The vocabulary, defined precisely enough to argue with.

Several of these terms are used loosely across the category, and the looseness is doing work — “visibility” and “localised” in particular. Here is what each one means here, including where our definition is narrower than the industry's.

GEOGenerative Engine Optimization

The practice of improving how a brand appears in AI-generated answers. It is the term that has mostly won: around 42% of practitioners choose it when forced to pick one, and roughly 84% recognise it. There is no academic consensus distinguishing it from AEO or LLMO, and practitioners stack the terms freely.

AEOAnswer Engine Optimization

A synonym for GEO, preferred by several large vendors. Usable, and a minority label. If you are writing for an audience rather than for search, GEO is the safer choice.

AI visibilityThe tool category's name

What this class of product is called: measuring, rather than doing. Distinguished from GEO, which is the work of changing the outcome.

The denominator problemYour score is a share of prompts you chose

Every AI-visibility figure is computed over a prompt set someone authored. Choose flattering prompts and the score is meaningless in one direction; choose irrelevant ones and it is meaningless in the other. No tool solves this for you; a published recipe and a visible prompt list make it defensible.

Mention rateHow often the brand is named at all

The proportion of runs in which the brand appeared. No competitor denominator, so it can rise for every brand in a category at once — which is itself useful information.

Share of voiceThe brand's share of all competitor-set mentions

Zero-sum by construction: for one brand's share to rise, another's must fall. Meaningless without a published competitor set, which is why ours is printed with the report.

Citation shareCited sources that are brand-owned

The proportion of an answer's cited sources belonging to domains the brand owns. Only as correct as the declared owned-domain set — getting that set wrong understates it silently and by a lot.

ProminenceWhere in the answer the mention fell

First-named, in the body, or in a trailing list. Averaging these into one visibility figure discards the distinction that matters commercially.

SentimentTone of the language around a mention

Off by default, because classifying every captured answer is a real cost line. Counter-intuitively the most stable thing in the system: how a brand is described, once mentioned, varies less than whether it is mentioned.

Stability classAlways, contested, or never

A per-prompt, per-engine classification drawn from the same samples that produce the bands. Roughly 77% of cells are near-deterministic; the contested minority is where intervention can actually move something.

RunOne prompt, one engine, one sample

The atomic unit. One prompt, one engine, one language, one clean session, one moment. Every figure in the product is an aggregate over runs, and every figure states how many.

CycleOne period's measurement for one brand

Every prompt × every engine in the market mix × every sample, orchestrated as a batch workload on a durable worker tier. Tolerates partial returns rather than failing whole.

BandA figure reported as a range with its run count

A binomial confidence interval over the period's runs, e.g. `28–36% · n=150`. The reporting contract of the whole product: nothing renders as a bare percentage.

Prompt groupOne intent, several phrasings

Two or three paraphrases of the same question, tracked as one rolled-up question with independently reportable members. Past around five samples, a new phrasing reduces error four to fifteen times faster than another repeat.

ProvenanceHow a given number was acquired

Method, engine, engine version, locale and geo parameters, language, timestamp, and whether a web search actually fired. Recorded on every capture, not derived at report time.

CaptureThe stored raw answer

The verbatim answer an engine returned, written to object storage before anything is derived from it. Reachable from every metric it contributed to. Around 57 KB each; roughly 47,000 per brand-year at the default configuration.

Declared instrumentSaying exactly how you measured

The honest replacement for “our numbers are what consumers see”, which no vendor can defend. Instead: here is the method, the run count, the cadence, the geography and the model version, and here is the answer behind the number.

Grounded APIA model API with web search enabled

The acquisition method for the assistant engines. Scalable and compliant, and lower-fidelity than the consumer interface: fewer words, fewer sources, and search fires on only a fraction of runs. Recording which runs searched is what keeps the gap honest rather than hidden.

Clean sessionNo memory, no history, no personalisation

Each sample is issued without carried-over context, so what is being measured is the engine's answer to the question rather than the engine's answer to a conversation.

CoverageWhich engines actually contributed to a period

Vendors fail. A cycle that reached five of six engines reports five, widens its bands accordingly, and says so on the report — rather than presenting a five-engine aggregate as a six-engine one.

MarketA configuration object, not a locale string

Engine mix, geo parameters, languages, AI Overview availability profile, device treatment and local-intent rules. Adding a country is a configuration exercise plus a validation pass.

TombstoneA deleted capture that still answers

When a capture's payload is erased, the row survives so a link to it returns “this was deleted” rather than “this never existed”. The distinction matters to anyone auditing a historical report.

AI OverviewsGoogle's generated summary above results

Reaches billions of users monthly. Triggers on roughly 36% of informational queries against 5–8% of transactional ones, so coverage varies enormously with prompt intent rather than with brand performance.

AI ModeGoogle's conversational search surface

Live in South Africa since August 2025 and available in around 200 countries. Afrikaans, isiZulu, Sesotho and Setswana were added to it and to AI Overviews in March 2026.

Tri-state presencePresent, absent, or not detected

Third-party detection of AI Overviews runs at roughly 68% at best, so about a third of genuinely-present overviews return nothing. Recording that as “absent” corrupts a trendline permanently; the third state keeps the uncertainty visible.

Engine mixPer-market weighting of engines

The aggregate figure is a weighted roll-up across the engines a market actually uses, not a flat average across every engine that exists.

Model eventA provider update that moves everything at once

Model updates cause step-changes: one release redistributed roughly 47% of citations within 48 hours. Engine version is recorded on every capture so these appear as annotations on a trendline rather than as an unexplained cliff.

AI crawlerGPTBot, ClaudeBot, PerplexityBot, and friends

The agents that fetch pages for training and for live answering. They are separate bots with separate purposes, none of them executes JavaScript, and default CDN bot-management frequently blocks them without anyone deciding to.

llms.txtA proposed standard that nobody reads

A file intended to guide AI systems around a site. Across half a billion observed bot events, requests for it numbered in the hundreds, and Google has stated it ignores the file. We check for it, report it, and do not score it.

A term missing, or defined wrongly?

This page is meant to be argued with. If we have a definition wrong — or are using a term to do work it should not be doing — say so.