, a plugin can report indexability from the settings themselves.

Related: SEO audit · Canonical

Retrieval

The step where a system selects documents or passages to put in front of a language model in response to a query. It sits between crawling and generation, and it is the least documented link in the chain.

Being crawled makes you eligible for retrieval and nothing more. Being retrieved makes you eligible to be cited and nothing more. No major provider publishes how its retrieval component ranks or filters candidates for generative answers, so claims that a specific on-page technique causes retrieval are, at present, hypotheses. What can be observed from outside is the end of the chain: whether a source appeared in an answer.

Related: Grounding · Crawled is not cited

Grounding

Constraining a model's answer to retrieved source material rather than letting it answer from parameters alone. In practice it usually means a web-search tool the model can call during generation, with the retrieved sources returned as citations.

Whether grounding is enabled changes almost everything about an observation: which sources appear, how current the answer is, and whether citations are returned at all. It is therefore recorded on every observation SiteRank AI stores, and grounded and ungrounded runs are never pooled into one metric. Comparing a grounded result with an ungrounded one measures the configuration, not the site.

Related: Retrieval · Methodology

RAG

Retrieval-Augmented Generation: the general architecture in which a system retrieves relevant documents and supplies them to a language model as context before it generates an answer. Grounded AI search products are one application of it.

RAG matters to site owners for one structural reason: what gets supplied to the model is usually a passage, not a whole page. That makes passage-level structure — self-contained sections, clear headings, claims that survive being read out of context — a plausible practical concern. Plausible is the operative word; no provider documents the chunking or selection behaviour of its production systems, so this remains reasoning about mechanism rather than confirmed guidance.

Related: Passage retrieval · Retrieval

Passage Retrieval

Selecting and returning a section of a document rather than the whole document. Retrieval systems generally index and return passages, which is why passage length and self-containment are worth measuring at all.

In SiteRank AI a passage is the text between two headings, and passage structure is what the retrievability audit rule examines: sections long enough to carry a complete idea, short enough to be usable as context, and comprehensible without the paragraph above them. The evidence tier here is emerging practice with a plausible mechanism — not a documented requirement of any system.

Machine Readability

How easily an automated consumer can extract a page's meaning without rendering, guessing or scraping. It covers valid HTML semantics, a sensible heading hierarchy, text that exists in the markup rather than only after JavaScript executes, structured data that parses, and stable URLs.

Machine readability is the least speculative part of GEO, because it is mostly conventional technical quality restated for a different consumer. A page whose JSON-LD fails to parse is not partially understood — it is ignored entirely. That is a real, observable defect with an established fix, which is why it is graded as one rather than as a hypothesis.

Related: GEO · Structured data inspection

Source Transparency

The extent to which a page makes plain who produced it, when, and on what basis: a named author with real credentials, a visible publication and update date, an about page, a contact route, and identifiable organisation details.

These are useful to human readers regardless of any retrieval system, which is why they are worth doing. What must never happen is manufacturing them. Invented authors, fabricated credentials, made-up reviews, backdated timestamps and fictional organisation details are fabricated trust signals. If a field is unknown, it stays empty and the site owner is asked — SiteRank AI never fills one in.

Evidence Tier

The label attached to every finding and recommendation, saying how well supported it is. Six tiers, in descending order: ESTABLISHED STANDARD, OFFICIAL PROVIDER GUIDANCE, STRONG EVIDENCE, EMERGING PRACTICE, EXPERIMENTAL, HYPOTHESIS.

The tier is not decoration. Findings at the bottom two tiers never contribute to a status band, are never phrased as fixes, and are shown separately from standards compliance — and that exclusion is enforced in the query rather than left to a caller to remember. The tier system exists so that a genuinely speculative suggestion can be offered without being disguised as a defect.

Related: Methodology · Editorial policy

Observation

One recorded response to one prompt, under fully recorded conditions: provider, the model id the API actually returned, prompt id and prompt version, whether web search was enabled, locale, timestamp, and the raw response snapshot, with its extracted mentions and citations attached.

Provenance is not metadata here — it is the definition. A response stored without it cannot be compared, audited or reproduced, so it is not an observation and is not written. Only observations with a status of ok are valid for metrics: errors and timeouts are excluded from both numerator and denominator, and the excluded count is always reported.

Related: Methodology · Citation monitoring

Volatility

The variation between samples of the same prompt under identical conditions — specifically, the proportion of samples that disagree with the majority answer. Range 0 to 0.5.

It is a first-class reported metric, not an error term to be smoothed away, because it tells you how much a single observation can be trusted. The condition key includes the prompt id, prompt version, provider, returned model and grounding flag, because pooling different prompts that happen to share a model measures between-prompt variation instead. With fewer than two samples it reports as not measurable rather than as zero.

Of all citations to any domain in a tracked set — your own and the competitors you chose — the proportion that went to yours. It is the citation-level counterpart to share of voice, which counts mentions.

Like share of voice, it is a proportion within your own prompt set and your own competitor list, and both denominators are always displayed. It is not a share of anything the market measures. It is most useful held constant over time on a fixed prompt set and competitor list, and least useful the moment either changes.

Prompt Coverage

The proportion of currently enabled prompts that produced at least one valid observation within the reporting window.

This is a data-quality metric rather than a visibility metric, and it belongs next to every other number in a report. If only half your prompt set ran — because of rate limits, a budget ceiling, provider errors or a queue that never ticked — then every rate you are reading came from half the instrument. A high mention rate over four of twelve prompts is not a high mention rate. Showing coverage prevents that misreading rather than relying on the reader to check.

Related: Prompt monitoring · Methodology

User-triggered Agent

An agent that fetches a specific URL because a person asked for it inside a product — pasted a link, asked a question that needs a page, added a source. ChatGPT-User, Claude-User and Perplexity-User are the documented examples.

Provider positions on robots.txt differ here and the differences matter. OpenAI states that robots.txt rules may not apply to ChatGPT-User because the action is user-initiated. Perplexity states that Perplexity-User generally ignores robots.txt for the same reason. Anthropic documents Claude-User as respecting standard robots.txt exclusions. One blanket rule for "AI bots" erases all three distinctions.

SEO

GEO

AEO