Definitions

SEO, AEO, GEO, AIO, LLMO — defined

How AI search visibility is measured, term by term. One page, one canonical definition each, written plainly enough to quote and specific enough to check against the run a number came from.

The figures here are our own measurements, taken by a method we define and publish — a different source from the numbers the platforms issue themselves. That is why the method and the judging rules come first on this page.

See how the grid works Check the source of every number

  1. The problem

    Every vendor means something different by the same words.

    If one quote counts a mention and the next counts a citation, putting them side by side tells you nothing.

  2. How we fix it

    Each term here is pinned to one sentence.

    These are the words that appear in your contract and in your report. Every term has its own address, so you can quote it back to us.

  3. What you get

    You can check what any number in the report actually counted.

    Trace the wording next to a figure and you find the rule behind it. It also gives you one standard to hold other vendors to.

The cost of a loose word. If two vendors define citation differently, the higher number wins the decision and the loss only shows up a quarter later when neither figure reproduces. CiteAngle reports run on these exact definitions, sealed with every reading, which is why a number of ours can be checked against the word it used.

There is a second failure, quieter than the first. Without a fixed vocabulary a team cannot brief its own agency, so work is commissioned against one meaning and delivered against another, and the cost lands as a wasted quarter that nobody argues about.

Why the words have to be pinned first

Two vendors can report the same brand at 41.5% and 9.6% in the same month, and both be telling the truth — we measured both figures on one dataset. One counted any appearance of your name. The other counted only answers that cited a page of yours. Without a fixed definition you cannot tell which you bought, so the higher number wins the decision and the loss surfaces a quarter later. Nothing in the meeting will reveal that. The numbers look comparable, the decision gets made on the higher one, and the disagreement only surfaces a quarter later when neither figure reproduces. Undefined vocabulary is not a documentation problem here. It is where the money goes wrong.

Each entry below fixes one term to one meaning, with the unit and the denominator attached, so a number carrying that word can be checked instead of believed. Mention and citation are kept apart, because they are different events with different causes. A planned observation we could not read gets its own word, so it can be counted and kept in the denominator. And a rate arrives with the interval around it, since a percentage without one is a guess wearing a decimal point.

These are the definitions our own reports run on — the same ones behind the 50 questions, 21 engines and 7 runs of the widest panel — and every public figure we quote is registered against them in the claims registry. Take any of them into a conversation with another vendor. If a number cannot survive being asked which definition it used, you have learned something worth knowing while you can still act on it.

Why take the definitions from us

Definitions are cheap to assert and expensive to keep. The wording here is the wording in the method we publish, in the registry where every public figure carries its source and its date, and in every report we put out. Nothing gets a looser meaning on the way from a measurement into an ad.

That is what makes a definition here checkable rather than merely stated: quote a sentence, then follow it to the measurement method and to the entry in the public claims registry that holds the matching figure, its source and its date.

What is AI search visibility?

AI search visibility is the degree to which a brand or its content is named or cited inside the answers generated by AI search systems, rather than in the ranked links beside them.

CiteAngle measures it by putting a fixed query set to the same answer surfaces, repeatedly, and recording for every observation whether the brand appears in the answer body and whether a page from its domain is offered as a source. Rates come back with an interval attached, and the conditions behind them are published in the methodology.

Which surfaces the questions go to What an engagement changes first Where each published figure came from

How do SEO, AEO, GEO and AIO differ?

They name four different targets on the same screen: ranked results, direct answers, generated answers, and the AI layer a search engine places above its own results.

The work overlaps and the measurement does not. A page can rank without ever being cited, and it can be cited by an assistant that never sends a visitor. LLMO belongs in the same table because assistants increasingly answer without a search page in front of them.

TermWhat it targetsWhat gets countedWhere the work lands
SEO
Search engine optimization		TargetsRanked organic resultsCountedIndexation, position, organic entriesWhere the work landsTechnical health, content, entity consistency, internal links
AEO
Answer engine optimization		TargetsSurfaces that resolve a question directlyCountedWhether the answer is adopted and which sources travel with itWhere the work landsAnswers that finish the question, evidence, comparisons, entity consistency
GEO
Generative engine optimization		TargetsGenerated answersCountedDirect citations and brand mentions inside the generated textWhere the work landsOriginal data, first-hand expertise, source material worth quoting
AIO
AI-driven search as a category		TargetsThe AI layer above search resultsCountedAppearance, mention, citation and the traffic that followsWhere the work landsSearch fundamentals plus evidence a layer can attribute
LLMO
LLM optimization		TargetsAssistants answering without a search pageCountedWhether the facts are retrieved, attributed and restated correctlyWhere the work landsMachine-readable facts, consistent naming, sources an assistant can resolve

Five lenses, one grid The work that sits under each lens Figures filed lens by lens

What is AEO?

Answer engine optimization is the practice of shaping content so that a surface which answers a question directly adopts that answer and links the page it came from.

Rank is not the unit here. The question is which answer gets used, and which sources are carried along when it is.

Where answer adoption gets counted Rewriting a page so it finishes the question What we can show for answer surfaces

What is GEO?

Generative engine optimization is the practice of building source material that a generative answer can quote, so the brand is cited and named inside the generated text itself.

Our reports keep those two events apart. A citation and a mention are counted separately, on their own denominators, because they move for different reasons.

How a citation is told from a mention Building material worth quoting Citation figures with their sources

What does AIO refer to?

AIO covers visibility and traffic across AI-driven search as a whole, including the AI layer a search engine places above its own results, and it is kept on a denominator separate from that layer's own metrics.

The abbreviation gets used for two things — AI Optimization and AI Overview — so we keep the reporting fields apart. A word should never quietly change what a number is divided by.

Google answers in two places, not one. AI Overviews sit above the classic results page; AI Mode is a separate screen where people keep asking follow-up questions. The same question can surface different brands in each, so we count them separately and never fold the two into a single number.

Counting one Google query as two screens Shaping pages for the layer above results Readings from the AI layer

What is LLMO?

LLM optimization is the practice of making a brand's facts easy for a language model to retrieve, attribute and restate correctly, whether or not a search page sits in front of the answer.

GEO and LLMO overlap in the work and differ in the measured event. GEO asks whether the generated answer cited or named you; LLMO asks whether what it said about you was accurate and traceable to something you own.

Reading assistants on the same grid Straightening facts, names and sources What we can show for assistant answers

How do mention and citation differ?

A mention is the brand name appearing in the body of an answer; a citation is the answer presenting a URL from that brand's domain as a source.

Citations resolve deterministically against the domain. Mentions need entity resolution, because a name can belong to more than one company, and counting a namesake as you is how a report ends up flattering. Blend them into one score and neither event can tell you what moved.

The rule that keeps the two events apart Getting the domain into the source list Both events published on their own lines

What is an unobserved AI answer?

An unobserved AI answer is a planned observation where the request and the capture both worked, but the surface produced no AI answer for that query at that moment.

There is nothing in such an observation to judge, so it stays out of the citation and mention denominators and is published as its own count instead. Full-grid accounting means every planned observation is accounted for in one of three states rather than only the ones that came back clean. That is what stops a thin run from reading like a strong result.

How the three observation states are separated Cutting down the observations that come back empty Unobserved answers printed as a count

What is answer presence rate?

Answer presence rate is the share of queries for which an AI answer surface appeared at all, and it moves with platform policy rather than with anything a brand does.

No answer means no denominator for citation or mention, so presence belongs next to the numbers as a stated condition. We print it as context and read it as a signal about the surface, not as a scoreboard for the brand.

The surfaces presence is read on Widening the questions that get an answer Presence stated beside the rates

What is answer variance?

Answer variance is how far the citation and mention results move when the same query is run again under the same conditions.

One reading cannot tell you whether a difference is real movement or the spread you would have seen anyway. Repeats exist for exactly that reason, and a report shows the spread instead of a single flattering number.

Why the same question runs seven times Work that steadies a moving result Figures that carry their spread

What is a confidence interval?

A confidence interval states the range of values consistent with the observed sample, given how many usable observations there were.

Printed beside a citation rate, it makes the difference between a stable finding and a small-sample artifact visible on the page. When two intervals overlap, we write that down and say what would settle it, which is usually another round on the same grid. The interval we publish is a Wilson 95% confidence interval, taken over the seven repeated runs a paid diagnostic makes for each prompt, on the share of runs that could be judged. The formula is named on purpose: a different one over the same counts moves the edges of the range.

How wide an interval is allowed to be Rounds that narrow the interval Intervals printed next to the numbers

What is a verifiable measurement?

A measurement is verifiable when its query set, surfaces, repeat count and measurement dates are published, so the same conditions can be run again and the result checked by someone else.

A figure built on conditions nobody can see is a figure nobody can check, and that is a thin thing to sign a contract on. Ours are listed in the public claims registry, each with the kind of source behind it and the date it was established.

The conditions a rerun needs Running the same conditions again Figures anyone can re-check

What is a sealed hash ledger?

A sealed hash ledger records a fingerprint of the measurement setup and of its results at the moment of measurement, so later edits and backdating become detectable.

Our own canonical runs are sealed this way. The identifier of the run, the configuration fingerprint and the query-set version are all fixed at the moment of measurement, not when the report gets written, so a single changed character shows up. The t0 run behind our published figures holds every observation it planned, and each one was sealed on the day it was measured. Not a sample of the run. The run.

Why measure the same question more than once? Because the answer is not fixed. Ask an AI assistant the same thing on Tuesday and on Wednesday and the brands it names can change, along with the order they arrive in. A single reading tells you what happened once. We ask 7 times and keep the whole spread, so what you get is a range you can act on rather than a snapshot that may not survive the week.

And we read the answer everywhere it can carry your name. These are the places a sealed run covers in the United States:

What gets fingerprinted, and when Sealing before the site is touched Numbers drawn from sealed runs

What is a t0 baseline?

A t0 baseline is the starting measurement, sealed before any change is made, and it is the reference every later comparison is judged against.

Once a change goes live, the state before it cannot be recreated. A baseline reassembled afterwards is a weaker object than one sealed on the day, and sealing is cheap while the measurement is still running.

Fixing the starting point in month one What follows once the baseline is sealed Baselines kept on the record

What belongs in a measurement receipt?

A measurement receipt is the record that ties a published figure back to the queries, surfaces, repeat count and dates behind it.

It is what lets a reader walk a number backwards to the run that produced it. Our published measurements carry those fields in the body of the text, where a buyer can act on them, rather than in a footnote written to be skipped.

The fields a receipt has to carry Receipts that come with every round Published figures with their receipts

Same wording, same denominators, everywhere we publish. Read how the grid is built and why the repeats are there in the measurement methodology, or open the claims registry and check any figure we publish against the source and date registered with it.

Each term above is one we publish a number against. The audits that produce those numbers are listed with their prices. See the audits and what they cost →

Each term on its own page

Ready to turn US visibility evidence into a market plan? Tell us the US buyer, target state and decision date.

Request a written proposal