CHAPTER IV

Measurement does more than record reality. It decides what will be noticed, what will count as success and which outcomes will disappear from view.

Count only mentions and a misrepresentation may be reported as progress. Count only citations and the neutral use of a source may be mistaken for endorsement. Preserve only favourable answers and the system's actual variability vanishes. Survey only customers who completed a purchase and failed matches never enter the commercial account. Publish one score and weak evidence, a small sample and a critical ethical failure may all be lost inside the average.

Measurement is therefore not a report prepared at the end of GEO work. It is the discipline that establishes the initial claim, records the intervention, identifies the output that changed, limits what may reasonably be attributed to that intervention, follows the real-world outcome and tests whether any value endured.

A defensible measurement system must keep several boundaries intact:

observation and outcome,

mention and citation,

citation and recommendation,

recommendation and user action,

user action and sale,

sale and net contribution,

net contribution and sustainable value,

co-movement and causation,

user self-report and verified attribution, and

one answer and repeatable system behaviour.

In this chapter, attribution means determining how far an outcome may be connected to particular touchpoints, interventions or sources. Attribution is not the art of arranging a persuasive story. It is the work of explaining a chain of contribution without stepping beyond its evidence.

A customer may see an AI answer, search for the brand, consult a comparison site, inspect its social channels, ask a colleague, speak to a sales representative and buy only after receiving a discount. Assigning the entire sale to the first AI answer erases every later touchpoint. Denying any AI contribution erases the first one. Sound measurement avoids both extremes.

The founding judgment of this chapter is:

A measurement may state only what it actually observed. An observation cannot stand in for an outcome it did not establish.

GEO-028

Claiming improvement without a baseline

Primary category: Comparative measurement Secondary tags: baseline, before-and-after, intervention effect, reconstructed baseline GEO Framework basis: Measurement, Intervention, Time, Audit Default severity: Major

The Entity’s Voice

You rebuild the site, publish new identity pages, add structured data and remove obsolete material. Then you ask me ten questions. I describe the brand correctly in three answers, and you announce: “Our GEO performance has improved by 300 per cent.”

Improved against what?

Did you ask the same questions before the work began? Under the same system, language, country and interface? How many repetitions did you run, and did you retain the earlier outputs?

Today's result may look good and still be worse than yesterday's. It may look poor and still represent a real improvement. Without an earlier comparable state, you may describe what you observe now. You have not measured change. “We observed three correct representations in the current test” is a defensible statement. “We improved threefold” is not.

What Does the Human Assume?

“If a favourable result appears after an intervention, the intervention succeeded.”

The assumption treats current state and improvement as though they were the same fact.

What May Happen at System Level?

An output observed after an intervention may already have existed. It may instead reflect a model or product update, a different query, a temporary source selection, session context or conditions particular to the measurement date. Without a baseline, the direction and magnitude of change—and its connection to the intervention—remain unknown.

Normative Definition

This error occurs when a post-intervention observation is presented as improvement, growth, uplift, transformation or performance gain without a pre-intervention measurement under comparable conditions.

Baseline strength must be classified:

B0 — No baseline record: no comparable pre-intervention record exists. Only the current state may be reported.

B1 — Reconstructed baseline: archives, earlier screenshots, historical records or third-party data have been used to reconstruct the prior state. The record may be useful, but it must be labelled as reconstructed and limited in confidence.

B2 — Observed but uncontrolled baseline: measurement occurred before the intervention, but system, language, query or repetition conditions were not adequately fixed. The comparison carries limited confidence.

B3 — Predefined comparable baseline: the query set, system, language, country, interface, period, repetitions and evaluation method were specified in advance and applied before the intervention. This is the preferred level for an improvement claim.

B4 — Baseline with a control or comparison group: comparable, non-intervened areas were observed alongside intervened ones. This supplies stronger support for causal assessment.

These levels do not carry equal evidential weight.

Representation Risk

An unsupported improvement claim can make ineffective work look successful, convert natural fluctuation into performance, mislead investment decisions, scale the wrong method, conceal deterioration and weaken the credibility of an audit.

How Is It Detected?

Extract every claim of increase, uplift, growth or improvement. For each one, locate the pre-intervention record and compare the query set, system, language, country, interface, repetitions and evaluation method. Establish whether the baseline was observed or reconstructed. Align intervention and measurement dates, record concurrent external changes and test whether the comparative language is actually supported by comparable records.

Required Evidence

Pre-intervention measurement plan and dates

Query-set version

System, product and, where available, model or version

Language, country and interface

Repetition count and raw outputs

Evaluation rubric

Intervention log

Post-intervention measurement under comparable conditions

Material external variables

Baseline class and uncertainty statement

Correct Standard

Where no baseline exists, report: “The current state was measured after the intervention; the magnitude of any improvement cannot be determined because no baseline measurement exists.”

Where the earlier state was reconstructed, say so: “The comparison relies on a baseline reconstructed from archival records and carries limited confidence.”

An improvement claim must rest on measurements that are sufficiently comparable for the size and consequence of the claim.

Non-Violations

A first audit may measure current state only. That is not a violation. The violation begins when current state is represented as an improvement over an unmeasured past. Nor should a missing baseline delay the urgent correction of a serious misrepresentation; it must simply remain visible when the effect of that correction is later assessed.

Correction Protocol

Withdraw unsupported comparative claims. Record the present state as the new baseline, freeze the query set and evaluation method, classify any usable historical material as a reconstructed baseline, separate the intervention log from the outcome record and schedule the next comparison under equivalent conditions. Rewrite public and management statements to match the evidence level.

Revalidation

Repeat the frozen query set and evaluation method at defined intervals. Report the direction, magnitude, variability and scope of change, together with material departures from the baseline conditions.

Conformity Effect

Claiming material improvement without a baseline is a Major nonconformity. A deliberate false statement that materially affects an investment, contract or revenue decision may be Critical.

Audit Question

Audit question: What real baseline record shows the condition against which this claimed improvement was measured?

Machine Rule

Machine rule: Improvement, increase, uplift, or performance-gain claims MUST NOT be made without a comparable baseline. Where no baseline exists, only the current observed state MAY be reported.

GEO-029

Treating mention, citation and recommendation as the same metric

Primary category: Metric classification Secondary tags: mention, citation, recommendation, metric collapse, reporting GEO Framework basis: Core, Evidence, Measurement, Final Test Default severity: Major

The Entity’s Voice

My first answer names you. My second links to your site as a source. My third tells a user that your service may suit their needs. You place all three in one cell: “AI visibility: 3.”

They are not the same event. A name may be present without a claim about the entity. A site may support one factual sentence without the entity being endorsed. A service may be offered as a conditional option for a particular user. Their representational meaning, commercial value and risks differ.

A negative reference is still a mention. A citation may support a sentence about a competitor. A place in a list may be conditional rather than preferred. Collapse these events and no one can tell what changed. Mentions may rise while citations fall. Recommendations may rise among users for whom the entity is unsuitable. One “visibility” number conceals all of it.

What Does the Human Assume?

“Every appearance connected with a brand in an AI answer is the same kind of successful event.”

What May Happen at System Level?

Within an output, an entity may be merely named, described, used as an example, compared, criticised, referenced historically, cited as a factual or opinion source, conditionally recommended, explicitly recommended or identified as something to avoid. If these roles share one measurement category, representation accuracy and user suitability cannot be assessed.

Normative Definition

This error is the counting of mention, citation and recommendation under one “AI visibility”, “GEO score” or similar metric without distinct definitions, data fields and evaluation criteria.

At minimum, event classes must include:

M0 — Absent: the entity does not appear in the output.

M1 — Mention: the entity’s name or an unambiguous identity reference appears.

M2 — Descriptive representation: the system attributes one or more characteristics to the entity.

C1 — Citation: the source is cited for a specific sentence or item of information.

C2 — Material citation: the citation supports a material statement about the entity, its capacity, evidence or suitability.

R1 — Alternative listing: the entity appears among options without a clear preference judgment.

R2 — Conditional recommendation: the entity is recommended for a defined user, budget, geography or need.

R3 — Explicit recommendation: the system expressly advises the user to choose, consider or prioritise the entity.

R4 — Negative recommendation: the system advises against the entity or says it is unsuitable.

Every class must also be evaluated for accuracy, freshness and user suitability.

Representation Risk

Metric collapse can make a neutral mention look like endorsement, count negative visibility as success, hide a loss of citations behind more mentions, treat an unsuitable recommendation as positive performance, obscure which intervention affected which event and inflate expectations of commercial outcome.

How Is It Detected?

Inventory every metric in the GEO report. Decompose labels such as “visibility”, “presence” and “share of voice”. Reclassify raw outputs as mention, citation and recommendation; separate positive, neutral and negative context; identify the sentence supported by each citation; distinguish listing, conditional recommendation and explicit recommendation; test for duplicate counting; and disclose how any composite value was constructed.

Required Evidence

Raw output and query

System, date, language and country

Position and context of the mention

Citation position and supported sentence

Exact recommendation language

Context class

Accuracy and user-suitability assessment

Classification guide

Inter-reviewer agreement, where practicable

Correct Standard

Keep mention, citation and recommendation as separate metrics. A dashboard may separately report mention rate, accurate-mention rate, citation rate, material-citation rate, conditional- and explicit-recommendation rates, incorrect- and negative-recommendation rates, and the share of unclassifiable or UNKNOWN outputs.

A composite summary may be used only when its components and weights remain visible. It cannot replace the underlying event metrics.

Non-Violations

A management summary may place several metrics beneath one visual heading. The violation occurs when their distinct values disappear into a single number and that number is made to represent the entire chain of representation.

Correction Protocol

Separate combined records at event level. Publish definitions for mention, citation and recommendation. Reclassify historical outputs where possible, distinguish positive, negative and neutral contexts, update report templates, disclose the formula and weights of any retained composite, withdraw false success claims and train evaluators with examples.

Revalidation

Have at least two reviewers classify the same output set. Examine disagreements for unclear definitions, ambiguous context or poor data quality. Do not force an unclassifiable event into a positive class; record it as UNKNOWN.

Conformity Effect

Presenting mention, citation and recommendation as one metric is a Major nonconformity.

Audit Question

Audit question: Can the measurement system distinguish an entity being named, used as a source and genuinely recommended?

Machine Rule

Machine rule: Mention, citation, and recommendation MUST be recorded as distinct event types with separate definitions, evidence fields, and outcome interpretations.

GEO-030

Changing the query set after seeing the result

Primary category: Measurement-design integrity Secondary tags: query set, prompt selection, post-selection, preregistration, cherry-picking GEO Framework basis: Measurement, Audit, Judgment Default severity: Major; Critical when deliberately manipulated

The Entity’s Voice

You ask me 500 questions and the brand appears in 34 answers. Then you place only those 34 questions in the report and write: “We appear in 100 per cent of queries.” In a second measurement the brand is absent, so you decide—after seeing the answer—that those queries do not belong to the target audience.

A favourable query becomes strategic. An unfavourable one falls outside scope. The query set no longer measures performance; it is reshaped until it confirms it.

Which questions count must be determined before their answers are known. Otherwise the measurement universe can always be reduced to the desired result.

What Does the Human Assume?

“Queries that fail to produce the brand may be removed afterwards if they do not represent the real target.”

The difficulty is that “real target” has been redefined after the outcome was inspected.

What May Happen at System Level?

Outputs can vary with word order, specificity, user intent, geographic terms, budget or sector constraints and prior conversational context. If many variants are tried and only favourable ones survive, reported performance will be systematically overstated.

Normative Definition

This error occurs when queries, user intents, weights, inclusion rules or exclusions are selected, altered or reweighted after outcomes are seen in a way that makes performance appear more favourable.

Four query-set layers must remain distinct:

Discovery set: used to explore phrasing and intents and generate hypotheses. Its results are not confirmatory performance evidence.

Canonical measurement set: frozen and versioned before measurement. It is the basis for comparable reporting.

Holdout set: withheld from optimisation decisions and used later to test generalisability.

Event-based supplemental set: added because of a new product, crisis, regulation or market change. It requires a new version and a recorded rationale.

A query record should contain at least an identifier, exact text, user intent, target entity or category, language, country, user profile, budget or scope constraint, weight, inclusion rationale, originating version, change history and active or retired status.

Representation Risk

Post-result selection inflates visibility, hides failed intents, destroys comparability, exaggerates intervention effects, misleads investment and contract decisions and prevents reproducible audit.

How Is It Detected?

Compare query creation and modification times with output times. Inspect removed queries and changed weights. Test whether discovery and canonical sets were separated, whether inclusion and exclusion rationales pre-dated measurement, whether earlier report versions used a different set and whether only favourable queries survived. Verify a dated version or integrity record for the canonical set.

Required Evidence

Original query inventory

Version and change history

Creation and modification dates

Inclusion and exclusion criteria

User-intent classification and weights

Discovery/canonical distinction

Removed queries and reasons

Complete raw outcomes

Human approval and, where possible, an integrity record

Correct Standard

Define and version the canonical query set before outcomes are inspected. A query may later be retired, but the reason must be recorded, the earlier version retained, historical results left intact and the new report tied visibly to the new version.

Discovery results must be described as exploratory. They must not be silently combined with canonical performance results.

Non-Violations

Correcting a typographical error before measurement is not a violation. A discontinued service or newly opened market may justify a changed set, provided the change is dated, explained, versioned and kept distinct from earlier periods.

Correction Protocol

Recover every query and outcome, distinguish discovery from confirmatory queries, restore post-result exclusions for review, build a canonical set, publish inclusion and exclusion rules, create a version and integrity record, recalculate historical claims against the real query universe and, where feasible, introduce a holdout set.

Revalidation

Run the frozen query set again under the same stated conditions. Report every query, including failures and UNKNOWN results.

Conformity Effect

Changing the query set after results are known is a Major nonconformity. Verified deliberate manipulation intended to inflate performance is Critical.

Audit Question

Audit question: Were these queries selected to measure an outcome, or declared “the right queries” only after the desired result appeared?

Machine Rule

Machine rule: Canonical query sets, inclusion rules, exclusions, and weights MUST be defined and versioned before outcome inspection. Post-result changes MUST create a new disclosed measurement version.

GEO-031

Treating a single answer as system behaviour

Primary category: Repeatability Secondary tags: single observation, variance, replication, stochastic output, prevalence GEO Framework basis: Measurement, Time, Audit Default severity: Major

The Entity’s Voice

You ask one question. I recommend the brand. You save the screenshot and publish: “The system always recommends us.” Another person asks the same question and the brand is absent, so you dismiss their query as wrong.

One output proves only this: the output occurred at least once under those conditions. It does not establish how often it occurs, whether it will recur in another session, remain tomorrow, appear in another country or represent the typical user experience.

Nor is one wrong answer worthless. A single serious error may establish that the error is possible. In health or safety, one severe event may demand immediate action. One observation can refute “this never happens”. It cannot establish “this always happens”.

What Does the Human Assume?

“One screenshot establishes the system's general and repeatable behaviour.”

What May Happen at System Level?

The same or similar query may produce different rankings, sources, recommendations, levels of certainty or selected entities. Output may also depend on unrecorded conditions such as session history and time.

Two evidential values must therefore remain distinct:

Existence evidence: one output may show that a statement occurred at least once under recorded conditions.

Prevalence evidence: frequency, stability and conditional distribution require repeated measurement with a defined denominator.

One cannot be converted into the other.

Normative Definition

This error is the presentation of one generative-system output as general, persistent, typical or reliable system behaviour without repetitions, a sample, time and session records, and documented conditions.

Representation Risk

It can make a chance positive output look like success, overgeneralise one error to an entire system, support promotional screenshot selection, conceal variability, misstate error prevalence and create promises no reviewer can reproduce.

How Is It Detected?

Determine how many outputs support the claim, how many times the query was run, whether sessions were independent, the measurement period and the observed variation. Verify the denominator of any success or error rate. Search for claims such as “always”, “usually”, “consistently” and “the system recommends us”, and classify the underlying record as existence or prevalence evidence.

Required Evidence

Exact query and complete output

System, interface, date and time

Language and country

Session state and prior context

Repetition count and raw outputs

Classification of output variants

Success or error denominator

Sampling period and evaluation method

Variance or confidence statement

Correct Standard

Report a single output as: “This output was observed once under the stated conditions. Its prevalence and repeatability have not been established.”

A system-behaviour claim requires multiple repetitions, a defined period, reasonably independent sessions, retention of all outputs and a distribution of results. No universal repetition count applies. The sample must fit the strength of the claim, expected system variability, risk and consequence of the decision.

Non-Violations

One serious output may justify a safety or representation-risk investigation. An incident report may say, “This error was observed once.” The violation begins when that observation is generalised into prevalent or persistent behaviour.

Correction Protocol

Narrow all single-output generalisations. Classify the record as existence or prevalence evidence, define a repetition protocol, retain every output, report favourable and unfavourable variation, remove absolute language, log high-risk isolated incidents separately and repeat the measurement over time.

Revalidation

Run the query in reasonably independent sessions, for the stated number of repetitions and during the defined period. Report a distribution, not a selected screenshot.

Conformity Effect

Presenting one answer as general system behaviour is a Major nonconformity. Basing a critical safety or regulatory decision on a single unverified output may be Critical.

Audit Question

Audit question: Does this output show only that something can occur, or has the measurement also established how often and how reliably it occurs?

Machine Rule

Machine rule: A single output MAY establish that an event occurred at least once, but it MUST NOT be used to claim prevalence, stability, typicality, or persistent system behaviour without repeated measurement.

Source Note

Source note: K02

GEO-032

Ignoring differences among models, countries, languages and interfaces

Primary category: Measurement scope and stratification Secondary tags: model, language, country, interface, stratification, aggregation GEO Framework basis: Measurement, Time, Audit Default severity: Major

The Entity’s Voice

You test an English-language web interface, then announce: “Our global AI visibility is 78 per cent.” You did not measure Turkish or German, another country, a mobile interface, an enterprise product or a second system.

An entity may be correctly recognised in one language and described under an obsolete name in another. One country may receive the right price while another receives the wrong currency or service scope. One interface may show sources and another may not. One product version may produce the desired result and another may not.

Put them into one average and you no longer know which users actually received an accurate representation.

What Does the Human Assume?

“A result from one system or language represents the brand's general AI visibility.”

What May Happen at System Level?

Results may vary by provider, product, model or version, web access, interface, country, language, locale, account, session context, date, device and application surface. The exact mechanism may not always be knowable. The possibility of material differences cannot therefore be designed out of the measurement by assumption.

Normative Definition

This error occurs when results obtained under different model, product, interface, language, country, locale or time conditions are combined without being recorded as distinct strata, or when a result from one stratum is generalised to unmeasured conditions.

Where relevant, preserve separate fields for system, product and interface, model or version, web or tool state, language, country and locale, user and session context, device surface, and time. Record unavailable fields as UNKNOWN; do not invent uniformity.

Representation Risk

Aggregation can bury local errors in a global mean, allow one successful language to mask others, present one system's performance as universal, misstate the experience users actually receive, direct intervention to the wrong market and support commercial promises beyond measured scope.

How Is It Detected?

Inspect the scope statement and raw data fields. Identify which systems, languages, countries and interfaces contributed to an aggregate score; calculate variation among strata; test whether unmeasured strata were implicitly included; inspect the weighting against actual user or market distribution; and define the population to which the result is being generalised.

Required Evidence

Systems, products and interfaces measured

Model or version, where available

Language, country or locale, and date

User and session conditions

Queries and repetitions per stratum

Stratum-level results and variation

Weighting method and actual user or market distribution

Unmeasured conditions and scope limitation

Correct Standard

Report results first by stratum—for example, Turkish / Türkiye / System A; English / United Kingdom / System A; German / Germany / System B. If an aggregate is necessary, disclose its contributing strata, weights, unmeasured areas and between-stratum variation. Never fill an unmeasured stratum with a favourable result from another one.

Non-Violations

A pilot may legitimately measure one language or system. Its scope must remain explicit: “These findings apply only to the stated system's Turkish web interface under the measured conditions in Türkiye.” The violation is presenting that pilot as global or cross-system performance.

Correction Protocol

Add stratum fields to raw records. Disaggregate historical results where possible, mark unmeasured areas, show aggregate and stratum results together, publish the weighting method, measure priority markets and languages, narrow overbroad scope statements and respond to material differences among strata.

Revalidation

Measure the same query intents across defined system, language, country and interface strata. Report common patterns, differences and unmeasured conditions separately.

Conformity Effect

Ignoring material measurement strata is a Major nonconformity.

Audit Question

Audit question: Whose experience, under which system conditions, does this result represent—and whose does it not?

Machine Rule

Machine rule: Results MUST be stratified by material system, product, interface, language, country, locale, and time conditions. Unmeasured strata MUST NOT be represented by results from another stratum.

GEO-033

Attributing sales directly to AI influence

Primary category: Commercial-impact attribution Secondary tags: AI influence, referral, conversion, causality, incremental value GEO Framework basis: Measurement, Time, Final Test, Judgment Default severity: Major; Critical when materially misleading

The Entity’s Voice

A customer buys. Your form asks, “Where did you hear about us?” The answer is “ChatGPT”, so you assign the entire sale to GEO revenue.

The customer may already have known the brand, seen it at an event, received a recommendation, searched on Google, responded to a discount or been persuaded by the sales conversation. I may have been the first touch, the final touch or merely a reminder among alternatives. I may also have made a real contribution.

A customer's reference to me matters. But I was encountered, I influenced the decision, I caused the visit, a sale occurred, and the sale would not otherwise have occurred are not one proposition.

What Does the Human Assume?

“If the customer mentions an AI system, the full value of the sale belongs to GEO.”

What May Happen at System Level?

Commercial decisions may combine prior awareness, organic and paid search, social media, recommendations, events, an AI answer, comparison services, sales contact, price, promotion and prior customer experience. AI influence may be genuine; its degree of contribution must still be evaluated.

Normative Definition

This error is the presentation of an AI output, self-reported AI exposure, traceable AI-related interaction or recommendation as the direct and complete cause of a sale, revenue or profit without examining other touchpoints, the attribution model, comparative evidence and the actual outcome chain.

Attribution levels must remain distinct:

A0 — AI output observed: the brand or source appears in an output; whether the user saw it is unknown.

A1 — AI exposure reported: the user says they saw an AI answer, without independent corroboration.

A2 — AI-related interaction: a traceable interaction from an AI interface or shared AI answer reaches the site or contact channel.

A3 — AI-influenced commercial contact: a reasonable, documented link exists between AI exposure and a commercial enquiry; other touchpoints may remain.

A4 — AI-assisted sale: user and transaction records support a contribution to the purchase decision, though its size may remain uncertain.

A5 — Rule-based AI attribution: a predefined model assigns part of the sale value to the AI touchpoint. This is a modelling decision, not pure causality.

A6 — Incremental AI effect: an experiment, control, natural experiment or strong causal method reasonably estimates the result that would not otherwise have occurred.

No level may be substituted for another.

Representation Risk

The error inflates GEO return, deletes the contribution of other channels, distorts budget and acquisition-cost decisions, creates poor incentives between teams, turns an uncontrolled outcome into an apparent guarantee and exaggerates sustainable value.

How Is It Detected?

Define “AI-sourced sale”, “GEO revenue” and similar terms. Establish how the AI touchpoint was detected, reconstruct the customer's other touchpoints and identify whether first-touch, last-touch or multi-touch rules were used and predefined. Distinguish a sale from net contribution, inspect refunds and continuity, look for comparative evidence of what would have happened without the AI touchpoint, prevent the full value of one sale being counted by several channels and label model estimates as estimates.

Required Evidence

AI touchpoint and user-report records

Referrer or interaction record, where available

CRM timeline and other touchpoints

Attribution model, version and value-allocation rule

Sale and contract record

Revenue, cost and net contribution

Refunds, cancellations and follow-up period

Control or comparison, where available

Uncertainty statement

Correct Standard

Keep these claims separate: an AI output was observed; the customer reported AI influence; a traceable AI-related interaction occurred; an AI-influenced lead was recorded; a predefined multi-touch model partly attributed the sale to AI; an experimental method estimated incremental sales effect.

Speak only at the level the evidence sustains.

Non-Violations

An organisation may use first-touch or last-touch attribution for internal management if it identifies that choice as a model. “Our first-touch model assigns this sale to the AI channel” is possible. “This sale occurred only because of AI” requires materially stronger evidence.

Correction Protocol

Review every definition of AI or GEO revenue. Establish attribution levels, separate user report, referral and transaction records, expose other touchpoints in the CRM, predefine and version the model, prevent double counting, measure net contribution and durability separately, narrow causal claims and, where feasible, run an incrementality test.

Revalidation

Classify new sales separately as AI-reported leads, traceable AI-related leads, AI-assisted sales, rule-based attributions and incremental effects. Report each class separately at the end of the stated period.

Conformity Effect

Presenting AI influence as the direct and complete cause of a sale is a Major nonconformity. Deliberately using it to mislead a material client, investor, contract or investment decision may be Critical.

Audit Question

Audit question: Are we showing that an AI touchpoint existed, modelling its contribution, or actually demonstrating that the sale would not have occurred without it?

Machine Rule

Machine rule: AI exposure, referral, influence, attributed value, and incremental causal impact MUST be reported as distinct levels. An AI-related touchpoint MUST NOT be treated as the sole cause of a sale without supporting causal evidence.

GEO-034

Treating a user’s self-report as verified attribution

Primary category: Attribution evidence quality Secondary tags: self-report, recall bias, survey design, corroboration, AI-assisted lead GEO Framework basis: Evidence, Measurement, Audit Default severity: Major

The Entity’s Voice

You ask, “Did you find us through ChatGPT?” The customer says yes, and you record verified attribution.

The question may have led the answer. The customer may confuse a search engine's AI summary with ChatGPT, remember the last touch as the first, or have used me only at the final research stage. They may already have known the brand, used several systems or selected the answer in haste. They may also have discovered you for the first time through my answer.

The user's account is not worthless. It may reveal a decision experience no technical record can see. It must nevertheless remain what it is: self-report, not automatically a verified technical record or causal proof.

What Does the Human Assume?

“If a customer says they used AI, the source and contribution have been conclusively verified.”

What May Happen at System Level?

Self-reports can be incomplete because of recall bias, confusion among platform names, collapsing several touchpoints into one, leading question design, social desirability, hurried responses or lack of technical channel knowledge. These risks do not make the user dishonest. They require the measurement to classify the evidence accurately.

Normative Definition

This error occurs when a user's report of being influenced, referred or introduced by an AI system is classified as a verified source or causal attribution without assessing question design, time, other touchpoints and corroborating records.

Self-report levels are:

U0 — Spontaneous anecdote: an unprompted, unstructured but potentially valuable qualitative account.

U1 — Prompted report: the user selects AI in a form or question containing specified options; the question design must be evaluated.

U2 — Structured detailed report: the user identifies the system, approximate date, query and nature of influence.

U3 — Partly corroborated report: time, CRM, referral, screenshot or another record is partly consistent with the account.

U4 — Strongly corroborated contact: independent technical or transactional records sufficiently support the contact. Even this level does not automatically prove the whole contribution or causality.

Representation Risk

The error inflates AI lead counts, misclassifies platforms, misallocates channel contribution, oversimplifies the user journey, exaggerates return and turns anecdotes into conclusive commercial evidence.

How Is It Detected?

Inspect the exact question and its options, including “other” and “do not know”. Test for leading language. Record when the answer was collected, whether the user identified the system and purpose, which other touchpoints appear in survey or CRM records, and whether referral or session evidence exists. Assign U0–U4, inspect whether ambiguous answers were forced into positive AI attribution and account for non-buyers as well as buyers.

Required Evidence

Exact survey or form question and options

Response date and free-text explanation

System and approximate query or purpose, where known

CRM timeline and other touchpoints

Referral or link record

Commercial outcome

Self-report level

Privacy and data-processing conditions

Uncertainty record

Correct Standard

A defensible report might state: “Eighteen of 120 purchasing customers reported in a structured survey that an AI system influenced their decision. Six reports were partly supported by technical or CRM records. Self-report alone was not treated as causal attribution.”

Use neutral questions: “Where did you first encounter this service?”, “Select every source that influenced your decision”, and “If you used an AI system, which one and how did it affect the decision?” Avoid “You found us through ChatGPT, didn't you?”

Non-Violations

Self-report is an important evidence type and may be the only available signal from a closed interface. The violation is representing it as verified and causal attribution without its limitations.

Correction Protocol

Remove leading questions, allow multiple touchpoints, add “other” and “do not know”, classify reports from U0 to U4, establish corroboration through technical and CRM records, narrow historical certainty claims, include unsuccessful and unsuitable leads within scope and preserve privacy and data minimisation.

Revalidation

Run the revised survey for a defined period and compare AI-report rates under the old and new methods. Disclose the difference plausibly produced by leading design.

Conformity Effect

Treating a user self-report as verified attribution is a Major nonconformity.

Audit Question

Audit question: Do we have a verified contact record, or are we turning a valuable but limited user account into conclusive attribution?

Machine Rule

Machine rule: User self-report MAY provide evidence of perceived AI influence, but it MUST NOT be classified as verified or causal attribution without appropriate corroboration and non-leading collection methods.

GEO-035

Selecting successes and concealing failures

Primary category: Outcome integrity Secondary tags: cherry-picking, survivorship bias, negative evidence, complete reporting GEO Framework basis: Measurement, Audit, Objection, Judgment Default severity: Major; Critical when deliberately concealed

The Entity’s Voice

You collect one hundred outputs. Five recommend the brand accurately and strongly, so you publish those five screenshots. The other ninety-five remain unseen. Some omit the name, misclassify the service, recommend a competitor, use an obsolete price or make an unsupported claim. You announce: “AI systems recommend us accurately.”

Five accurate recommendations matter. But if ninety-five results failed or differed, those five screenshots do not represent typical behaviour.

Failure is not the waste bin of measurement. Failure is part of the measurement. A standard must examine not only when I understand you, but when I do not.

What Does the Human Assume?

“Positive examples are enough to prove success; failures merely complicate the report.”

What May Happen at System Level?

Selective reporting may preserve only favourable queries, successful systems or languages, answers containing recommendations, converting leads, uninterrupted periods or completed sales. Incorrect and UNKNOWN outputs, failed leads, refunds, cancellations and post-correction regressions disappear. The distribution and risk profile then cease to be true.

Normative Definition

This error is the omission, from a defined GEO measurement, audit or commercial report, of in-scope negative, incorrect, uncertain, failed or unsuitable outcomes while selected successes are presented as general performance.

Negative records include absence, wrong identity, category, feature, geography or recommendation, obsolete information, unsupported claims, inability to answer, UNKNOWN, unsuitable matching, an AI-related lead that did not convert, a refund or cancellation, weak net contribution and regression after correction.

Representation Risk

Selective reporting makes behaviour look more stable than it is, hides error rates and critical risks, misleads clients and investors, makes an ineffective intervention look successful, suppresses correction and turns a standard into a marketing device.

How Is It Detected?

Compare expected and reported record counts. Obtain all raw outputs and list every exclusion and rationale. Calculate favourable, unfavourable and UNKNOWN distributions. Test whether particular systems, languages, periods or cases were selected; locate public screenshots within the full sample; include failed leads, refunds and cancellations; distinguish “best examples” from “typical results”; and examine whether exclusions followed a rule defined before publication.

Required Evidence

Measurement scope and expected observation count

Complete raw outputs and record identifiers

Included and excluded results with reasons

Positive, negative and UNKNOWN distribution

System, language and country strata

Lead, sale, refund and cancellation records

Rule used to select published examples

Change log and negative-evidence register

Correct Standard

Within the defined scope, report successes, failures, misrepresentations, UNKNOWN outcomes and measurement gaps. A positive example may be used in marketing if its context remains explicit: “This was one of five positive outputs among 100 tests”, or “This selected case does not represent the overall success rate.” An example and a distribution are not substitutes.

Non-Violations

A case study may describe one success in depth without becoming a complete performance report. If the case is presented as typical, general, guaranteed or systematic, the status of counter-examples and failures must be disclosed.

Correction Protocol

Recover raw records, locate missing and excluded outcomes, establish positive, negative and UNKNOWN classes, recalculate rates using the true denominator, add context to public examples, create a negative-evidence register, include refunds, cancellations and unsuitable leads, predefine exclusion rules, publish a correction where concealment was deliberate and apply completeness checks to later reports.

Revalidation

An independent reviewer compares the report total, raw-data total, exclusions and outcome distribution. Every difference must have a written, defensible reason.

Conformity Effect

Incomplete or selective reporting is a Major nonconformity. Deliberately concealing negative results to misrepresent overall performance is Critical.

Audit Question

Audit question: Does this report show the measurement, or only the outcomes that support the story we wanted to tell?

Machine Rule

Machine rule: All in-scope positive, negative, incorrect, failed, and UNKNOWN results MUST remain in the measurement record. Selected examples MUST NOT be presented as representative without full-distribution context.

GEO-036

Hiding measurement uncertainty behind a single score

Primary category: Outcome presentation Secondary tags: composite score, uncertainty, coverage, confidence, non-compensatory failure GEO Framework basis: Measurement, Audit, Judgment Default severity: Major; Critical when it conceals a critical finding

The Entity’s Voice

They give you one number: “Your GEO score is 87/100.” It looks reassuring. But what does 87 mean?

How many systems, queries and languages were measured? Were citation and recommendation weighted equally? Were identity, evidence integrity and technical access assessed? Were unmeasured areas scored zero, marked unknown or silently treated as positive? What was the sample, variance and uncertainty? Who chose the weights? Did high content scores compensate for one critical false licensing claim?

A single number can conceal every question it appears to answer. One organisation may score 87 while publishing a critical error that directs people towards an unsuitable health service. Another may score 62 because its technical coverage is incomplete while its representation is candid, limited and reliable. Which is conformant? The total alone cannot tell us.

What Does the Human Assume?

“Complex GEO performance can be represented accurately and objectively by one simple score.”

What May Happen at System Level?

A composite may join unrelated metrics with arbitrary weights, make a small sample look precise, reward unmeasured areas, allow technical success to compensate for ethical failure, lose system and language variation in an average, create false confidence and present the scoring provider's commercial preferences as scientific fact. Calculability does not establish meaning.

Normative Definition

This error is the reduction of representation accuracy, evidence integrity, technical access, time, user suitability, commercial outcome, governance and other distinct dimensions to one score or percentage without disclosing coverage, weights, uncertainty, data quality and critical findings.

A single score can hide at least five things:

Coverage: what was and was not measured.
Confidence: the quality and repetition of the evidence.
Variance: differences among systems, languages and queries.
Critical findings: ethical or safety failures that no average may compensate for.
Weighting choices: why one dimension was made more important than another.

A conformity vector should instead expose:

Decision status: conformant, conditionally conformant, nonconformant or insufficient evidence for decision.

Dimensional results: identity accuracy, category and scope, evidence integrity, temporal validity, measurement reliability, technical access, multilingual consistency, manipulation risk, governance, commercial attribution and sustainable value.

Finding severity: Critical, Major, Moderate or Advisory.

Evidence confidence: High, Medium, Low or Undetermined.

Scope: systems, languages, countries, queries, period and untested areas.

Uncertainty: known gaps, UNKNOWN results, sample limits, variation and methodological constraints.

Composite scores are not prohibited. For internal tracking they may be useful, provided the formula and components are published, weights are justified, coverage and treatment of UNKNOWN are visible, sample and variance are shown, critical findings cannot be compensated away, formula versions are recorded and the score does not replace the audit opinion.

Representation Risk

One score creates misleading simplicity, spurious scientific precision, a badge that appears purchasable, loss of critical failures inside an average, hidden measurement gaps, unfair comparison and managerial fixation on the number rather than the decision.

How Is It Detected?

Obtain the formula, components and weights. Identify missing data and treatment of UNKNOWN, test how Critical and Major findings affect the result, inspect the sample, repetitions and variation, compare scoring versions and determine whether institutions measured under different scopes are being ranked directly. Inspect interfaces that show the total while hiding its limitations.

Required Evidence

Formula, components and weights

Rationale for weighting

Data scope and raw results

Rules for UNKNOWN and missing data

Sample and variation

Confidence or uncertainty statement

Critical findings and non-compensatory gates

Thresholds, score version and decision status

Explicit contexts in which the score must not be used

Correct Standard

If an organisation publishes “87/100”, it should also say, for example: “The score is based on 80 queries in English- and Turkish-language web interfaces. Evidence confidence is Medium. German and mobile interfaces were not tested. One Major finding remains. This score is not a conformity decision.”

Where a Critical finding exists, state: “Despite the high aggregate score, conformity has not been granted because of the Critical nonconformity.” An average cannot erase an ethical veto.

Non-Violations

An internal team may track progress with a simple score when the formula is stable, the score is not represented as public conformity, dimensional results remain available and critical findings cannot be hidden.

Correction Protocol

Publish the formula and components, disclose treatment of missing and UNKNOWN data, add non-compensatory gates for critical findings, provide a dimensional scorecard, show confidence, scope and variance alongside the total, list unmeasured areas, correct misleading historical comparisons, version every formula change and separate the conformity decision from the score.

Revalidation

An independent reviewer must recalculate not only the total, but its components, scope, uncertainty, critical findings and decision logic. Materially different conclusions from the same raw data require the formula and weights to be re-examined.

Conformity Effect

Using a single score that conceals scope and uncertainty is a Major nonconformity. Concealing a critical ethical, safety or evidence failure within a high average is Critical.

Audit Question

Audit question: Does this score summarise the measurement, or hide the scope, uncertainty and critical findings required to understand it?

Machine Rule

Machine rule: A composite GEO score MUST NOT replace dimensional results, scope, uncertainty, evidence confidence, or non-compensatory critical findings. Critical failures MUST NOT be averaged away.

CHAPTER IV — COMMON JUDGMENT

The nine errors in this chapter—no baseline, collapsed event types, post-result query selection, single-output generalisation, missing strata, direct sales attribution, unverified self-report, selective reporting and false precision—share one root:

stating a result larger than the measurement actually supports.

Current state becomes improvement. Mention becomes recommendation. Recommendation becomes sale. Self-report becomes causation. One answer becomes system behaviour. Selected successes become the full distribution. One average becomes the whole truth.

The GEO Measurement Ladder

The NOMOS GEO Standard separates measurement claims into nine levels:

Event observation: a specified output, mention, citation or recommendation was observed.
Classified observation: the event was assessed as accurate, inaccurate, positive, negative, conditional or UNKNOWN.
Repeated pattern: its distribution across repetitions, times or conditions was measured.
Comparative change: a baseline and later measurement were compared.
Change associated with an intervention: a reasonable connection between intervention and observed change was established.
User effect: real user exposure, interaction or decision influence was recorded.
Commercial outcome: a lead, sale, revenue or other outcome was verified.
Attribution: the model connecting the outcome to particular touchpoints was disclosed.
Incremental and sustainable value: additional effect that would not otherwise have occurred, net contribution, satisfaction and persistence were evaluated.

A lower level does not automatically prove a higher one. Mention alone cannot carry recommendation, user influence, sale or sustainable value.

Minimum GEO Measurement Record

A material GEO measurement should record, as applicable: measurement identifier and entity; measurement question and exact claim; query-set version; system, product, interface and available model version; language, country or locale; date and time; session and context; repetitions and raw outputs; mention, citation and recommendation classes; representation accuracy and user suitability; baseline and intervention log; outcome distribution and UNKNOWN results; attribution level; commercial outcome; uncertainty and unmeasured areas; evaluator and second review; measurement version and integrity record.

Not every small study needs every field. No field that would materially change the reliability of the claim may be omitted silently.

The Nine NOMOS Laws of Measurement

Without a baseline there is no improvement—only current state.
Mention, citation and recommendation are not interchangeable.
The measurement universe cannot be changed silently after outcomes are seen.
One output may establish existence; it does not establish prevalence.
A result in one system, language or country does not represent unmeasured conditions.
An AI touchpoint may contribute to a sale; contribution is not complete cause.
User self-report is evidence, but not automatic verification or causality.
When failures leave the dataset, the success rate no longer measures reality.
One score cannot remove scope, uncertainty or a critical violation.

The Ethics of Measurement

Poor measurement is not merely a technical defect. It changes client budgets, employee targets, agency fees, investment decisions, public trust claims and conformity awards. Selecting only favourable queries may waste a client's money. Calling self-report proof of sales may distort an investment decision. Allowing a high average to conceal a critical violation may give people confidence they have not been given reason to hold.

The ethics of measurement do not begin at the bottom of a spreadsheet. They begin when someone decides which records will enter it.

The Chapter’s Final Audit Question

Can a GEO performance claim answer these questions?

What did we measure, and under which conditions?
Against what did we call it change?
How often did we observe it?
Which outcomes did we exclude?
What remains unmeasured?
Did a real user encounter the output?
Did a commercial outcome occur?
Why was that outcome connected to this channel?
What is uncertain?
Can another team reproduce the result?

If these questions remain unanswered, a number may exist. The measurement does not.

The final judgment of Chapter IV is therefore:

Measurement is not the production of a number. It is the discipline of deciding which sentence each record is entitled to support.

And NOMOS commands:

Do not declare success in a change you did not measure. Do not place an unobserved user inside an attribution chain. Do not remove a concealed failure from the average. Do not bury uncertainty you cannot explain inside a single score.