GeoZ Metrics Dictionary: AI Search Visibility, Citation Quality, and Intent Resolution

TL;DR

• An AI-search metric is trustworthy only when its denominator and eligibility rule are visible. “We appeared 62% of the time” means little until you know which prompts, products, markets, dates, repeats, and failed observations were included.


• Measure answer roles separately. Presence, citation, fit recommendation, accurate recommendation, claim accuracy, traffic, leads, and pipeline answer different business questions. A single unexplained GEO score hides those differences.


• Use a governed prompt portfolio as the observation base. A stable panel converts screenshots into comparable data, but the panel describes the selected buyer questions—not all AI searches or market share.


• Report distributions and variance, not one-day verdicts. AI answers change by wording, product, mode, locale, source environment, and time. The operating question is what stayed stable, what changed, and what action the pattern supports.


• Treat proprietary scoring as an auditable contract, not a magic number. GeoZ does not publish its proprietary algorithm formulas here. Buyers should still be able to inspect metric purpose, input scope, missing-data handling, version, confidence, and decision use.


• Connect answer visibility to business outcomes without pretending the journey is fully observable. GA4 can show attributable AI Assistant sessions and conversions. It cannot reconstruct every answer exposure, copied URL, direct return, or later branded search.


• GeoZ combines measurement with execution. Its Value as a Service model helps agencies and in-house teams use proprietary metrics to choose, implement, and review the next SEO/GEO action—not merely watch a dashboard.

What Is the GeoZ Metrics Dictionary?

The GeoZ Metrics Dictionary is a buyer-facing contract for interpreting AI-search measurement. It defines which observable outcomes should be counted, which denominators make a percentage meaningful, which limitations belong beside the number, and which business decision each metric can support.

It is deliberately not a disclosure of GeoZ’s proprietary algorithms. A company can protect the formula that creates differentiated value while still giving customers enough methodological transparency to challenge a score, reproduce its public inputs, understand changes, and decide whether an action is justified.

That distinction matters because AI-search dashboards frequently place unlike signals beside each other:

• a brand mention in one answer;


• a visible citation to an owned page;


• a recommendation for a buyer with a specific constraint;


• an accurate or inaccurate description of a product claim;


• a click from an AI assistant;


• a form completion on the landing page;


• an opportunity recorded in a CRM;


• a weighted vendor score whose inputs are not shown.

All 8 can be useful. None is a substitute for the other 7.

For CMOs and VPs, the dictionary prevents a reporting number from outrunning its evidence. For agency leaders, it makes client reporting more defensible across accounts. For in-house teams, it creates a shared language between SEO, content, product marketing, analytics, RevOps, and leadership.

Why AI-Search Measurement Needs a Contract

Traditional analytics already requires definition discipline. “Traffic,” “conversion,” and “pipeline” change meaning when the channel rule, event setup, identity resolution, or CRM stage changes. AI search adds another layer because part of the buyer’s experience can happen inside an answer surface before a click exists.

The result is a measurement environment with 4 recurring problems.

Problem 1: The unit is unclear

One dashboard counts prompts. Another counts answers. A third counts products multiplied by prompts. A fourth counts citations. If 50 prompts are run on 3 products with 2 repeats, there are 300 planned observations—not 50. A percentage calculated over 50 and one calculated over 300 are not comparable.

Problem 2: Missing observations disappear

A failed run, blocked page, unavailable product, or unreviewable answer can be removed from the denominator without disclosure. The resulting percentage improves even though the data quality became worse.

Problem 3: Answer roles are blended

A mention, citation, and recommendation are different states. If a score assigns 1 point for a mention, 2 for a citation, and 3 for a recommendation, the weighting may be reasonable for one use case and wrong for another. Without the component distribution, leadership cannot tell what moved.

Problem 4: Correlation becomes causation

A new page goes live on April 1. Accurate recommendations improve in the April 15 run. That sequence is compatible with an effect, but model changes, competitor changes, new third-party sources, prompt variance, and product-mode changes may also contribute.

The Community’s weather-system model for AI-search visibility provides the right mental model: record conditions, observe a panel, and report a distribution. One answer is an event. Repeated, conditioned observations become a measurement system.

The 10 Fields Every Metric Must Declare

Before accepting a metric, ask for its metric card. The card is a compact contract that allows an analyst, client, or executive to understand what the number includes.

#### Add a version to every metric card

A metric definition can change. A taxonomy may add “overbroad” as an accuracy state. A product may introduce a new answer mode. GA4 may reclassify traffic. Record definition version 1.0, 1.1, or 2.0 and the effective date.

Preserve the raw component values

If a proprietary score changes from 54 to 68, the customer should be able to see whether the movement came from more presence, better citation coverage, more accurate recommendations, a different prompt set, or a revised weighting model. The formula can remain private; the diagnostic components should not disappear.

Keep the decision attached

A metric with no possible action becomes reporting inventory. Each metric should route to a choice: investigate a source, repair a claim, create a comparison page, improve a landing page, check tracking, review sales qualification, or leave the content unchanged until more observations accumulate.

Build the Observation Base Before Calculating Coverage

Coverage metrics inherit the strengths and weaknesses of the prompt portfolio underneath them. A carefully formatted percentage cannot repair a biased set of questions.

GeoZ recommends starting from buyer decisions and recording the panel design before the first run. The detailed workflow is in the 50-prompt AI-search evaluation panel guide.

Prompt Portfolio Coverage

Prompt Portfolio Coverage is the share of required prompt slots that contain an approved, active prompt.

Prompt Portfolio Coverage = approved active prompts / required prompt slots

If a design requires 50 slots and 47 contain approved prompts, coverage is 47 / 50 = 94%. The metric describes design completion. It does not say the brand appears in 94% of answers.

#### Observation Eligibility Rate

Observation Eligibility Rate measures how much of the planned collection produced reviewable observations.

Observation Eligibility Rate = eligible observations / planned observations

Suppose 50 prompts are run across 3 products with 2 repeats. The plan contains 50 × 3 × 2 = 300 observations. If 9 fail collection and 3 cannot be reviewed, 288 remain eligible. Eligibility is 288 / 300 = 96%.

Report the 12 missing observations beside every coverage metric. Otherwise the result can rise merely because unfavorable or difficult cases disappeared.

Intent-Route Coverage

Intent-Route Coverage checks whether the panel includes the buyer routes defined in the design. A B2B panel might require problem framing, category discovery, comparison, fit, implementation, risk, and measurement. An ecommerce panel might require product discovery, attribute fit, comparison, availability, policy, and purchase confidence.

The Community’s Hidden Intent Map illustrates why route design matters. Its DevOps corpus contained 74,346 responses, 984 prompt templates, 35 intent categories, and 14 retrieval routes. Those distributions belong to that corpus, not every industry. The transferable lesson is that a keyword does not capture the evidence route behind a buyer question.

Answer Presence Metrics

Presence is the lightest observable answer state. It tells you whether the monitored entity appears, but not why it appears, how prominently it appears, or whether the description is correct.

Answer Presence Coverage

Answer Presence Coverage is the share of eligible observations in which the target entity appears in the answer under the documented matching rule.

Answer Presence Coverage = observations with target entity present / eligible observations

The matching rule should define:

• exact brand-name matches;


• accepted abbreviations;


• product names that count toward the parent brand;


• common misspellings;


• whether a cited URL with no visible brand mention counts;


• whether a competitor comparison that names the brand counts;


• whether an answer in a non-target language is eligible.

If 288 observations are eligible and the entity appears in 126, presence coverage is 126 / 288 = 43.8%.

Qualified Presence Coverage

Not every presence belongs in a positive KPI. Qualified Presence Coverage limits the numerator to appearances relevant to the intended category, use case, or buyer constraint.

If 126 observations contain the brand but 18 attach it to the wrong category, qualified presence is 108 / 288 = 37.5%. The 18 mismatches become a positioning or semantic-integrity queue.

Presence by intent family

A company can appear in 80% of definition prompts and 10% of comparison prompts. The blended rate may look stable while the commercial layer remains weak.

Presence supports category-recognition analysis. It does not prove citation, preference, recommendation, traffic, or revenue.

Citation and Source Metrics

A citation is a visible attribution event. It can reveal which pages or domains the answer surface credits, but it should not be treated as complete evidence of source influence. Some answer modes show citations; others do not. A source can influence wording without visible credit, and a citation can appear without carrying the brand’s most important claim.

Citation Coverage

Citation Coverage is the share of eligible observations where at least 1 approved target URL or domain is visibly cited.

Citation Coverage = observations with target citation / citation-eligible observations

The denominator must be citation-eligible. If a product or mode does not expose citations in the observed interface, placing those answers in the denominator penalizes the brand for a product-design difference.

Owned Citation Coverage

Owned Citation Coverage counts observations citing the company’s own domain. It helps diagnose whether product pages, documentation, research, and editorial assets are being visibly selected.

Corroborating Citation Coverage

Corroborating Citation Coverage counts approved third-party sources that accurately support the target entity or claim. The goal is not to collect arbitrary mentions. It is to understand whether credible external evidence exists where buyer questions require it.

Source Domain Coverage and concentration

Source Domain Coverage is the number or share of distinct relevant domains appearing across a panel. Source Concentration shows how much visible citation activity comes from the top 1, top 3, or top 5 domains.

Do not assign automatic authority merely because a domain ends in .gov or .edu. Domain suffix is not a substitute for topical relevance, claim fit, authorship, recency, methodology, or accurate representation.

Recommendation Metrics

Recommendations sit closer to a buyer shortlist than mentions or citations. They also create more reputational risk because an answer can recommend a brand for the wrong company size, industry, integration requirement, or governance constraint.

Fit Recommendation Coverage

Fit Recommendation Coverage is the share of recommendation-eligible observations where the brand is suggested for the target buyer and stated constraint.

Fit Recommendation Coverage = observations with a fit recommendation / recommendation-eligible observations

The coding guide should distinguish:

• recommended as a primary option;


• recommended as a conditional option;


• included in a list with no fit rationale;


• mentioned as an alternative but not recommended;


• explicitly rejected for the constraint;


• absent.

Accurate Recommendation Coverage

Accurate Recommendation Coverage requires both fit and correct representation.

Accurate Recommendation Coverage = accurate fit recommendations / recommendation-eligible observations

If 72 comparison and fit observations are eligible, 24 recommend the brand, and 18 are accurate for the stated constraint, fit recommendation coverage is 24 / 72 = 33.3% while accurate recommendation coverage is 18 / 72 = 25.0%.

The 6-point gap is commercially important. It means 25% of recommendations are overbroad, outdated, or otherwise inaccurate: 6 / 24 = 25.0%.

Recommendation Position

A recommendation can appear first, third, or in an unranked list. If position is recorded, preserve it as an observation rather than converting it into an arbitrary universal value. Product interfaces format lists differently, and the apparent order may not represent a stable preference.

## Claim Accuracy and Semantic Integrity Metrics

Visibility without accuracy can create a larger problem than invisibility. An answer may use the brand name while changing the audience, entity, evidence, condition, comparison, or time boundary of a claim.

The Community’s claim-drift framework explains why a canonical claim card matters. The monitored answer should be compared with a source-of-truth claim that includes scope and boundaries—not with a vague memory of what marketing intended.

Claim Accuracy Rate

Claim Accuracy Rate is the share of coded claim observations that match the approved source of truth within the defined tolerance.

Claim Accuracy Rate = accurate coded claims / eligible coded claims

Use an explicit taxonomy:

• accurate;


• accurate but incomplete;


• overbroad;


• outdated;


• wrong entity;


• wrong capability;


• unsupported comparison;


• unverifiable;


• not applicable.

If 84 claim observations are eligible and 59 are accurate, the rate is 59 / 84 = 70.2%. Do not hide the remaining 25 in one “inaccurate” bucket. The remedy for outdated information differs from the remedy for a wrong capability.

Boundary Preservation Rate

Boundary Preservation Rate asks whether the audience, condition, geography, time period, sample, and outcome boundary survive summarization.

Suppose 40 answers repeat a decision-critical claim. If 28 preserve the audience and condition, boundary preservation is 28 / 40 = 70%. The 12 failures may require the company to move boundaries closer to the claim across product pages, research, FAQs, comparison pages, and third-party materials.

Semantic Integrity Issue Rate

Semantic Integrity Issue Rate is the share of eligible answers containing at least 1 material error that could change a buyer decision.

Do not count punctuation, harmless paraphrase, and material product error equally. Define materiality before review.

## Intent-Resolution Metrics Without Invented Session Behavior

“Did the content resolve intent?” is a valuable question. The mistake is turning it into a precise session metric when the required behavior is not directly instrumented.

Ordinary GA4 data does not reveal the user’s full conversation before or after a site visit. It generally cannot tell whether the person returned to an AI assistant, reformulated the prompt, received another answer without clicking, or continued on another device. A formula that counts “sessions with no reformulation” therefore requires product-level conversational data, an instrumented study, or another declared proxy—not a standard landing-page report.

Use an Intent-Resolution Evidence Ladder

Instead of one invented score, report the available evidence by level.

#### Decision-Block Coverage

Decision-Block Coverage is an editorial QA metric: the share of required buyer questions answered on the target page or cluster.

If a vendor-evaluation page requires 12 blocks and 10 pass review, coverage is 10 / 12 = 83.3%. It supports an editorial decision. It does not prove 83.3% of visitors resolved their intent.

Next-Step Completion Rate

Next-Step Completion Rate is the share of eligible landing-page sessions that complete the declared next event: view methodology, compare approaches, start a demo request, submit a form, or another meaningful action.

Keep the event specific. A scroll or generic click may indicate engagement, but it should not automatically become “intent resolved.”

Assisted Resolution Evidence

Use sales notes, survey responses, or research interviews to understand whether content answered objections. State sample size and selection. Twelve positive interviews can produce insight; they do not establish a population-wide conversion rate.

Variance and Stability Metrics

AI-answer variance is not noise to delete. It is part of the observed system. The measurement job is to distinguish a durable pattern from a single favorable or unfavorable event.

Cross-Surface Range

Cross-Surface Range is the difference between the highest and lowest product-level rate for the same metric and period.

If accurate recommendation coverage is 42% on Product A, 28% on Product B, and 17% on Product C, the range is 42 - 17 = 25 percentage points.

The range shows heterogeneity. It does not explain the cause. Products can differ in model, mode, source access, citation behavior, locale, and interface.

Repeat Agreement Rate

Repeat Agreement Rate is the share of prompt-product pairs where repeated observations receive the same coded state.

If 150 prompt-product pairs have 2 repeats and 96 pairs agree, repeat agreement is 96 / 150 = 64%. The remaining 54 pairs require a distribution, not deletion.

Stable Accurate Coverage

Stable Accurate Coverage counts cases that meet the desired accurate state across the required repeat rule. A stringent rule might require 2 of 2 repeats; a broader rule might require 2 of 3. Publish the rule beside the result.

#### Change With Confidence Label

Every reported movement should carry a label such as:

• observed change;


• repeated pattern;


• plausible contribution after a documented intervention;


• quasi-experimental evidence;


• controlled causal evidence;


• unresolved.

This prevents “coverage increased after publication” from becoming “the article caused a 19% visibility lift” without the necessary design.

GA4 AI-Assistant Traffic Metrics

Answer measurement and site analytics are separate instruments. One observes what appears in an answer environment. The other observes what happens when a trackable visitor reaches the site.

The current GeoZ GA4 AI-traffic guide explains the operating setup. The Community’s native AI Assistant channel update explains the classification change and the historical reporting break.

AI Assistant Sessions

AI Assistant Sessions are sessions classified under the relevant GA4 AI Assistant channel or validated source/medium rule for the reporting period.

Track at least:

• sessions;


• users where appropriate;


• landing page;


• source or product when available;


• engagement metrics;


• key events;


• qualified lead events;


• revenue or pipeline only when the downstream identity and stage rules support it.

AI Assistant Key-Event Rate

AI Assistant Key-Event Rate = AI Assistant sessions with target key event / eligible AI Assistant sessions

If 420 attributable sessions produce 34 demo-request starts, the observed rate is 34 / 420 = 8.1%. If 21 requests are submitted, the submit rate is 21 / 420 = 5.0%. Do not collapse start and submit into one event.

Landing-Page Qualified Lead Rate

If 21 submissions generate 9 accepted qualified leads, the submission-to-qualified rate is 9 / 21 = 42.9%. The session-to-qualified rate is 9 / 420 = 2.1%.

These numbers are illustrative. Their purpose is to show the denominators that disappear when a dashboard reports only “AI conversions.”

Dark-Funnel and Attribution Confidence Metrics

GA4 records observable visits. It does not record every buyer who reads an answer, remembers a brand, later searches directly, asks a colleague, or returns through another channel. The Community’s AI-search dark-funnel analysis separates answer exposure, referral behavior, demand movement, on-site quality, and causal confidence.

Attribution Confidence

Attribution Confidence should be a declared evidence label, not a secret multiplier. A practical scale can use 5 levels:

Do not convert every Level 1 observation into pipeline. Do not dismiss Level 1 because it lacks a click. Each level answers a different question.

Qualified Pipeline From Observable AI Sessions

Qualified Pipeline From Observable AI Sessions is the sum of accepted opportunity value where the attribution rule connects an AI Assistant session to the opportunity within a declared window.

Declare whether the report uses first touch, last non-direct, position-based, data-driven, or a custom influence rule. A dollar amount without the rule is not reproducible.

Demand Correlation Monitor

A team may monitor answer coverage beside branded search, direct entry, sales mentions, and pipeline. Call this a correlation monitor unless the design supports causal inference. Movement across 3 or 4 layers can justify investigation and investment without being presented as proof.

The GeoZ AI-search ROI framework shows how to connect visibility, observable behavior, qualified demand, program cost, and confidence without forcing them into a false linear funnel.

How to Evaluate a Proprietary GeoZ Metric

GeoZ uses proprietary algorithms and metrics as part of its SEO/GEO Value as a Service model. Proprietary does not have to mean unchallengeable. A buyer can evaluate a method without receiving the source code, full weighting formula, or trade-secret implementation.

Ask what the metric is for

Is the score designed to summarize visibility, diagnose a content gap, prioritize an action, compare time periods, compare products, or estimate business impact? One score should not quietly perform all 6 jobs.

Ask what the score can be unpacked into

A useful score should allow review of the underlying prompt family, answer states, citation sources, accuracy labels, time period, and missing observations. If the composite moves, the component story should remain available.

Ask how inputs are governed

The input contract should cover:

• prompt inclusion and retirement;


• product, mode, country, and language;


• repeat count;


• brand and entity matching;


• citation eligibility;


• recommendation taxonomy;


• accuracy rubric;


• missing-data treatment;


• outlier and duplicate handling;


• human review and disagreement resolution;


• effective version date.

Ask how the score changes decisions

A diagnostic metric earns its place when a low or changing value routes to a specific investigation. A score that always recommends “publish more content” is not sufficiently diagnostic.

## The Legacy Acronyms GeoZ Is Retiring

This canonical page replaces 4 legacy articles because their acronyms were inconsistent or insufficiently observable. Redirecting them is a quality decision, not a claim that measurement should become simpler.

PCV is ambiguous

One legacy page used PCV for Page Citations View. Another used it for Prompt Coverage Velocity. A metric name cannot support governance when two readers can apply different numerators.

Use the explicit phrase that matches the decision: Citation Coverage, Prompt Portfolio Coverage, Answer Presence Coverage, or a time-series change in one of those metrics.

WSU is ambiguous and domain suffix is not authority

One legacy page introduced Web Semantic Understanding. Another used Weighted Source Utilization to prioritize .gov and .edu mentions. Neither label had a stable, owner-approved GeoZ definition. The suffix-based weighting also risks confusing domain type with topical evidence quality.

Use Source Domain Coverage, Source Concentration, Corroborating Citation Coverage, and a documented source-review rubric.

CPD lacks an approved definition

Content Potential Dynamics was introduced without a verified operating definition. A metric should not survive merely because its acronym sounds technical.

Use the observable diagnostic that supports the decision: Decision-Block Coverage, prompt-family gap, claim-accuracy issue, or change in eligible answer outcomes after a documented intervention.

CCR, IBC, and SSR imply data that may not exist

The old Clarification Capture Rate, Intent Branch Coverage, and Session Stabilization Rate framework included useful editorial questions: Does the page answer follow-ups? Does it cover important variants? Does the visitor take the next step? The formulas, however, implied direct knowledge of follow-up and reformulation behavior.

Retain the questions. Replace the unsupported formulas with the Intent-Resolution Evidence Ladder and instrument only what the team can actually observe.

The permanent redirects preserve the useful destination while preventing contradictory definitions from continuing to circulate.

A Worked Executive Scorecard

An executive scorecard should summarize the program without erasing the components. The following example uses a 50-prompt panel, 3 products, 2 repeats, and a monthly review. All figures are illustrative.

#### The 5-minute CMO reading order

Read the scorecard in this order:

1. Eligibility: did the measurement system work?


2. Accuracy: is the brand represented correctly?


3. Recommendation: is it entering relevant shortlists?


4. Citation and source: what evidence environment is visible?


5. Behavior: are attributable visitors taking meaningful steps?


6. Commercial outcome: are qualified opportunities appearing under the declared rule?


7. Confidence: what can leadership responsibly conclude?

The 30-minute operating review

The working team should then inspect the prompt family, product, market, page, source, and issue-type cuts. Every priority issue receives an owner, next action, expected observable change, and review date.

The quarterly method review

Do not redesign the panel every month. Review whether buyer language, products, competitors, markets, or decision routes changed enough to warrant a versioned panel update. Preserve historical prompt IDs and document replacements.

What Agencies and In-House Teams Should Do Differently

The measurement contract is shared, but the operating model changes by team.

For SEO and GEO agencies

Use one definition system across accounts, but do not force every client into the same prompt set or benchmark. Normalize the methodology, not the outcome.

An agency should preserve:

• client-specific buyer decisions;


• separate market and language panels;


• evidence and brand-claim sources of truth;


• collection conditions;


• reviewer training and disagreement logs;


• client-visible components behind proprietary prioritization;


• content, digital PR, technical, analytics, and conversion actions;


• the difference between agency contribution and causal proof.

For in-house SEO and GEO teams

Connect the metric owner with the person who can change the underlying evidence. A claim-accuracy problem may belong to product marketing or documentation. A citation problem may require editorial or third-party corroboration. A qualified-lead problem may belong to landing-page conversion, analytics, RevOps, or sales follow-up.

For CMOs and VPs

Do not ask only whether “the GEO score went up.” Ask what changed in the buyer journey, which evidence supports the interpretation, what the team will do next, and what would falsify the current explanation.

## Where GeoZ Fits

Teams do not usually fail because they cannot display another chart. They fail when measurement, diagnosis, execution, and review live in different queues with no shared operating contract.

The How GeoZ Works operating loop connects the company’s Value as a Service model for SEO and GEO with in-house tools, proprietary algorithms, proprietary metrics, and the execution needed to improve the underlying content and evidence system. The intended value is not a promise that one score reveals every hidden model decision. It is a disciplined route from observation to prioritized action.

Measurement

Build and maintain a governed view of buyer prompts, answer roles, citations, accuracy, model or surface variance, traffic, and business outcomes.

Diagnosis

Separate a presence gap from a citation gap, a fit problem from a claim problem, a data-quality failure from a true decline, and a traffic issue from a qualification issue.

Execution

Translate the diagnosis into content refreshes, industry pages, comparison evidence, claim cards, source work, technical fixes, analytics instrumentation, or conversion improvements.

Review

Repeat the observation under documented conditions, preserve the change log, and report what improved, what stayed unresolved, and how confidence changed.

This model can support an agency that needs repeatable delivery across clients, an in-house team that lacks a dedicated GEO research operation, or an executive who needs a decision system rather than a vanity dashboard.

A 30-Day Metrics Implementation Plan

A credible dictionary becomes valuable only when it is used. Start with a small, complete measurement loop.

Days 1–5: Define the decision

• Name the executive and operating owners.


• Choose 1 business decision for the first 30-day cycle.


• Define the buyer, category, market, and language boundary.


• Approve the 10-field metric-card template.


• Freeze metric version 1.0 for the cycle.

Days 6–10: Build the panel and sources of truth

• Approve 25–50 prompts across buyer-decision routes.


• Assign stable prompt IDs.


• Record products, modes, repeats, and expected answer roles.


• Create canonical claim cards for the top 10 decision-critical claims.


• Identify owned pages and approved corroborating sources.

Days 11–15: Run and review

• Collect planned observations.


• Report failed and missing runs.


• Code presence, citation, recommendation, and accuracy separately.


• Resolve reviewer disagreement on a sample before full scoring.


• Calculate public component metrics and preserve raw evidence.

Days 16–20: Connect behavior

• Validate the GA4 AI Assistant channel and historical classification boundary.


• Check landing-page and key-event instrumentation.


• Define qualified-lead acceptance with RevOps or sales.


• Confirm CRM source and influence rules.


• Label dark-funnel indicators as correlation, not direct attribution.

Days 21–25: Prioritize action

• Rank issues by buyer value, evidence strength, effort, and reversibility.


• Assign no more than 5 priority actions for the cycle.


• Name an owner and expected observable change for each action.


• Avoid rewriting pages because of 1 volatile answer.


• Preserve a control or comparison where practical.

Days 26–30: Publish the operating review

• Show eligibility before outcome metrics.


• Report distributions and component changes.


• Attach confidence labels.


• Document method, product, content, and market changes.


• Schedule the next repeat and the quarterly panel review.

If your team wants help building this measurement-to-execution loop, talk with GeoZ. Bring the current dashboard, prompt list, GA4 setup, or client reporting template. The useful starting point is the decision you need to make—not the score you want to display.

FAQs

What is the most important GEO metric?

There is no universal single most important GEO metric. For a buyer-evaluation program, Accurate Recommendation Coverage may be more useful than raw presence. For brand integrity, Claim Accuracy Rate may matter most. For an acquisition team, qualified leads from attributable AI Assistant sessions may be the relevant observable outcome. Start with the business decision, then select the metric whose unit and denominator support it.

Does GeoZ publish its proprietary metric formulas?

No proprietary algorithm formulas are disclosed in this public dictionary. The page defines the transparency contract a buyer should expect: purpose, input scope, unit, eligibility, missing-data rule, version, confidence, component drill-down, and decision use. That allows a customer to evaluate and challenge a proprietary score without requiring GeoZ to publish trade-secret weighting or source code.

Can a 50-prompt panel measure AI-search market share?

No. A 50-prompt panel is an internal observation portfolio, not a census of all real prompts, users, products, modes, markets, or time periods. It can compare a stable set of buyer questions across runs and diagnose patterns. It should not be described as total market share unless a separate, defensible sampling design supports that claim.

Should mentions, citations, and recommendations be combined into one score?

They can contribute to a proprietary prioritization model, but the underlying components must remain visible. A mention tests recognition, a citation tests visible source credit, and a recommendation tests fit in a buyer context. Claim accuracy should also remain separate. If the composite changes from 54 to 68, the team should be able to see which component and scope changes produced the movement.

Can GA4 show whether a visitor reformulated an AI prompt after reading a page?

Not through ordinary site analytics alone. GA4 can observe attributable sessions and instrumented on-site events. It generally cannot observe the visitor’s full conversation inside an external AI assistant, later no-click answers, copied URLs, cross-device returns, or prompt reformulation. Use on-site behavior, research, sales evidence, and controlled studies as separate evidence levels.

Why were the old PCV, WSU, CPD, CCR, IBC, and SSR pages consolidated?

The acronyms had conflicting meanings, lacked an owner-approved operating definition, or implied access to behavior that ordinary instrumentation could not observe. GeoZ retained the useful questions—coverage, source quality, follow-up needs, intent variants, and next-step behavior—but replaced the ambiguous formulas with explicit metric cards and an intent-resolution evidence ladder.