What should an AI visibility audit measure?

Brand visibility in AI search is often reduced to a single question: does ChatGPT, Google AI Overviews or Perplexity mention the company?

That is useful as a preliminary observation, but it is too shallow a standard for an audit.

A brand may appear frequently but be assigned to the wrong category. It may be cited as a source without being recommended as a provider. It may be represented accurately in one formulation of a question and disappear after a minor change in wording. A system may also repeat outdated information consistently, confuse two similar entities or attribute a feature to a product that has never offered it.

A professional AI visibility audit should therefore examine five separate questions:

• Is the brand present?

• What role is it given?

• Which visible sources support its representation?

• Are the claims accurate?

• Does the result persist across runs, prompts, platforms and time?

A mention is an observation. A reliable audit explains the representation behind it.

This article defines those five dimensions. The next article in the series explains how to run an AI visibility audit across public AI search surfaces.

An AI visibility audit is not a prompt check

The simplest form of AI visibility testing usually follows four steps:

1. prepare a list of questions;

2. enter them into several answer systems;

3. count how often the brand appears;

4. compare the total with competitors.

This can reveal obvious absences or errors. It does not establish why the questions were selected, which commercial intentions they represent, whether web search was active, whether the runs were repeated or whether the brand appeared as a recommendation, a source or an incidental example.

It also risks treating one probabilistic output as a stable result.

In Quantifying Uncertainty in AI Visibility, Ronald Sielinski tested repeated samples from Perplexity Search, OpenAI SearchGPT and Google Gemini. Citation distributions varied substantially across daily and high-frequency measurements, and some apparent differences between domains fell within the statistical noise of the measurement process. The paper is a 2026 preprint, so its numerical findings require further replication, but it directly challenges single-run visibility reporting.

Don’t Measure Once: Measuring Visibility in AI Search reaches the same broader conclusion: answers can vary across runs, prompts and time, so visibility should be treated as a distribution of possible outcomes rather than a fixed rank.

This does not imply one universal number of repetitions. Sample size should depend on the platform, scenario and observed variability. It does support a minimum rule:

An audit in which each prompt is run only once is a collection of illustrative observations, not a reliable measurement of visibility stability.

The 5P AI representation audit model

Brand Semantics uses the 5P AI representation audit model:

1. Presence

2. Position

3. Provenance

4. Precision

5. Persistence

The underlying concerns are not new. The original GEO research, later published at KDD 2024, formalised visibility and position-adjusted measures for generative answers. Subsequent research has examined repeated measurement, citation support and claim-level reliability.

Comparable elements also appear in practitioner methodologies. David Cosgrove’s Five Layers of AI Brand Knowledge covers entity recognition, factual accuracy, positioning, knowledge gaps and source attribution. Digital Applied’s AI Search Visibility Score includes Position and Persistence, although it defines them more narrowly and combines them into a composite score. Yotpo also uses the term brand persistence for repeated visibility across sessions.

The value of the 5P model lies in integrating these concerns as five operationally separate dimensions of a representation-focused audit. It does not claim that mentions, positioning, provenance, accuracy or stability were newly discovered.

Presence: does the brand appear?

Presence is the most basic layer. It establishes whether the brand, product, domain or related entity appears in the answer.

Useful measures include:

• mention rate across relevant scenarios;

• platform coverage;

• product presence;

• co-occurrence of the brand and its category;

• presence in branded and non-branded questions.

A useful audit should distinguish at least four forms of presence:

• prompted presence — the brand appears because the question names it;

• spontaneous presence — the brand appears in a category, problem or recommendation scenario without being named;

• source-only presence — the domain is cited, but the brand is absent from the answer text;

• product-only presence — a product is mentioned without a clear association with the parent brand.

This matters because a brand that is recognised when explicitly requested is not necessarily discoverable in non-branded decision scenarios.

Mention rate is therefore not market share, recommendation probability or commercial value. It is one observation about inclusion.

Position: what role is the brand given?

Position is broader than the physical location of a brand name in the answer. It covers:

• the role assigned to the brand;

• recommendation status;

• category assignment;

• suitability for particular use cases;

• comparative relationship with competitors;

• prominence within the answer.

A practical taxonomy can include:

• primary recommendation;

• secondary recommendation;

• shortlist inclusion;

• category example;

• information source;

• background mention;

• caution or negative comparison;

• irrelevant inclusion;

• entity confusion.

A brand mentioned in 70 per cent of answers but almost always only as a source has a different visibility problem from a brand present in 30 per cent of answers but regularly presented as the strongest recommendation.

This is why Position should not be reduced to how early a mention appears. Textual order may be useful, but it does not capture category, role or recommendation status.

Provenance: which visible sources shape the answer?

Provenance concerns the observable source environment around an answer.

An audit can establish:

• which domains are cited;

• which sources belong to the brand;

• which come from media, directories, forums, partners or competitors;

• whether sources are current;

• whether a cited page concerns the correct entity;

• whether it supports the specific claim attributed to it;

• whether different platforms rely on different source ecosystems.

The precise term is visible provenance. An interface displaying several links does not reveal every document considered or the complete internal retrieval process. The absence of a citation also does not prove that no external information contributed to the answer.

Provenance is therefore not a synonym for citation count. It should include source type, independence, currency, entity match, contradictions and claim-level support.

The preprint From Citation Selection to Citation Absorption separates the selection of a cited page from the degree to which that page contributes language, evidence, structure or factual support to the final answer. Across 602 prompts and more than 21,000 search-layer citations, the authors found that citation breadth and measured depth of influence could diverge. The study does not justify a permanent ranking of platforms, but it supports measuring source contribution separately from citation volume.

Precision: are the claims accurate?

Precision concerns the factual and semantic accuracy of claims about the audited entity. It should not be confused with citation precision, which evaluates citations rather than the truth of the brand representation.

An answer should be decomposed into atomic claims.

The sentence:

“Company X is a British SaaS platform offering predictive analytics in its entry-level plan”

contains at least four claims:

1. Company X is a SaaS platform.

2. Company X is British.

3. It offers predictive analytics.

4. The feature is available in the entry-level plan.

Each may be true, false, outdated, partially true, unverifiable, missing an important limitation or attributed to the wrong entity.

The DeepTRACE framework applies statement-level analysis and builds matrices linking claims, citations and factual support. Its authors found that generative search and deep-research answers can contain material proportions of statements unsupported by their listed sources. The paper is a preprint and part of its evaluation used a model-based judge validated against human ratings, but the claim-level method is directly relevant to brand auditing.

Useful Precision measures include:

• claim accuracy;

• unsupported claim rate;

• hallucination rate;

• outdated claim rate;

• entity confusion rate;

• missing-limitation rate;

• unsupported recommendation rate.

A hallucination rate cannot be calculated credibly without a defined reference truth. The audit first needs an approved claim map, current product documentation, pricing, organisational data and other verified sources.

Persistence: does the representation hold?

Persistence is an umbrella dimension for several forms of stability:

• repeated-run stability;

• prompt-variant stability;

• cross-platform consistency;

• cross-language consistency;

• citation stability;

• claim stability;

• temporal persistence.

Earlier industry methods also use the term, sometimes more narrowly. Digital Applied defines Persistence through continued weekly citation presence, while Yotpo applies brand persistence to repeated sessions. The 5P model uses it across presence, role, sources, claims and competitors rather than reducing it to the lifespan of one citation.

This distinction matters because a system may:

• mention the brand consistently but change its role;

• recommend the brand consistently while changing its evidence;

• repeat the same factual error in every run;

• represent the brand correctly in English but incorrectly in Polish;

• remain stable within one platform while diverging sharply elsewhere.

Stability is not accuracy. An inaccurate representation can be highly persistent.

Why one AI visibility score is not enough

A composite score may be useful for executive reporting, but it is a poor substitute for diagnosis.

High Presence can coexist with low Precision. High Persistence may mean that a false claim is being repeated consistently. Strong citation share can coexist with weak recommendation visibility. A brand may dominate prompted questions while remaining absent from non-branded discovery scenarios.

Combining these outcomes into one number hides the mechanism that should guide action.

A 5P scorecard can still use numerical summaries, but the five dimensions should remain visible and separately interpretable. The audit should answer not only whether visibility is high or low, but what kind of visibility exists, whether it is accurate and what evidence supports it.

What this does not mean

A mention is not a recommendation

A brand may appear as a source, example or peripheral reference. Mention rate does not automatically indicate recommendation share.

A citation is not proof of absorption

A visible link does not establish how deeply the page influenced the answer. Citation selection and citation absorption are different observations.

No citation does not prove no retrieval

The absence of a visible link does not reveal the complete generation process. An audit evaluates observable outputs and visible sources.

Stability is not accuracy

A system may repeat an outdated or false claim consistently.

One score does not represent the entire audit

Aggregation may conceal a material error, unstable recommendation or source dependency.

AI visibility does not prove business impact

Presence in answers may support discovery or purchasing decisions, but revenue impact requires separate traffic, conversion, attribution and qualitative evidence.

Measure the representation, not only the mention

An AI visibility audit should not answer only whether a brand appears in ChatGPT or Google AI Overviews.

It should establish:

• where the brand appears;

• what role it receives;

• which competitors surround it;

• which sources are visible;

• whether those sources support the claims;

• whether the representation is accurate;

• whether the result persists across relevant conditions.

The 5P model organises those questions without collapsing materially different outcomes into one visibility score.

The next step is methodological: define the entity, build intent scenarios, select the relevant product surfaces and preserve enough evidence for the findings to be reviewed. That process is covered in How to run an AI visibility audit across AI search platforms.

Brand Semantics applies this distinction through AI Strategic Consulting, connecting technical visibility, source analysis, claim accuracy and representation monitoring.

Sources and methodological notes

• Pranjal Aggarwal et al., GEO: Generative Engine Optimization, peer-reviewed at KDD 2024. Used for early formalisation of visibility and position-adjusted measures in generative answers.

• 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 visibility as a distribution across runs, prompts and time.

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

• Pranav Narayanan Venkit et al., DeepTRACE, arXiv preprint, September 2025. Used for statement-level auditing and citation-support analysis.

• David Cosgrove, The Five Layers of AI Brand Knowledge; Digital Applied, AI Search Visibility Score; and Yotpo, How to do an AI search audit for brand mentions. These practitioner materials document earlier public use of overlapping concepts; they are not independent validation of proprietary metrics.

Methodological note: The 5P AI representation audit model is a Brand Semantics organising framework. Its contribution is the integration and operational separation of five established audit concerns. The terminology is not official platform terminology or an established academic standard.