Definitions

The AI search vocabulary, defined once.

Every term below gets used loosely somewhere on the internet. These are the meanings I use in reports, so there is never any ambiguity about what a number refers to — and each one says how it is actually measured, which is where most definitions stop short.

Generative engine optimization (GEO)

The practice of getting a brand named inside AI-generated answers — the paragraph ChatGPT, Perplexity, Google AI Overviews or Copilot returns instead of a list of links. Distinct from SEO in that the target is a sentence in a paragraph rather than a position in a list.

Why it matters
It is the discipline everything else on this page belongs to. Also called answer engine optimization or AI search optimization; the three names describe the same job.
How it is measured
By citation share and recommendation rate across a defined prompt set.

Answer engine optimization (AEO)

A synonym for GEO, more common in B2B marketing teams. Where the two are distinguished at all, AEO usually implies the narrower job of formatting content so a machine can lift a direct answer from it.

Why it matters
Mostly a vocabulary difference. If a vendor claims a methodological distinction between GEO and AEO, ask them to state it in one sentence.
How it is measured
Identically to GEO.

Citation share

The percentage of prompts in a defined set where a brand is named or linked in the generated answer.

Why it matters
The primary scoreboard metric, and the one most tools report. On its own it is misleading — see recommendation rate.
How it is measured
Prompts cited ÷ prompts tracked, per engine, weighted by the revenue behind each prompt.

Recommendation rate

The share of prompts where a brand is named as the answer rather than used as a footnote.

Why it matters
The more valuable of the two numbers and the one that correlates with pipeline. Ahrefs found 43% of answers citing a page never mentioned that page's brand — a gap invisible to citation share alone.
How it is measured
Prompts recommended ÷ prompts tracked. Tracked separately from citation, and joined to the CRM.

Prompt set

A fixed list of 150–400 real buyer questions, each weighted by the revenue of the customers who ask it.

Why it matters
The GEO equivalent of a keyword list, and the only stable way to measure something that answers differently every time. Narrow questions outperform broad ones badly: 66.4% mention rate against 15.8% in the Ahrefs data.
How it is measured
Built from support tickets, reviews and won/lost CRM notes, clustered by a model, then weighted by revenue.

Entity

The machine-readable identity of a business: its name, category, location, offering and relationships, as expressed consistently across your site, your schema and every third-party profile.

Why it matters
Models hedge or miscategorise you while those sources disagree. Reconciling them is consistently the cheapest movement available and usually shows within three weeks.
How it is measured
By counting how many distinct category descriptions are in circulation across your own properties and third-party profiles.

Corroboration

Independent sources describing a brand the same way.

Why it matters
Assistants weight consensus above self-description. 94% of an established brand's new AI mentions came from third-party pages in the Ahrefs study, against 6% from its own content.
How it is measured
By the count of third-party pages that state your category claim, and where those pages rank.

Retrieval pool

The set of pages an assistant actually searches and quotes for a given question.

Why it matters
The pool and page one are largely different sets: roughly 90% of ChatGPT citations come from pages ranking 21 or lower, and AI Overview links overlap the organic top ten only 20–26% of the time.
How it is measured
By recording the ranking position of every page cited across the prompt set.

Extractable answer

A passage written so a retriever can lift it whole and have it still make sense: the direct claim in the first 100 words, specific numbers, a named source, and no dependence on the paragraph before it.

Why it matters
Retrievers lift passages, not pages. A paragraph that only makes sense in sequence cannot be quoted, and a claim sitting apart from its category term gets separated from it.
How it is measured
By reading the exact passage an assistant quoted and checking whether the claim survives on its own.

Citation volatility

How inconsistently a cited page reappears for the same prompt.

Why it matters
Measured at roughly one day in three in the Ahrefs experiment. It is why a monthly manual check has about a one-in-three chance of describing your real position.
How it is measured
Days appearing ÷ eligible days, between a page's first and last appearance for a prompt.

AI Overview

Google's generated answer above the organic results.

Why it matters
It appeared on 6.49% of searches in January 2025 and 24.61% by July, and cuts clickthrough for the top-ranking page by about 34.5% when present. Your ranking can hold still while the traffic behind it drains.
How it is measured
By AI Overview prevalence across your tracked keyword set, and clickthrough on affected terms.

llms.txt

A plain-text file at the root of a site listing its most useful pages for language models, in the spirit of robots.txt.

Why it matters
Support is inconsistent and it is not a ranking factor. Genuinely low priority — worth an hour, after the items above, not before.
How it is measured
Not meaningfully measurable in isolation. Treat it as hygiene rather than a lever.

Want these numbers for your own site?

The scorecard takes two minutes, or send your domain and I will run an automated read and email you a ranked fix list.