GEO glossary

Generative Engine Optimization borrows words from SEO, statistics, and machine learning, then uses them slightly differently. This glossary defines 16 terms the way they are actually used when measuring AI visibility — each definition is self-contained, linkable, and consistent with our published methodology.

Generative Engine Optimization (GEO)

The practice of making a brand appear — and get cited — in AI-generated answers from engines like ChatGPT, Gemini, Claude, and Grok. Unlike SEO, which optimizes for a ranked list of links, GEO optimizes for inclusion in a synthesized answer: the outcome is measured as a rate across repeated runs, not a position. What is GEO? →

AI visibility

How often and how prominently a brand shows up in AI answers to the questions its buyers actually ask. Because the same prompt can produce different answers on different runs, visibility is a sampled statistic — a percentage of runs — rather than a fixed rank.

AI Visibility Score

A 0–100 composite measuring how present, prominent, cited, and positively framed a brand is in AI answers. Geoptimizer's version is published in full: EngineScore = 100 × (0.35 · mention rate + 0.25 · citation rate + 0.20 · prominence + 0.20 · sentiment), averaged across engines and computed over a 7-day rolling window. The full methodology →

Mention rate

The share of sampled answers that name the brand at all. A mention requires no link — the engine simply talks about the brand. Mention gaps usually point to off-site positioning problems: the sources engines read don't associate the brand with the category.

Citation rate

The share of sampled answers that cite the brand's own domain as a source. Citations and mentions move independently: an engine can recommend a brand without citing its site, or cite a page while naming a competitor. Tracking them separately is what makes the gap diagnosable. How citations work, engine by engine →

Ghost citation

A citation of a brand's page in an answer that never mentions the brand by name. The engine used the page as evidence but attributed the recommendation elsewhere. The inverse — mentioned but never cited — is equally common; the two mismatches call for different fixes.

Prominence

Where in an answer a brand first appears. Being the first brand named is worth more than being the ninth: readers and downstream models both weight early positions. Geoptimizer scores prominence as 1/√rank of the first mention, so the value decays smoothly rather than falling off a cliff.

Sentiment (in AI answers)

How an answer frames a mentioned brand: promoted, neutral, caveated, or negative. Most brand mentions are neutral-to-positive, so sentiment mainly matters as an alarm — a single engine flipping to caveated ('users report hidden fees') can silently reprice every answer it gives about you.

Share of voice (SoV)

A brand's mentions as a share of all brand mentions detected across sampled answers. With an open denominator — every brand the AI names counts, not just a preset competitor list — unknown competitors surface automatically the first time an engine recommends them.

Buyer-intent prompt

A question a real buyer would ask an AI assistant while choosing — 'best X for Y', 'alternatives to Z', 'is X worth it for a small team'. Buyer-intent prompts are the unit of GEO measurement; brand-name prompts ('what is Acme?') mostly measure whether the engine can read your homepage. Choosing prompts to track →

Prompt tracking

Re-running a fixed set of buyer-intent prompts against AI engines on a schedule and recording who gets mentioned and cited. Keeping the prompt set frozen between periods is what makes the trend honest — change the questions and you change the denominator.

AI crawler

A bot an AI company uses to fetch web content. They come in distinct categories with separate user agents: training crawlers (GPTBot, ClaudeBot), search/retrieval crawlers (OAI-SearchBot, Claude-SearchBot), and user-triggered fetchers (ChatGPT-User, Claude-User). Blocking a retrieval crawler removes you from that engine's cited answers; blocking a training crawler does not. Check your crawler access →

llms.txt

A proposed plain-markdown file at /llms.txt that gives AI systems a curated map of a site's most important pages. It is cheap to add and harmless, but unproven: no major engine has committed to reading it, so it belongs at the end of a technical-GEO checklist, after crawler access and server-side rendering. The 20-minute setup →

Grounding (web search retrieval)

An engine's step of running live web searches and reading the results before answering, instead of relying only on what it memorized in training. Grounded answers carry citations and can change week to week with the underlying sources; ungrounded answers reflect the model's training snapshot.

Rolling window

Scoring over the last N days of samples (Geoptimizer uses 7) rather than a single run. Individual AI answers are noisy — the same prompt can name different brands on consecutive runs — so rates over a window are the signal, and any single scan is just one sample from the distribution.

Confidence band

The uncertainty range around a sampled score, driven by how many runs it aggregates. A 3-point move inside a ±9-point band is not a finding. Reading the band before the point estimate is the difference between reacting to noise and reacting to change. Why tools disagree →