GEO Metrics Glossary
Every metric in generative engine optimization, defined the way we actually compute it in production: one-sentence definition, exact formula, worked example. Formulas use pooled window math and count only completed model runs — failed runs are excluded from every denominator.
GEO (Generative Engine Optimization)
GEO is the practice of improving how AI assistants and answer engines present a brand — whether it appears in answers, how it is framed, which sources are cited, and whether the claims are true.
Notes: Where SEO optimizes for ranked links on a results page, GEO optimizes for being named inside the generated answer itself.
Example: A brand that moves its ChatGPT mention rate from 20% to 45% on category prompts has done GEO, even though no search ranking changed.
Prompt Set
A prompt set is the fixed collection of buyer-style questions run against each AI model on a schedule to measure brand visibility consistently over time.
Notes: A balanced set splits roughly 35% category prompts, 35% comparison prompts and 30% use-case prompts, and never contains the brand name itself.
Example: "Best email platform for paid newsletters" belongs in a prompt set; "Is Beehiiv good?" does not — it leaks the brand into the question.
Mention Rate
Mention rate is the percentage of completed AI model runs in which the brand is named at least once.
mentionRate = mentioned runs ÷ completed runs × 100
Notes: The denominator counts only completed runs — runs where the model errored or never finished are excluded, because they are neither a mention nor a miss.
Example: If 40 of 80 completed DeepSeek runs name your brand, your DeepSeek mention rate is 50%.
Citation Rate
Citation rate is the percentage of your brand mentions that are backed by a cited source attributed to your brand.
citationRate = cited mentions ÷ brand mentions × 100
Notes: A mention only counts as cited when the citation set attributed to the brand is non-empty — arbitrary URLs elsewhere in the answer do not count. With zero mentions the metric is null (no signal), not 0%.
Example: If 12 of your 40 mentions carry a brand-attributed source link, your citation rate is 30%.
Citation Share of Voice (Citation SoV)
Citation SoV is the percentage of all brand-attributed cited URLs in a scan that are attributed to your brand rather than competitors.
citationSov = brand cited URLs ÷ (brand + competitor cited URLs) × 100
Notes: Mention SoV measures who gets named; citation SoV measures who gets sourced. The gap between the two is a strategy signal — see mention rate vs citation rate below.
Example: Holding 26% of mentions but only 7% of citations means models name you from memory while grounding answers in competitors’ pages.
Mention Rate vs Citation Rate
Mention rate measures how often a brand is named in AI answers; citation rate measures how often those namings are backed by a source — the two diverge, and the divergence is diagnostic.
Notes: Mention-strong / citation-weak means the model remembers you but cannot find authoritative pages to cite — invest in the source layer. Citation-strong / mention-weak means you are used as a reference but not surfaced as a vendor — invest in positioning content.
Example: A brand with a 50% mention rate and 15% citation rate wins recall but loses grounding; its mentions can vanish with the next model update.
Sentiment Score
Sentiment score is the percentage of a brand’s mentions classified as positive (or negative), measuring how the brand is framed when it does appear.
sentimentPosRate = positive mentions ÷ brand mentions × 100 sentimentNegRate = negative mentions ÷ brand mentions × 100
Notes: Each mention is classified positive, neutral or negative. With zero mentions the score is null — "no signal" is deliberately distinct from 0%.
Example: If 24 of 40 mentions are positive and 2 negative, sentiment is 60% positive / 5% negative, with the rest neutral.
Answer Accuracy / Hallucination Rate
Answer accuracy is the percentage of a brand’s mentions not flagged as potential hallucinations; hallucination rate is its complement.
answerAccuracy = (brand mentions − flagged mentions) ÷ brand mentions × 100 hallucinationRate = 100 − answerAccuracy
Notes: Flags come from cross-model voting: when models disagree on a prompt, an LLM judge rates the factual conflict; at medium/high severity every mention under that prompt is flagged — a deliberately conservative rule. Computed pooled over the window, not averaged per day.
Example: With 40 mentions of which 7 sit under conflict-flagged prompts, answer accuracy is 82.5% and hallucination rate 17.5%.
Cross-LLM Reach
Cross-LLM reach is the percentage of active AI models that mentioned the brand at least once in the measurement window.
crossLlmReach = models with ≥1 mention ÷ active models × 100
Notes: Measures breadth where mention rate measures depth: a brand can have a high mention rate on one model and zero presence on the other six.
Example: Mentioned in 4 of 7 tracked models during the window → cross-LLM reach is 57%.
Query Fan-out
Query fan-out is the hidden step where an AI assistant expands a user’s question into multiple sub-queries before searching, so the final answer is stitched from sources the user never asked for directly.
Notes: Individual sub-queries are unstable across runs (research pegs stability around 27%), but the themes recur — optimize theme coverage, not specific sub-queries.
Example: "Best CRM for startups" may fan out into pricing, integrations and alternatives sub-queries; a brand missing a pricing page loses every fan-out that fires one.
See these metrics live for your brand
Run a free scan and get mention rate, SoV, sentiment and citations across 7 AI models — computed exactly as defined above.
Start free