How to interpret and report an AI visibility audit

Collecting answers is the easy part of an AI visibility audit. The value lies in how those answers are classified, verified and converted into decisions.

A report based only on mention share, screenshots or one visibility score may hide the most important findings. A brand can be frequent but inaccurate, cited but not recommended, or consistently represented through an outdated claim.

The analytical task is to determine whether the brand appears, what role it receives, which visible sources support it, whether the claims are accurate and whether the representation persists.

These are the five dimensions of the 5P AI representation audit model: Presence, Position, Provenance, Precision and Persistence. The testing procedure is covered separately in How to run an AI visibility audit across AI search platforms.

The value of an audit is not the volume of collected answers. It is the quality of the classification and the decisions that follow from the evidence.

Classify every answer across the five dimensions

Each answer needs separate classifications. Labels such as “positive” or “visible” are too imprecise for diagnosis.

Presence classification

Presence establishes whether the brand or a related entity appears.

Useful statuses include:

• brand present;

• product present without the parent brand;

• domain cited without a brand mention;

• brand absent;

• ambiguous name;

• wrong entity present.

Record whether presence was prompted or spontaneous: explicit brand questions test recognition; non-branded questions test discovery.

A domain in a source panel, a passing mention and shortlist inclusion are different outcomes.

Position classification

Position describes the role assigned to the brand.

A practical taxonomy includes:

• primary recommendation;

• secondary recommendation;

• shortlist inclusion;

• category example;

• information source;

• background mention;

• caution or negative comparison;

• irrelevant inclusion;

• incorrect category;

• excluded despite explicit fit.

Position should capture the use case, category, comparison set and stated advantages or limitations, not only textual order.

A brand that appears first but is described as unsuitable does not have a stronger result than a brand appearing later as the preferred option for the user’s criteria.

Provenance classification

Provenance analyses the visible source environment.

Classify sources as, for example:

• owned;

• earned media;

• institutional or governmental;

• academic;

• partner;

• directory;

• review platform;

• forum or social;

• competitor;

• unidentified or unavailable.

For each visible source, assess:

• entity match;

• publication or update date;

• source type and ownership;

• whether it supports the relevant claim;

• whether the answer extends beyond the source;

• whether several claims depend on one domain;

• whether contradictory sources are present.

Visible provenance matters because a public interface does not reveal every retrieved page or the complete generation process. The audit evaluates shown sources, not hidden retrieval.

Citation count alone is insufficient. The preprint From Citation Selection to Citation Absorption separates source selection from the degree to which a cited page contributes evidence, language or structure to an answer. Across the authors’ dataset, citation breadth and measured influence did not move together consistently. The practical implication is straightforward: count sources, but also verify what they support.

Precision classification

Precision evaluates factual and semantic accuracy against the audit’s verified claim map.

The unit of analysis should be the atomic claim, not the entire answer.

For example:

“Brand X is a German project-management platform with predictive financial forecasting included in every plan.”

This statement contains several independently testable claims:

• country association;

• category;

• feature availability;

• pricing or plan coverage.

One may be correct while the others are false or outdated.

The DeepTRACE framework uses statement-level decomposition and citation-support matrices to audit whether generated claims are supported by listed evidence. The study focuses partly on debate and deep-research queries and uses a model-based judge validated against human ratings, so its numerical results should not be transferred mechanically to brand audits. Its claim-level approach is nevertheless a useful methodological precedent.

A practical error taxonomy

Also use contradictory for internally conflicting answers and unverifiable where evidence is insufficient. Neither is automatically a hallucination.

Do not overuse the term hallucination

A hallucination should refer to a false or fabricated claim for which the system has no adequate factual basis. It should not become a catch-all label for every absence, weak recommendation, omitted detail or difference in wording.

A brand may be absent because it does not meet the criteria, the scenario is broad, the answer is limited to a few examples or the system selected a different interpretation. A claim may be unverifiable because the brand itself has not published clear evidence.

The error register should distinguish output errors from weak or conflicting source conditions.

Persistence classification

Persistence measures whether a representation holds across repeated runs, variants, platforms, languages and time.

Report stability separately for:

• presence;

• recommendation role;

• competitor set;

• cited domains;

• individual claims;

• language versions;

• product surfaces;

• measurement periods.

Research on repeated sampling supports this separation. Quantifying Uncertainty in AI Visibility found substantial citation variability across repeated measurements, while Don’t Measure Once argues that visibility should be understood as a distribution rather than a one-off observation.

Do not interpret high stability as proof of quality. A wrong category or invented feature may recur consistently.

A stable error is a more persistent representation risk, not a more accurate answer.

A hypothetical B2B audit example

Assume that the audited brand provides a B2B platform for analysing operational processes.

The verified claim map establishes that:

• the product is available in Europe;

• it integrates with several ERP systems;

• it offers advanced access controls;

• it does not include a standalone financial-forecasting module;

• it is not project-management software.

The tests produce the following observations:

A shallow report might state that “the brand appears in 60 per cent of answers”.

The 5P interpretation is more useful:

• Presence: moderate but uneven across scenarios;

• Position: rarely the primary recommendation;

• Provenance: dependent on different source types by platform;

• Precision: affected by an invented feature and wrong category;

• Persistence: core facts are stable, while recommendations and sources vary.

The priority is not simply more mentions, but clearer category signals, corrected external descriptions, a defined functional scope and monitoring of the invented feature.

Turn findings into prioritised recommendations

Every recommendation should connect a problem to evidence and a measurable follow-up.

A recommendation should record the problem, affected scenarios and surfaces, evidence, plausible mechanism, confidence, intervention type, priority, owner and date for remeasurement.

The distinction between direct control, indirect influence and observed outcomes follows the GEO control surface. An audit can identify a likely intervention area without claiming full causal access to a platform’s internal process.

Technical interventions

Use technical recommendations when evidence indicates an access or discoverability problem, such as:

• blocked crawlers;

• non-indexable pages;

• incorrect canonicalisation;

• important information unavailable in HTML;

• broken internal links;

• obsolete pages still discoverable.

A technical fix may improve the conditions for retrieval. It does not guarantee future selection, citation or recommendation.

Content, claim and entity interventions

Use these when the information estate is incomplete or ambiguous:

• create a precise product definition;

• publish missing facts and limitations;

• update pricing or documentation;

• separate the company from its products and sub-brands;

• clarify category and geographical scope;

• align language versions;

• disambiguate similar entities.

The recommendation should identify which tested claims or scenarios justify the change.

Source ecosystem interventions

Use source actions when external descriptions are outdated, conflicting or absent:

• correct a partner or directory profile;

• request a factual amendment;

• update integration documentation;

• publish a sourceable report;

• develop relevant earned media;

• improve independent comparison coverage;

• resolve contradictions across prominent domains.

This is influence rather than control: a brand can provide evidence or request correction, but cannot dictate editorial or platform decisions.

Monitoring interventions

Some findings require observation rather than immediate remediation:

• a critical false claim that appears intermittently;

• changing competitor sets;

• cross-language drift;

• unstable recommendation scenarios;

• new sources entering the citation set;

• a platform-specific error after a product update.

The report should specify what will be monitored, how often and what threshold would trigger action.

What an AI visibility audit report should contain

A professional report should make conclusions traceable to evidence.

1. Executive diagnosis

Summarise the most material visibility gaps, representation errors, source risks, unstable scenarios and three to five priorities. Avoid presenting a composite score without the underlying dimensions.

2. Scope and methodology

Document the entities, claim map, scenarios, prompt variants, surfaces, dates, languages, locations, search states, number of runs and classification rules. State known limitations.

3. The 5P scorecard

Report Presence, Position, Provenance, Precision and Persistence separately. Numerical summaries are acceptable when definitions, denominators and samples are explicit.

4. Scenario and platform matrix

A useful structure is:

scenario × surface × brand role × sources × claim accuracy × stability

This reveals whether a problem is platform-wide, language-specific, limited to purchase-intent scenarios or associated with one source.

5. Claim audit

For each material claim, provide the reference version, generated variants, status, supporting or conflicting sources, affected surfaces, frequency and recommended response.

6. Source ecosystem analysis

Show owned, earned, institutional, partner, directory, review, forum and competitor sources. Identify dominant domains, source gaps, contradictions and outdated material.

7. Error register

Record the error type, evidence, scenario, surface, frequency, business significance, plausible mechanism, confidence level and proposed intervention.

8. Prioritised action plan

Prioritise actions by risk, scenario importance, frequency, degree of control, cost, implementation time and need for remeasurement.

Why a single score can mislead

A composite score can simplify executive communication, but it should not replace the diagnostic view.

Consider three brands:

• Brand A is mentioned frequently but carries an inaccurate product category.

• Brand B is mentioned less often but is usually the primary recommendation.

• Brand C is cited widely as a source but is rarely included as a provider.

A single score may rank them, but it cannot explain what each should do next. The five dimensions need to remain visible even when a summary indicator is used.

What this does not mean

A mention is not a recommendation

Presence and Position are different measurements.

A citation is not proof of source influence

A visible source may support one claim, provide only background context or be listed without measurable absorption into the answer.

No citation does not prove no retrieval

The public interface does not expose the complete process.

Accuracy is not completeness

An answer may contain no false statement while omitting an important limitation or differentiator.

Stability is not accuracy

Repeated error increases persistence risk; it does not validate the claim.

An audit does not prove causality

It identifies patterns, visible evidence and plausible intervention areas. It does not reconstruct the complete internal mechanism of an answer system.

AI visibility does not prove revenue impact

Business impact requires separate traffic, conversion, branded-search, sales and attribution evidence.

A practical reporting checklist

Scope and evidence

• Are the audited entities and competitors explicit?

• Is there a verified reference claim map?

• Are full prompts, answers, sources and conditions preserved?

• Are web-search and non-search results separated?

Classification

• Is presence separated from recommendation role?

• Are source types and claim support recorded?

• Are material answers decomposed into atomic claims?

• Are errors reviewed against evidence rather than inferred from tone?

• Is stability reported separately for different outcomes?

Reporting

• Are definitions and denominators provided for every metric?

• Can each major conclusion be traced to answer records?

• Are limitations and uncertain interpretations visible?

• Does the report avoid hiding the diagnosis inside one score?

• Does every recommendation specify evidence and remeasurement?

Audit the representation, not only the mention

The purpose of an AI visibility audit is not to produce the largest prompt collection or the cleanest dashboard.

It is to determine:

• where the brand is present;

• how it is positioned;

• which visible sources support the representation;

• whether the claims are correct;

• whether the result persists;

• which intervention is justified by the evidence.

That is what turns AI answer monitoring into an audit.

Brand Semantics applies this approach through AI Strategic Consulting, connecting technical access, source analysis, claim verification and representation monitoring.

Discuss an AI visibility audit with Brand Semantics.

Sources and methodological notes

• Zhang Kai, He Xinyue and Yao Jingang, From Citation Selection to Citation Absorption, arXiv preprint, April 2026. Used to separate citation selection, breadth and measurable source absorption.

• Pranav Narayanan Venkit et al., DeepTRACE, arXiv preprint, September 2025. Used for statement-level decomposition and citation-support analysis. Its empirical scope and model-assisted evaluation limit direct generalisation.

• Ronald Sielinski, Quantifying Uncertainty in AI Visibility, arXiv preprint, revised June 2026. Used for repeated measurement, citation variability and uncertainty.

• Julius Schulte, Malte Bleeker and Philipp Kaufmann, Don’t Measure Once, arXiv preprint, April 2026. Used for treating visibility as a distribution across runs, prompts and time.

Methodological note: The 5P model is a Brand Semantics organising framework. It integrates established and emerging concerns around visibility, positioning, source provenance, factual accuracy and repeated measurement; it is not official platform terminology or an established academic standard.