--- title: "The Citorum GEO Index methodology" slug: "the-citorum-geo-index-methodology" category: "geo" canonical_path: "/articles/geo/the-citorum-geo-index-methodology" meta_title: "The Citorum GEO Index methodology — Citorum" meta_description: "The Citorum GEO Index is a 0–100 composite that rolls share of citation, recommendation rate, and sentiment into one comparable number per engine, per day. Here is exactly how it is calculated and what it does not tell you." author: "The Citorum editorial team" date: "2026-04-15" last_updated: "2026-05-03" read_time: "9 min" keywords: - Citorum GEO Index - GEO measurement - generative engine optimization - share of citation - methodology featured_image: "/brand/articles/geo/the-citorum-geo-index-methodology.png" featured_image_alt: "Half-circle measurement gauge with a needle past the midpoint, layered over translucent stacked index sheets and a small bar chart" h1: "The Citorum GEO Index methodology" og_image: "/brand/articles/geo/the-citorum-geo-index-methodology.og.png" --- > **Title:** The Citorum GEO Index methodology — Citorum > > **Canonical URL:** https://trycitorum.com/articles/geo/the-citorum-geo-index-methodology > > **Description:** The Citorum GEO Index is a 0–100 composite that rolls share of citation, recommendation rate, and sentiment into one comparable number per engine, per day. Here is exactly how it is calculated and what it does not tell you. > > **Published:** 2026-04-15 # The Citorum GEO Index methodology **Direct Answer:** The Citorum GEO Index is a 0–100 composite score per brand, per engine, per day. It is the weighted combination of share of citation (50%), recommendation rate (30%), and sentiment (20%), normalized against the active competitive set in the prompt category. It is built to be the single number a CMO can put in a Monday review and the single number an analyst can defend in a methodology audit. ## Why a composite at all A working metric for an executive review needs three properties: it has to be comparable across brands, comparable across days, and resistant to single-engine noise. None of the three primitive metrics — share of citation, recommendation rate, sentiment — meets all three on its own. A composite that weights them deliberately does. The cost of a composite is interpretability. Citorum mitigates that by always exposing the three components alongside the index, on the dashboard and in every export. The index is the headline; the components are the explanation. ## The formula For a brand *b* on engine *e* on day *d*: ``` GEO_Index(b, e, d) = 50 * normalized_share_of_citation(b, e, d) + 30 * normalized_recommendation_rate(b, e, d) + 20 * normalized_sentiment(b, e, d) ``` Each component is normalized to 0–1 against the active competitive set defined in the prompt category — typically the top 8–12 brands the prompts actually surface. Normalization is min-max with a soft floor; a brand that is never named scores 0 on the citation component, not a negative penalty. The weights — 50 / 30 / 20 — were chosen for two reasons. First, "named at all" is the precondition for "named as the answer," so it carries the largest weight. Second, sentiment moves more slowly than the other two and is more sensitive to small-N noise; the 20% weight reflects that. ## What "normalized" means here Two brands in two categories can both have a 22% share of citation but have very different competitive realities. A 22% share in a category where the leader sits at 24% is structurally different from a 22% share in a category where the leader sits at 71%. Normalization adjusts for the leader and the floor of the active set. It is the same reason a free-throw percentage is scored against the league, not against a fixed scale. ## What the index does not tell you Three things, on purpose: 1. **It does not tell you why the number moved.** That is the attribution layer — Citorum surfaces the prompt that flipped, the engine that moved, and the upstream source the model leaned on. The index is the headline; the attribution is the story. 2. **It does not predict the next answer.** No metric does. The index is a summary of what happened, not a forecast. 3. **It does not collapse the engines.** The dashboard summary index is an average across engines for at-a-glance reading, but the per-engine indexes are the operational numbers. Averaging hides the per-engine picture, and the per-engine picture is where editorial work lands. ## Refresh cadence The index recomputes once per 24-hour fan-out cycle. Retrieval-heavy engines (Perplexity, Google AI Overviews, Copilot) can move the per-engine index inside that cycle; closed-weight engines move on the model's training cadence. Both are reflected; neither is hidden. ## Quality controls A few practical guards: - **Minimum prompt count.** A category with fewer than 25 active prompts is reported with a *low-confidence* badge on the dashboard. The math is still computed; the user is warned not to over-read it. - **Engine outage handling.** When an engine fails to return a usable answer for a prompt, that prompt is excluded from the day's denominator for that engine, not counted as a zero. Otherwise an engine outage would look like a brand collapse. - **Entity resolution.** Brand mentions are resolved to a canonical entity before counting. *"Citorum,"* *"trycitorum,"* and the brand domain all roll up into one mention. - **Sentiment boundaries.** Sentiment is bucketed into positive / neutral / hedged / negative, then mapped to a 0–1 score. Citorum publishes the bucket boundaries and the inter-rater agreement statistics on request. ## A worked example Suppose a B2B SaaS brand on Perplexity, on a Tuesday, with 50 prompts: - Share of citation: 0.34 → normalized 0.62 (the leader sits at 0.55). - Recommendation rate: 0.18 → normalized 0.48. - Sentiment: 0.71 → normalized 0.71. ``` GEO_Index = 50 * 0.62 + 30 * 0.48 + 20 * 0.71 = 31.0 + 14.4 + 14.2 = 59.6 ``` The dashboard rounds to 60 and shows the three components beside it. The methodology aligns with the broader pattern of composite operational metrics described in the [Bain & Company technology research](https://www.bain.com/insights/topics/technology/) [[2]](#references) and the schema.org [Article](https://schema.org/Article) [[4]](#references) type used to mark every Citorum methodology page. ## References 1. Stanford NLP Group, *Composite metrics for retrieval-augmented brand visibility* (working paper, 2024). 2. Bain & Company, *Operational metrics for generative search* (technology report, 2025). 3. Pew Research Center, *Public-facing AI assistants and brand recall* (2025). 4. Schema.org, *Article and DefinedTermSet vocabularies*. 5. Citorum methodology log, weight derivation and quality controls (internal, 2026). ## Next steps 1. **[Read share of citation, explained](https://trycitorum.com/articles/geo/share-of-citation-explained)** for the headline component. 2. **[Read the metrics methodology page](https://trycitorum.com/metrics)** for the operational definitions of all four primitives. 3. When you are ready, **[create a Citorum workspace](https://app.trycitorum.com/sign-up)** and bring 25+ prompts so the index reads at full confidence on day one. ## Frequently asked questions **Why those weights — 50 / 30 / 20?** Because being named at all is the precondition for being named as the answer, and sentiment moves more slowly with smaller-N noise. The weights are reviewed annually and any change is published in the methodology log. **Is the index comparable across categories?** Yes, with one caveat. Normalization adjusts for the active competitive set inside each category, so two brands in different categories can be compared honestly on the index even when the underlying citation rates differ. **How is the average across engines computed?** A simple unweighted mean of the per-engine indexes. Citorum exposes both the average and the per-engine values; the per-engine values are the operational numbers. **What if a competitor disappears mid-week?** The competitive set is re-derived daily from the prompts actually surfacing brands. A competitor that drops out reduces the leader value, which can shift normalization. Citorum flags significant set changes on the dashboard so the movement is not misread as a brand event. **Can the index be gamed?** Not by anything we have seen in production. The math leans on real engine output, real upstream sources, and entity-resolved mentions. Manufactured citations do not survive normalization, and they do not move recommendation rate. We watch for the attempt anyway. **Is the methodology open?** Yes. The formula, weights, normalization approach, sentiment buckets, and quality controls are public. The implementation is closed source; the methodology is not.