The AI visibility glossary

Plain definitions for the metrics and terms used across AI visibility work. Every entry says what the term means and, just as importantly, what it does not tell you — because most bad reporting comes from a metric answering the wrong question.

See how these metrics are measured · Go to the measurement hub

In short

This glossary defines the vocabulary of AI visibility across five groups: measurement metrics, how assistants work, technical and retrieval terms, research and method terms, and risk terms. Each entry gives a one-sentence definition followed by the limits of that term — what it cannot tell you on its own. It exists because most disagreements in this category are definitional: two people say share of voice and mean different formulas. Use it as a reference while reading anything else on the site, or as a shared vocabulary for an internal discussion.

Who this is for

• SEO and growth leads who need to explain an AI visibility metric to someone who controls budget
• Agency strategists standardising the language they use in client reporting
• Founders trying to work out which of four dashboard numbers actually answers their question
• Analytics and RevOps people mapping AI terms onto measurement they already run
• Anyone auditing a vendor report and wanting to check whether its terms are used consistently

When to use this

• A report or demo used a term you cannot define precisely
• You are writing a measurement spec and need consistent definitions
• Two sources disagree and you suspect they are measuring different things
• You are new to the category and want the vocabulary before the tactics

How to use this glossary

Read the definition first, then read the limitation. The limitation is the part that keeps you out of trouble, because nearly every metric in this category is defensible on its own and misleading when quoted alone.

Two habits make the vocabulary usable. First, whenever someone reports a number, ask which definition they used and over which prompt set — the same word covers several different formulas in different tools. Second, never quote a figure without its sample size and run dates attached; a defensible figure ships with its prompt set, model versions, run dates, sample size, and variance, and links to its method.

Each entry here is complete on its own; a linked article gives the longer explanation where one exists.

Measurement metrics

These are the numbers that end up in a report. They are related but not interchangeable, and choosing the wrong one is the most common reason an AI visibility report misleads the person reading it.

• AI visibility — the umbrella term for how present a brand is in AI-assistant answers for a defined set of prompts. It does not tell you whether that presence was positive, whether it influenced a buyer, or whether it will hold next month.
• Mention rate — the share of sampled answers in which a brand name appears at all. It does not tell you whether the brand was recommended, dismissed, or listed as an also-ran, and a high mention rate can coexist with unfavourable framing.
• Citation share — the share of the sources an assistant cited across a prompt set that belong to a given domain. It does not tell you whether the brand was named in the answer text, and citation display differs between assistants, so it is not comparable across models without care.
• Share of voice — a brand's presence expressed as a proportion of all brand presence across the same prompt set. It does not tell you the size of the underlying opportunity, and it moves when competitors move even if nothing about you changed.
• Sentiment and framing — how a brand is characterised when named, including position in the answer and whether it is presented as a recommendation, a caveat, or a warning. It does not reduce cleanly to a single positive or negative score, and automated sentiment labels are frequently wrong on short comparative text.
• Citation vs mention vs link — three separate events: being named in the answer text, having a source attributed, and having a clickable destination shown. Conflating them breaks reporting, because a brand can be mentioned with no citation and cited with no mention.
• Baseline — the first documented measurement against which later runs are compared, including its prompt-set version, model versions, run dates, and raw responses. It does not become more accurate with age, and a baseline recorded without its method is not a comparison point at all.

How assistants work, and the words people use for it

These terms describe the system being measured. Getting them wrong leads people to reason about AI answers as though they were a ranked list of blue links, which produces the wrong metrics and the wrong tactics.

• Answer engine — a system that returns a synthesised answer rather than a list of documents, usually naming a small number of options. It does not have a stable position ordering in the way a search results page does, so asking for your position is the wrong question.
• Non-determinism — the property that the same prompt can produce different answers on different runs, even with the same model. It does not mean the output is random or unmeasurable; it means one answer is a sample, not a measurement.
• Grounding — the practice of tying a generated answer to retrieved source material rather than model recall alone. It does not guarantee accuracy, and a grounded answer can still misread or misattribute the source it cites.
• Training data — the corpus a model learned from, fixed at training time. It does not update when you change your website, which is why some answers stay wrong long after the underlying fact changed.
• AI Overview — a search-embedded generated summary shown above conventional results. It is not the same surface as a conversational assistant, has different triggers and different sourcing, and should be measured separately rather than folded into an assistant score.
• Entity — the machine-readable notion of a distinct thing, such as a company, product, or category, and the attributes attached to it. It does not exist because you asserted it once on your homepage; consistency across your site and third-party sources is what makes it stable.
• AEO and GEO — answer engine optimisation and generative engine optimisation, two names for the practice of trying to influence AI answers. They do not describe a settled discipline with proven techniques, and vendors use both terms inconsistently.

Technical and retrieval terms

This group covers the preconditions. If a retriever cannot reach, render, or parse your content, none of the measurement or influence work above can help you.

• Retrieval — the step where a system fetches external documents at answer time to inform its response. It does not mean your page was read in full, and being retrieved is not the same as being cited or named.
• AI crawler — an automated agent that fetches pages on behalf of an AI provider, either for training corpora or for live user-triggered fetches. It does not always identify itself consistently, and blocking one does not necessarily block the other.
• Rendering — whether page content is present in the fetched response or only appears after client-side JavaScript executes. It does not follow that anything a browser shows a human is available to a retriever.
• Chunking — the splitting of a document into passages that can be retrieved and quoted independently. It does not respect your intended narrative order, which is why self-contained sections get quoted correctly and dependent ones do not.
• Structured data — schema markup that states explicitly what a page and its entities are. It does not force a model to use those facts, and marking something up is not evidence that any assistant consumed it.
• llms.txt — a proposed file for signalling preferred content to language-model consumers. It does not have demonstrated, broad support, so treat it as an experiment rather than a lever with known effect.

Research and method terms

These describe how a number was produced. In a non-deterministic system they are not procedural details, they are the difference between evidence and an anecdote.

• Prompt set — the versioned, dated list of questions used in a measurement, ideally phrased the way buyers actually ask. It does not represent the whole of demand, and a set that names your brand in the question measures nothing useful.
• Sampling — running each prompt multiple times to characterise a distribution of answers rather than capture one. It does not remove uncertainty; it quantifies it, and too few runs cannot detect a small change at all.
• Variance — the spread of results across repeated runs of the same prompt. It does not indicate a broken measurement; a stated variance is what separates a defensible figure from a headline.
• Confidence range — the interval within which the true value plausibly sits given your sample size. It does not make a small sample reliable, and a range wide enough to include both a rise and a fall means the result is inconclusive.
• Methodology version — the numbered, changelogged specification a given measurement was produced under. It does not travel with a figure automatically, so comparing numbers from different method versions without noting the change is a reporting error.
• Drift — a change in measured visibility caused by the underlying system rather than by anything you did, such as a model update or a change in retrieval behaviour. It does not announce itself, so distinguishing drift from your own impact requires stable prompts, stable run windows, and honesty about confounders.
• Brand baiting — writing prompts that mention your brand and then reporting the resulting mention rate as visibility. It does not measure discovery; it measures the model's willingness to talk about a name you supplied.

Risk terms

These appear in this glossary so you can recognise them, evaluate exposure, and describe them accurately to a legal or comms colleague. RankRift covers all of them defensively — what they are, how they are detected, and what they cost — and never as instructions.

• False model claim — an assistant stating something untrue about a brand, such as a service it does not offer or an event that did not occur. It is not necessarily fixable at the model, so the practical response starts with correcting the sources the claim likely came from.
• Prompt injection — content crafted to alter how a model behaves when that content is retrieved. From a brand owner's perspective it is a threat model to monitor for, not a technique, and detecting it does not tell you who placed it.
• Astroturfing — coordinated inauthentic posting designed to look like organic sentiment. It does not stay hidden reliably, and platform enforcement and regulatory exposure both land on the brand rather than the agency that ran it.
• Review fraud — fake or improperly incentivised reviews. It does not sit in a grey area legally in many jurisdictions, and detection by review platforms is increasingly systematic.
• Parasite and doorway content — thin pages published on borrowed authority to capture visibility. It does not build a durable position, and the cleanup cost typically exceeds whatever short-lived presence it produced.

Anything on this page touching law or regulation is general information, not legal advice. Consult qualified counsel about your situation.

How to extend a definition

Use the short definition and its limitation as the stable reference. Follow a linked article for a worked method; if no article is linked, the glossary entry is the complete maintained explanation.

If a definition here conflicts with how a tool uses the term, ask the tool's vendor for its formula and sample size before comparing numbers. Definitional disagreements are worth resolving in public, and this glossary is meant to be the shared reference the category currently lacks.

• How method versions and changelogs work
• The correction protocol

Related pages

• Measurement hub
• Methodology
• AI search fundamentals
• Start Here
• On-page measurement resources

Frequently asked questions

Which metric should I actually report to leadership?

Usually one presence metric and one comparative metric, both with a confidence range. Mention rate answers whether you show up; share of voice answers how that compares with the other names in the same answers. Citation share is a diagnostic for content and PR work rather than a headline number. Reporting all four without explaining which question each answers is how a report ends up sounding impressive and meaning nothing.

Why do two tools give different values for the same term?

Because the term does not have an industry-standard formula. Two tools can both call something share of voice while differing on what counts as a brand appearance, how they handle repeat mentions in one answer, which models they query, how many runs sit behind the figure, and whether they normalise across prompts. Ask for the definition and the sample size before comparing any two numbers.

Is a mention the same as a recommendation?

No, and treating them as equivalent inflates most reports. An answer can name your brand while advising against it, listing it as an alternative someone else outgrew, or mentioning it only as context for a competitor. Framing and position within the answer carry the meaning. That is why sentiment and framing sit beside mention rate rather than underneath it as an optional refinement.

How often do these definitions change?

The definitions themselves should be stable, because a metric that gets redefined quietly destroys every historical comparison built on it. What changes is the methodology version underneath: how a study samples, which models it queries, how variance is reported. Those changes get a version bump and a changelog entry before any dependent content publishes, so you can tell whether two figures are comparable.

Do I need to understand the statistics to use these terms?

Not to use them, but yes to defend them. The plain-English definition is enough for a conversation. The moment you present a number to someone who controls budget, you need to be able to say how many runs produced it and how wide the uncertainty is, because that is the first question a skeptical CFO asks. The sampling guide and the calculator cover exactly that.

Why does every entry include what the term does not tell you?

Because the failure mode of this category is a technically correct number answering a question nobody asked. Each of these metrics is defensible in isolation and misleading when quoted alone, and the gap between the two is where reports go wrong. Publishing the limits alongside the definition is the cheapest available protection against confidently reporting the wrong thing.

Turn the vocabulary into a measurement

Definitions are only useful once you have run something. The measurement hub covers prompt-set design and sampling, and the manual snapshot guide gives you a procedure for a first sampled snapshot.

Go to the measurement hub · Use the manual snapshot