The words we use, defined once

The plain words used across the site come first. Then the precise terms our method pages use, each linked back to its plain word.

The words on the rest of the site

AI assistant: A chatbot people ask for advice before they buy. We measure ChatGPT, Claude and Perplexity, each asked through its maker's API. The technical pages call each one an engine. See Engine.
Named: The AI assistant names the business in the text of its answer, the part a reader sees. The share of answers that do is what the technical pages call answer presence, or the named rate. See Answer presence (named rate).
Used as a source: A page on the business's own website is among the sources an AI assistant lists behind its answer. Being used as a source and being named can point in opposite directions, and only answers with web search switched on list their sources. The technical pages say cited, and call the share citation presence, or the cited rate. See Citation presence (cited rate).
How sure we are: A headline rate comes with a range: where the true figure could plausibly sit, given how many answers we read. More answers make the range narrower. We say a number moved only when the change, compared question by question, is too large to be day-to-day wandering. The technical pages call the range a 90% confidence band, or 90% interval. See 90% band (90% interval) and Paired difference (verdict).
The same questions: The customer questions agreed with a client at the start, none naming the business. They are asked unchanged at every measurement, so the before and the after are the same test. The technical pages call this the frozen question set, identified by its fingerprint (a hash). See Frozen question set and Fingerprint (hash).
With web search switched on: The AI assistant searches the web before it answers, the way the consumer apps do for a buying question, and lists the sources it used. It is how we ask the questions, and it is what makes being used as a source measurable. The technical pages call this grounded. See Grounded.
AEO and GEO: Answer engine optimisation and generative engine optimisation: industry names for the work of being named and used as a source inside AI assistants' answers, rather than ranked in a list of links.

The precise terms on the method pages

Used on /methodology, /proof, /data and /pre-registration, where precision is the job.

Engine: An AI assistant as the method pages refer to it, called through its provider's API. The engines a measurement asks are its measurement set, and every score records which they were. See AI assistant.
Probe: One question put to one engine once, and the answer that comes back. A probe that fails, through a timeout or a rate limit, is missing data and leaves the denominator. See Denominator (answers sampled).
Brand-blind: No frozen question contains the client's name or domain. A question that names the brand guarantees a mention and measures nothing.
Frozen question set: The fixed list of buyer questions a measurement asks, frozen the day it is approved and never edited, so every later measurement asks the same ones. It is identified by its fingerprint. Coverage grows by freezing a second set, never by editing the first. See The same questions and Fingerprint (hash).
Fingerprint (hash): A short code computed from the exact text of a frozen question set. Change one word and the code changes, so two measurements with the same hash asked the same questions. Measurements with different hashes are never compared. See The same questions.
Grounded: The engine answered with its provider's web-search tool on. Ungrounded means answering from the model's memory alone. The two are different tests, so every score records its mode, and a change of mode starts a new baseline. See With web search switched on.
Analyst (grader): The separate model that reads each stored answer and records which businesses were named, in what order, and which sources the answer cites. A graded answer is one it has read. It is never the model that produced the answer, and every score records which model graded it.
Answer presence (named rate): The share of graded answers in which the brand is named in the answer text, taken as the mean of the per-question rates over the frozen set. See Named.
Citation presence (cited rate): The share of graded answers whose sources include a page on the brand's own domain, averaged the same way. Measurable only on grounded answers. See Used as a source.
Share of voice: A brand's share of all brand mentions in the graded answers, counting every business any engine named. It counts names, so it says nothing about which websites the answers were built from.
Visibility Score: A 0 to 100 score that blends how often the brand is named, how early in the answer, and whether it is recommended. At the same named rate, named first scores higher than named fifth.
Denominator (answers sampled): The number of graded answers a rate is divided by. A failed answer, from an engine that timed out or a grader that could not read it, leaves the denominator; it is never counted as a miss.
Headline, pooled and per-question rates: The headline rate is the mean of the per-question rates, one for each question on each engine, and it is the rate we publish. A pooled rate (all hits over all answers) drifts from that mean when probes fail unevenly, so we do not publish one. The headline rate carries a 90% band when every question has more than one graded answer on each engine; one question on one engine is shown as counts, such as 9 of 14, without a band.
90% band (90% interval): The range the true rate plausibly sits in, given how many answers were sampled, also called a 90% confidence interval. The headline band is a variance sum over the per-question rates (meanRateInterval in the code), centred on their mean, with the Wilson interval over all answers as the fallback when there are no hits or no misses. See How sure we are and Wilson interval.
Wilson interval: A formula for a rate's band that stays sensible near 0% and 100%, where the textbook normal approximation shrinks to zero width and claims a certainty the data cannot support. See 90% band (90% interval).
Paired difference (verdict): How a before and after is judged: the same frozen questions asked again and compared question by question, each question its own control. The verdict is the 90% interval on that paired difference. It reads moved up or moved down only when the interval excludes zero, and too close to call otherwise. See How sure we are.
Baseline: The first measurement that later ones are compared with. When the instrument changes, the next measurement becomes a new baseline instead of being compared across the change.
Instrument: Everything that decides how the answers were produced and read: the question set, the engines, grounding, the engine profile, the analyst, and the collection method (our own engine calls, the rig, or a retired third-party feed). A change to the questions, engines, grounding, engine profile or collection method starts a new baseline; a change of analyst is handled by re-grading the stored answers. See Baseline.
Engine profile: How much each grounded call is allowed to spend (how many searches, how much reasoning). Recorded on every answer and every score; a change starts a new baseline.
Difference-in-differences (DiD): The client's change minus the change in a control measured in the same answers at the same time. Whatever moved the whole category moves the control too, so what remains, the attributable lift, can be credited to the work. The control is written down, with its date, before the outcome is known. See Control set.
Control set: For a brand that needs to be named, the rivals named at scoping; for a comparison or affiliate site, the other domains cited on the same questions, because the brands an engine names there are the products being compared, not competitors.
Pre-registration: What we commit to before optimising our own properties: the questions, the primary metric the target is set on, the target that counts as real, the window (the dates measured) and the analysis, locked and published before the treatment, the change being tested, ships. Once locked the text cannot change; anything later is appended as a dated amendment.

The full method, with the maths: /methodology. Applied to our own properties, live: /proof.