← Back to GEO LibraryMETHOD / REPRODUCIBILITY

GEO measurement methodology: a reproducible AI-search visibility protocol

A credible GEO audit is a versioned research exercise: define the sample, repeat the observations, preserve the evidence and state what the result cannot prove.

Updated: 02/09/2026
NobleJackal Research: Kaan MURAZ — CEO, NobleJackal
Publisher: NobleJackal
Version: 1.1
Source review: Primary sources reviewed

Establish the technical baseline

Record crawler directives, status codes, canonicals, hreflang, sitemap coverage, structured data and visible source pages before testing model responses.

Design the prompt sample

Group prompts by brand, category, problem and comparison intent. Fix the language, geography and evaluation date so later runs can be compared honestly.

Publish the evidence trail

Store prompt, response, citations, screenshots or exports, and the audit version. Separate confirmed facts from interpretation and list any inaccessible systems or unknowns.

DIRECT ANSWER / MEASUREMENT STANDARD

A defensible GEO measurement programme uses a fixed prompt panel, repeated observations and separate outcome layers.

GEO cannot be measured honestly with one search, one screenshot or one proprietary score whose ingredients remain hidden. Generative answers vary by platform, product surface, model version, language, location, account state and time. The practical unit of evidence is therefore a dated observation: one defined prompt, run in a documented environment, with the answer, sources and classification retained. A useful programme combines many such observations under a versioned protocol.

The method begins before an answer is generated. Technical eligibility, verified retrieval, search impressions, brand mentions, linked citations, accurate descriptions, recommendations, referral sessions and commercial outcomes are different events. Each has its own evidence source and denominator. Combining them too early turns diagnosis into theatre: a crawler test cannot stand in for indexing, and a citation cannot stand in for a qualified enquiry.

NobleJackal's method uses a stable core panel and a smaller rotating discovery panel. The core allows like-for-like comparison; the rotating panel catches new language and market behaviour. Important prompts are repeated, raw evidence is preserved and results are reported as distributions and rates, with sample size and missing observations shown. The report can guide work, but it never converts an observed sample into a universal ranking claim.

EIGHT OUTCOMES / EIGHT EVIDENCE SOURCES

Keep the discovery chain intact instead of hiding it inside one score.

A result can advance at one layer while remaining unchanged at the next. Report the strongest confirmed layer and preserve the gaps.

StageOutcomeAcceptable evidenceIt does not prove

01Technical eligibilityPublic HTTP response, robots policy, canonical, index controls, rendered text, language signals and supported structured data checked on the live URL.That a search or answer platform has fetched, indexed or selected the page.

02Verified retrievalA genuine platform record, verified crawler log, webmaster-tool crawl record or a controlled user-requested fetch with the final body retained.Indexing, inclusion in an answer, a citation or favourable treatment.

03Search visibilityPlatform-owned performance data showing impressions, pages, countries, devices, queries or grounding-query samples for the defined period.That the page supplied the wording of an AI answer or that a user considered the result useful.

04Brand mentionThe named entity appears in the captured answer, with prompt, platform, date, locale and run identifier preserved.Which source produced the name, whether the statement is accurate or whether the brand was endorsed.

05Linked citationA visible source link or platform citation points to the canonical domain or page in the retained answer state.Prominence, answer-level influence, recommendation or repeatability across runs.

06Accurate useThe answer's claims are compared with the cited page and a dated factual reference; material errors and unsupported additions are recorded.That the system will repeat the description or select the source for another intent.

07Contextual recommendationThe brand is explicitly recommended for a defined need and market, and the stated reasons are accurate in the captured response.A durable preference, a global rank, purchase intent or a customer relationship.

08Commercial outcomeConsent-respecting analytics, a qualified enquiry record and the organisation's own CRM or sales evidence connect the visit to a defined outcome.Sole causation; other touches, attribution rules and untracked journeys may have contributed.

PROMPT PANEL / EXPERIMENT DESIGN

Build the sample before seeing the answers.

A panel is useful when it represents real decisions and remains stable enough to compare. It is not a bag of phrases chosen after the brand appears.

01
Intent families
Separate definition, problem, category, comparison, supplier-selection and brand-validation prompts. A brand query answers a different business question from an unbranded buyer query.
Keep a declared quota per intent and report each family separately.
02
Market and geography
Name the country or city only where location changes the user's need, law, language, availability or delivery conditions. Do not multiply place names merely to manufacture a larger sample.
Record requested geography, actual test location and any location setting the product exposes.
03
Language and phrasing
Write prompts as a competent person would ask them in each language. Include a small number of natural variants, but do not translate one English sentence word for word and call the result multilingual demand.
Lock the approved wording and keep locale-specific editorial notes with the panel version.
04
Audience and constraints
Where recommendations are tested, state the buyer type, scale, budget boundary or technical condition that makes an answer useful. Vague prompts create broad lists that are difficult to judge.
Use the same constraints in baseline and comparison runs; change one factor at a time.
05
Core and discovery sets
The core set remains unchanged for trend analysis. A smaller discovery set samples emerging terms, current questions and new product behaviour without rewriting history.
Never replace weak core results with more favourable discovery prompts in the headline rate.
06
Eligibility and exclusions
Define whether logged-out surfaces, paid plans, region-limited products, answer refusals, unavailable citations and safety blocks belong in scope. Missing access is data, not a zero invented after the fact.
Publish inclusion, exclusion and missing-data rules before collection starts.

SIX-PASS PROTOCOL

Make another reviewer capable of repeating the work.

01 / REGISTER

Freeze the question, scope and decision rule.

Assign the study an identifier and version. Record the site, entities, competitors, locales, markets, platforms, product surfaces, prompt panel, run count, collection window and evaluator. State the business question: discovery, source selection, factual portrayal, recommendation or demand. Without this, the same chart will be used to answer incompatible questions.

Write the classification rules before collecting responses. Decide how aliases, subdomains, publisher pages, PDF citations, link cards and repeated citations are counted. Define a qualified recommendation and a material factual error with examples. Predefined rules reduce the temptation to reinterpret an awkward answer after seeing it.

02 / BASELINE

Capture the technical and platform-owned baseline.

For each target route, record live status, final URL, robots result, canonical, index directive, rendered answer text, language alternates, sitemap membership and source accessibility. Save the evidence date and response signature. A local build or deployment log is not a substitute for the public response.

Export available first-party reports for the same period. Google Search Console and Bing Webmaster Tools use their own counting and aggregation rules; retain those definitions. Search Console clicks are not Analytics sessions, and Bing's citation count does not describe the placement or influence of a page inside an answer.

03 / COLLECT

Run a controlled, repeated observation schedule.

Run every core prompt the declared number of times on each included platform and surface. Spread repetitions across the stated window rather than firing them in a single minute. Preserve the exact prompt, full answer, visible citations, destination URLs, timestamp, locale, test location, account state and any product version that is exposed.

Do not silently regenerate an inconvenient answer. Retries are allowed only under a written failure rule, for example a network error or an explicit service failure. Keep the original failure record, the retry reason and the new run identifier. This prevents selective sampling from masquerading as improvement.

04 / CLASSIFY

Label each observation at claim and source level.

Mark whether the entity was absent, mentioned, cited, accurately described or contextually recommended. Resolve citation redirects to the final canonical URL and distinguish the organisation's own domain from third-party pages that merely mention it. Count a source only under the rule fixed in the register.

For cited answers, inspect whether the cited page supports the nearby claim. Source selection and answer absorption are not the same: a page may appear in the source tray while contributing little visible substance, or it may support a key definition. Record the distinction rather than assuming every citation has equal weight.

05 / ANALYSE

Calculate rates with denominators, uncertainty and segments visible.

Report the number of eligible observations beside every percentage. Show platform, locale, market, intent and device or surface segments when the sample can support them. For repeated runs, publish the distribution: how often the outcome appeared, how variable it was and whether the direction persisted across collection days.

Compare like with like. A changed prompt, product mode, geography, login state or collection window is a new cohort, not a clean before-and-after. Treat small movements cautiously, particularly when the panel is small or missing responses cluster on one platform.

06 / CONNECT

Join visibility to behaviour and revenue without inventing causation.

Review referral and landing-page sessions in consent-respecting analytics, then connect qualified enquiries to the organisation's own intake and CRM records where a lawful identifier exists. Preserve source, medium, landing page, conversion definition and attribution window. Direct traffic can include journeys whose referrer was unavailable, so label it honestly.

Close the report with confirmed findings, plausible interpretations, unresolved questions and the next test. A visibility lift matters commercially only when it reaches the appropriate audience and helps produce a valuable action. Until then it is an upstream signal, not revenue proof.

RATE CARD / EXACT DENOMINATORS

Use simple measures that a reader can recalculate.

Rates should be reported by cohort and with the raw count. A metric without its denominator conceals more than it explains.

MetricNumeratorDenominatorInterpretation boundary

Mention rateEligible runs in which the defined entity appearsAll eligible runs in the same cohortEntity presence only; no source or sentiment is implied.

Citation rateEligible runs with at least one qualifying citation to the entity's domain or accepted source setAll eligible runs in the same cohortSource selection, not placement, absorption or endorsement.

Citation shareQualifying citations attributed to the entityAll qualifying citations recorded for the declared competitor setSensitive to source-count behaviour and competitor-set design; never a universal market share.

Accurate-description rateEntity-present runs with no material factual error under the written rubricAll eligible runs in which the entity appearsAccuracy within captured answers, not completeness or future stability.

Recommendation rateEligible buyer-intent runs containing a qualifying contextual recommendationAll eligible buyer-intent runs in the same cohortDepends on the stated need; it is not a global supplier ranking.

Qualified-action rateQualified actions attributed under the declared analytics and CRM ruleMeasured referral or landing sessions included by that same ruleAn attribution result, not proof that GEO alone caused the action.

VARIANCE / UNCERTAINTY

Treat answer instability as part of the result.

Repeated measurement does not remove uncertainty; it makes the uncertainty visible enough to manage.

Repeat count

Declare repetitions per prompt, platform and collection day. One result is an example; repeated results allow a rate and a distribution. Increase repetition for high-value or highly variable intents.

Time and version

Retain exact timestamps and any visible model or product version. Index refreshes, interface experiments and model updates can change answers without a site change.

Location and account state

Logged-in history, subscription tier, interface language and apparent location may affect access or output. Record them; do not describe one test state as the experience of every user.

Prompt sensitivity

Small wording changes can alter retrieval and answer composition. Keep the core wording fixed and analyse natural variants as a separate segment rather than pooling them invisibly.

Missing observations

Timeouts, refusals, unavailable features and blocked citations require their own status. Excluding failures only from the weaker cohort can make performance look better than it was.

Reviewer agreement

Accuracy and recommendation labels involve judgement. Use a written rubric and, for material studies, a second review or documented adjudication of disputed cases.

MINIMUM EVIDENCE PACKAGE

A published result should travel with six records.

The evidence package is what separates an auditable finding from a sales claim.

Protocol manifest

Study ID, version, objective, owner, dates, included platforms, locales, markets, prompt-panel hash, run policy and classification rubric.

Prompt register

Stable prompt ID, exact text, intent family, audience, market, language, constraints, inclusion status and change history.

Raw observation ledger

Run ID, timestamp, environment, full answer or permitted export, citation targets, retrieval failures, screenshot or evidence reference and integrity signature where practical.

Classification table

Mention, citation, accuracy, recommendation and error labels with reviewer, rule reference, notes and any adjudication.

Cohort summary

Raw counts, denominators, rates, variation, missing data and segments; baseline and comparison cohorts remain distinct when their scope differs.

Outcome boundary

What was confirmed, what is inferred, what remains unknown and which later evidence would be needed to establish indexing, answer use, demand or revenue.

MEASUREMENT RED FLAGS

Eight shortcuts that manufacture confidence rather than knowledge.

A polished dashboard is not a method if the underlying observations cannot be inspected.

One prompt becomes a market conclusion

A single response cannot represent every wording, user, location, model state or date. Keep it as an observation, not a trend.

The denominator disappears

A percentage without eligible-run count, missing-run count and cohort definition cannot be audited or compared.

Prompts change after the baseline

Replacing weak questions with easier ones creates a different sample. Preserve the core panel and version legitimate additions.

Mention, citation and recommendation collapse

These outcomes answer different questions. Adding them into one visibility total erases the very diagnosis the report is meant to provide.

All citations receive equal weight

A source-tray link may support a minor detail or the core answer. Record selection and material use separately when the product allows inspection.

Synthetic bot access becomes platform proof

A request using a crawler name tests the delivery stack for that string. It does not prove a genuine visit, index record or answer selection.

Clicks and sessions are forced to match

Search Console and analytics use different systems, aggregation and attribution. Reconcile trends and explain differences instead of editing them away.

Correlation becomes causation

A later citation, enquiry or sale may follow many changes and touches. Use cohorts and controlled tests where possible, and describe the remaining uncertainty.

CONNECTED EVIDENCE

Move from outcome definitions to collection, diagnosis and commercial action.

These NobleJackal records keep measurement connected to technical eligibility, answer evidence and the service decision it is meant to support.

QUESTIONS / CLEAR ANSWERS

What should a credible GEO report answer?

01Is there one standard GEO visibility score?

No official universal score covers every search and answer platform. A provider may create a composite indicator, but it should disclose the component events, weights, prompt sample, repetitions, platform coverage, missing-data rule and date. The underlying rates should remain visible so another reviewer can recalculate them.

02How many times should each AI-search prompt be run?

There is no universal number that makes every study representative. The protocol should declare a repeat count before collection, distribute runs across the stated window and increase repetitions where outcomes are valuable or unstable. Publish the count and variation; do not imply that repetition eliminates platform or sampling bias.

03Can Google Search Console measure AI visibility?

Google provides platform-owned reporting for appearances within its generative AI features, including impressions, pages, countries, devices and dates in the current reports. That evidence covers Google's supported surfaces under Google's counting rules. It does not measure every answer platform, disclose the private selection algorithm or by itself prove a conversion.

04Does a citation mean the page influenced the answer?

Not necessarily. A visible citation confirms source selection in that captured state. The cited page may support a central claim, a minor detail or little visible language at all. Where evidence permits, assess claim support and answer-level use separately from citation count.

05When can GEO activity be linked to revenue?

Only when a lawful analytics and business record connects an included visit or enquiry to a defined commercial outcome under a stated attribution rule. Even then, report contribution rather than automatic sole causation unless the design can isolate the effect. Mentions and citations remain upstream indicators until that evidence exists.

Practical checklist

Version the audit scope
Preserve technical and response evidence
Disclose sampling limits
Compare like-for-like reruns