Which GEO metrics matter, and how should you measure them?

Generative engine optimization (GEO) measurement starts by defining what each number means. AI share of voice can refer to a brand’s portion of coded recommendations in a defined AI-answer panel, but that competitive share is only one part of a useful scorecard. Large language models (LLMs) and other AI search systems can produce answers in which a mention is not necessarily a recommendation, a citation does not necessarily support the answer, and a referral session does not prove that AI visibility caused a sale. Measure recommendations, mentions, citations, referral traffic and CRM outcomes separately, with a defined denominator for each. Then repeat a fixed prompt panel across engines and report uncertainty rather than treating a single result as a ranking.

Which GEO metrics are useful for B2B SaaS?

Recommendation rate, citation rate and citation relevance, qualified assistant-referral sessions, and CRM outcomes answer different questions. A fixed prompt panel with repeated runs makes answer-level rates easier to interpret, while pipeline contribution and incrementality require separate attribution or experimental methods. No single GEO metric captures the entire path from an AI answer to revenue.

A prompt-run cell is one defined prompt submitted to one specified engine/model configuration for one collection run.

Keep each layer separate in reporting. Recommendation and neutral or negative mention should be coded distinctly; citation rate, source ownership, citation relevance and factual support should not be collapsed into a single “visibility” score.

For referral measurement, disclose the referrers included and the GA4 channel rule used. Google Analytics currently defines an AI Assistant channel for sources such as ChatGPT and Gemini. It excludes Google’s AI Overviews and AI Mode from that channel, placing them in Organic Search instead. These channel rules may evolve and do not capture every untracked AI exposure Google Analytics’ default-channel documentation.

For business outcomes, report CRM-qualified opportunities, win rate or gross profit using explicit stage definitions. Cost per incremental opportunity is an outcome measure only when incrementality has been estimated with an appropriate experimental design. Google Analytics attribution paths describe observed paths and credit allocation, not causal impact Google Analytics’ attribution-path documentation. Google Ads describes Conversion Lift as an experimental approach for Google Ads campaigns. Check its eligibility and applicability; it is not a direct experiment on organic AI-answer visibility Google Ads’ Conversion Lift documentation.

What denominator belongs under a recommendation or citation rate?

For a per-prompt-run recommendation rate, divide prompt-run cells with an explicit recommendation for the brand by all eligible prompt-run cells. For citation rate, divide cells with a qualifying citation by all eligible cells. Define the prompt set, engine/model, run count and collection period; one answer is not a reliable trend by itself.

For example, in a hypothetical sample of 100 eligible cells, if a brand is recommended in 18, its observed recommendation rate is 18%. This is an illustration of the arithmetic, not a research finding or a benchmark. If 24 cells cite a source from the brand’s site, the citation rate is 24%, but that result says nothing by itself about whether those citations support the answer.

Recommendation share uses a different denominator:

recommendation share = brand recommendations / all coded brand recommendations

The denominator includes all coded brand recommendations across the specified engine, prompt panel and period. A single answer may recommend more than one brand, so count recommendations according to the predeclared coding rules.

Prompt-run cells with no brand recommendation remain in the denominator for per-cell recommendation rate. They contribute no recommendations to recommendation share. This distinction prevents a competitive share from being mistaken for the chance that a brand appears in an answer.

Which GEO metrics become vanity metrics?

Raw mention counts and a single visibility score are weak when they omit the prompt panel, engine, period, repetition and decision context. Even a carefully defined rate can move because of sampling variation rather than a meaningful change in the market or business. These measures are not inherently useless; they need the context required to interpret them.

A repeated-sampling preprint across three consumer-product topics and three generative search platforms found substantial citation variability and unstable rankings; some apparent differences fell within sampling noise the repeated-sampling preprint. That is a reason to be cautious about small differences, not evidence for a universal GEO KPI or a B2B benchmark. A trend line can be measured consistently and still reflect sampling variation, a changed model or a changed prompt panel rather than business progress.

How should teams connect AI visibility to pipeline?

Join trackable assistant referrals to qualified leads and opportunities, but label observed traffic and assigned attribution separately from incremental effect. Last-click reports and model-assigned credit can describe observed journeys; neither proves that AI visibility caused the outcome Google Analytics’ attribution-path documentation and Google Ads’ Conversion Lift documentation.

1. Document the referrers and GA4 rules used to classify assistant traffic, including what remains unknown or direct Google Analytics’ default-channel documentation.

2. Reconcile analytics sessions with CRM records using explicit lead and opportunity stage definitions. Keep unmatched visits visible rather than silently assigning them to a channel.

3. Test incrementality with a controlled design where feasible; check the method’s eligibility and applicability before relying on a platform experiment Google Ads’ Conversion Lift documentation.

Ahrefs reported in a June 16, 2025 first-party case study that, on its own site over the preceding 30 days, AI assistants supplied 0.5% of visits and 12.1% of signups, with a 23-times higher conversion-per-visit rate than traditional organic search Ahrefs’ first-party conversion case study. This is an SEO software company’s case study, not a B2B SaaS benchmark or causal estimate. Its definitions and other sites’ audiences may differ, so the figures should not be used as a forecast.

The operational handoff should make clear which records are observed, which receive attribution credit, and which outcomes have been tested for incrementality. Keep answer-level visibility in its own reporting layer rather than treating it as a substitute for analytics or CRM evidence.

What should a monthly GEO scorecard disclose?

A monthly scorecard should include the numerator, denominator, prompt set, engine/model configuration, run count, collection date and uncertainty for each answer-level measure. It should state which decision each row supports and keep answer-level results separate from analytics and CRM measures. If a field is unknown, label it as unknown rather than filling the gap with a blended score.

A scorecard is useful when a reader can reconstruct what was counted and why. It should not imply that answer visibility has been tied to qualified demand unless the analytics and CRM evidence supports that conclusion.

What does brand share of voice mean in AI answers?

Brand share of voice in AI answers can mean a competitive share of coded recommendations, but it should not be confused with the brand’s mention rate or citation rate per prompt-run. For a specified engine, prompt panel and period, recommendation share is brand recommendations divided by all coded brand recommendations. State which measure you mean whenever using “share of voice.”

A prompt-run with no brand in the answer still belongs in the denominator for mention, recommendation and citation rates if it was eligible and successfully collected. It contributes zero events to those numerators. It does not contribute a recommendation to the competitive recommendation-share denominator. One answer can mention several brands, and recommendations can also be multi-brand; define how each is coded before collection.

For a complete view, report recommendation share alongside per-cell mention rate, per-cell recommendation rate, citation rate and, where reviewed, factual-support rate. Those measures answer different questions: who is recommended among coded recommendations, how often a brand appears, how often it is cited, and whether the source supports the answer.

How do you build a comparable prompt panel across LLMs?

Hold the buyer-intent prompt set and observation window consistent, then report each platform/model configuration separately. Record exact prompt wording, date, region, language, intent and repeated-run count. A pooled score can conceal meaningful differences by platform or intent, while changing the panel can make a trend comparison difficult to interpret.

1. Define buyer-intent strata, such as research questions and vendor-comparison questions, and describe the question population the panel is intended to represent.

2. Freeze prompt wording for the comparison period. Keep any refreshed panel as a separately identified series.

3. Run repeated samples for each prompt and engine/model configuration, recording collection date and relevant platform details.

4. Code every eligible output, including zero-mention answers and uncodable results; retain an explicit status for collection failures.

5. Calculate segment-specific measures with their own denominators, separating mentions, recommendations and citations.

6. Report sample sizes and uncertainty at the level of the claim. Have a statistical reviewer assess the interval method and dependence assumptions.

Before collection, specify how ties, multi-brand responses, uncodable outputs and answers with no brand mention will be handled. Keep the panel balanced for the intended buyer-question population; a large set of paraphrases is not necessarily broad coverage. Available evidence does not establish a universal number of prompts or runs.

Why can apparent brand rankings shift between runs or engines?

Generative-search citations can vary across repeated samples, so small differences between brands may be sampling noise the repeated-sampling preprint.

The repeated-sampling preprint studied Perplexity Search, OpenAI SearchGPT and Gemini on three consumer-product topics over nine days. It found substantial citation variability and unstable rankings; some apparent domain differences were within sampling noise the repeated-sampling preprint. The study supports caution when interpreting close rankings, but it is not a B2B category benchmark and does not establish a universal share-of-voice standard.

Report intervals and avoid ordering brands when their uncertainty substantially overlaps. Available evidence does not establish a universal uncertainty formula or minimum sample size for every prompt panel. A statistical reviewer should assess the implementation, especially where multiple runs share prompts, time periods or model/index state.

How many prompts and runs are enough for AI visibility measurement?

There is no universal count. Distinct prompts determine how broadly the measurement covers buyer questions; repeated outputs per prompt/model cell help reveal variation in answers to the same question. More repetitions cannot repair a narrow or biased panel, and more prompts do not by themselves reveal within-cell variability.

A prompt/model cell is one defined prompt paired with one specified engine/model configuration in a fixed market and collection period; each collected repetition is a prompt-run observation. If the platform configuration changes, record the change rather than treating observations as directly interchangeable.

What does prompt coverage represent?

Prompt coverage represents the buyer-question panel and the population it is meant to approximate. Organize prompts by buyer stage, intent, language and engine, remove redundant paraphrases, and freeze the panel when measuring a trend. The panel should reflect the questions the team intends to study, not merely the questions that are easy to generate.

For an illustrative B2B SaaS panel, a team might separate prompts about defining a problem from prompts comparing vendors. Those are different intents and should be reported separately if they matter to the decision. This example describes a design choice, not observed performance or a recommended fixed panel size.

Why repeat a prompt on the same model?

Repeating a prompt on the same model configuration reveals whether answers and citations vary within that prompt/model cell. Runs collected close together may still be correlated because they can share model or index state, so the number of runs is not automatically the number of independent observations. The repeated-sampling study collected results over nine days across three platforms and three consumer-product topics; it informs measurement caution, not B2B generalization the repeated-sampling preprint.

What can a confidence-interval calculation tell you?

For a binary cited/not-cited rate, a conservative planning approximation uses an expected rate of p = 0.5. It gives about 97 independent runs for a 95% interval of roughly ±10 percentage points, or 385 for roughly ±5 points the evidence-aware prompt-sampling guide. These are independent Bernoulli planning counts, not a platform rule or a guarantee of achieved precision. If repeated answers share state or otherwise depend on one another, the effective information can be lower the evidence-aware prompt-sampling guide.

These approximations do not specify a B2B variance estimate and should not be applied as if every prompt-run were independent. They are not counts of distinct prompts. Prompt coverage and repeated runs answer different sampling questions.

At what level should uncertainty be reported?

Match the interval and denominator to the claim: a prompt-level result needs a prompt-level denominator, an engine-level claim needs observations from that engine, and a category-level claim needs a defensible category sample. Pooling all runs without accounting for shared prompts can overstate certainty. Cluster or bootstrap approaches may be considered for dependent observations, but they are methods to review, not a guarantee that the result is valid the repeated-sampling preprint and the evidence-aware prompt-sampling guide.

How should a B2B team design a first pilot?

Start with a broad, balanced panel and about 10 repeated runs per important prompt/model cell as a directional screen, not statistical certification the evidence-aware prompt-sampling guide. Use the pilot to inspect coding and answer variability, then decide whether the next study needs broader prompt coverage, more repetition or both. This figure is not a universal recommendation for every platform or decision.

1. Define the buyer-question universe and the decision the pilot should inform.

2. Stratify questions by relevant buyer intent, language, region and engine; deduplicate paraphrases and freeze the initial panel.

3. Record the engine/model configuration, collection time, prompt wording and citation outcome for each run.

4. Repeat important prompt/model cells, treating the approximate 10-run screen as directional only the evidence-aware prompt-sampling guide.

5. Publish the denominators, coding rules and uncertainty, while noting dependence and uncodable outputs.

6. Decide whether to increase prompt breadth or repeated runs based on the gap between the evidence and the decision requirement.

A frozen panel helps comparison over time, but model or index changes may call for a separately tracked refresh. Preserve the old panel as a comparison series rather than silently replacing it.

When should the sample size increase?

Increase prompt coverage when the buyer-question universe is undersampled; increase repeated runs when within-cell uncertainty is too wide for the decision. Before planning a full study, specify the minimum detectable change, estimate variance from a B2B pilot and set a decision threshold. A change in visibility over time is not, by itself, evidence of causality or proof that the brand changed while the model remained stable.

Separate platform or model drift from brand-level change by recording configurations and dates, retaining a frozen panel, and tracking any refreshed panel separately. These figures do not support a company-specific sample-size calculation or performance claim.

What should a measurement report disclose?

Report the prompt panel and strata, engine/model, collection dates, runs per cell, denominator, binary citation definition and uncertainty. Acknowledge correlated observations and platform changes, and avoid reducing the results to one aggregate score. The report should let another reader understand what was counted and what the evidence can support.

• Panel: prompt wording, intended question population, intent strata, language and region.

• Collection: platform/model configuration, dates and runs per prompt/model cell.

• Measure: numerator, denominator and operational definition for each mention, recommendation or citation outcome.

• Uncertainty: interval method, dependence caveats and the level at which the claim is made.

• Coding: tie and multi-brand rules, zero-mention handling, uncodable outputs and citation-support review status.

• Change log: panel refreshes, model/index changes and other collection differences.

Frequently Asked Questions

What is share of voice marketing?

Share of voice is a brand’s portion of a defined set of market mentions or recommendations. In AI answers, specify the unit and denominator: recommendation share divides a brand’s coded recommendations by all coded brand recommendations in the panel, while mention rate divides prompt-run cells mentioning the brand by all eligible cells.

How do you measure AI visibility?

Define a fixed prompt panel, engine/model configuration, collection period, and coding rules, then repeat the runs and report each metric with its numerator, denominator, and uncertainty. Keep mentions, recommendations, citations, referral sessions, and CRM outcomes separate because each describes a different layer.

What is a good AI visibility score?

Available evidence does not establish a universal score or benchmark. A useful score is a clearly defined measure tied to a decision, reported for a representative prompt panel with its denominator, platform, collection conditions, and uncertainty.

What are some examples of vanity metrics?

Uncontextualized mention or citation totals, a blended visibility score, assistant-referral volume without traffic-quality measures, and a single brand ranking can mislead when their samples, definitions, or uncertainty are hidden. They become more useful when reported with the relevant denominator and decision context.

What is the difference between KPIs and vanity metrics?

A KPI is useful when its definition and evidence support a specific decision; a vanity metric can look impressive but does not clarify performance or what to do next. The same count can be useful or misleading depending on whether its sample, denominator, limitations, and connection to the decision are disclosed.

How Nortav Can Help

For teams building a GEO measurement program, the first step is to clarify what each visibility, traffic, and pipeline measure can—and cannot—show. Nortav can help assess the measurement gaps and connect the scorecard to commercially relevant outcomes.

Our approach starts with understanding your current search visibility, identifying gaps against competitors, and developing a strategy focused on driving qualified traffic, generating leads, and delivering measurable business value.

Book a Discovery Call →

Sources

• “Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement.” 2026 preprint.

• Ahrefs, “Does AI Search Traffic Convert Better Than Traditional Search? For Ahrefs, Yes: 0.5% of Visitors Drove 12.1% of Signups.” June 16, 2025.

• Google Analytics, “Key events attribution paths report.”

• Google Ads, “About Conversion Lift.”

• “Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines.” 2025 study.

• Answer Engine Land, “How Many Times Should You Repeat an AI Search Prompt? An Evidence-Aware Sampling Guide.”

• Google Analytics, “Default channel group.”