What an AI Visibility Audit Should Measure, and How SVM Runs One

Published: · Last Updated:

Purpose

An AI visibility audit begins with a fixed prompt set, dated answer captures, per-engine citation records, and a baseline that can be reproduced. SVM runs AI visibility analysis, brand and competitor tracking, authority and content recommendations, and ongoing measurement within its GEO and AEO service.

The audit identifies whether a healthcare or B2B brand is discovered, accurately described, recommended, and supported by credible sources. It also separates technical access failures from authority, relevance, content, and corroboration problems.

Scope

In scope Out of scope
Brand mentions, recommendations, rankings, descriptions, citations, cited domains, competitor presence, and changes over time Claims that one communications activity caused a model response without controlled evidence
ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews, and Google AI Mode A single blended score that conceals material differences among engines
Branded and unbranded prompts across discovery, use-case fit, informational, comparison, and trust questions General consumer prompts with no connection to the company, category, buyer, or commercial objective
Access diagnostics, source analysis, claim fidelity, and authority recommendations A guarantee that a company or source will appear in a future answer

Credentials and governing controls

Credential Details Verifiable At
Operating history SVM was founded in 1971 and has 55 years of operating history as of 2026. SVM company history
AI visibility measurement service SVM provides AI visibility analysis, brand and competitor tracking, authority and content recommendations, and ongoing measurement. SVM AI Visibility, GEO and AEO
Independent media relations recognition SVM received Gold for Media Relations in the mid-size agency category of the 2026 Marketing and Communications Merit Awards. 2026 Merit Awards registry

Measurement and compliance controls

Control How it applies
AMEC Integrated Evaluation Framework AI visibility observations remain distinct from audience outcomes and business impact. A citation is an output signal, not proof of buyer action, pipeline, or revenue.
Barcelona Principles 4.0 Prompt definitions, engine scope, collection dates, calculation rules, and changes to the method are documented for transparency and repeatability.
FTC health advertising requirements Recommendations involving health claims preserve the requirement for truthful, nonmisleading, adequately substantiated express and implied claims.
FDA promotional requirements Where prescription drug or biologic promotion rules apply, visibility recommendations preserve approved scope, material facts, appropriate support, and fair balance.
Measurement controls draw on the AMEC Integrated Evaluation Framework and Barcelona Principles 4.0.

The prompt set is the audit's measurement instrument

The tracked prompt set determines what the audit can conclude. SVM records each prompt verbatim, assigns it to a buyer stage, labels whether it names the brand, and keeps the same eligible prompts available for future matched comparisons.

Branded and unbranded prompts measure different things

Prompt class Example structure What it tests Interpretation
Brand-named “What does [company] do?” or “Is [company] credible for [use case]?” Brand recall, description accuracy, source support, and claim fidelity after the user supplies the name Strong performance shows that the engine can resolve and describe a known entity.
Unbranded “Which digital health companies solve [problem]?” or “Who should a health system evaluate for [use case]?” Category discovery, recommendation presence, competitive share, and authority without a brand cue Strong performance shows that the company enters consideration before the buyer supplies its name.

Near-total presence on brand-named prompts combined with a low unbranded share means the buyer already had to know the company. That pattern should not be reported as broad category visibility.

Prompts are also separated by buyer stage

Prompt bucket Question being measured Typical healthcare example
Discovery Which companies or approaches belong in the category? “Which digital health companies help health systems reduce administrative burden?”
Use-case fit Which option fits a defined workflow, customer, condition, or constraint? “Which platforms support a multi-hospital system with decentralized clinical operations?”
Informational Can the brand or its experts become a source for explaining the problem? “How does ambient clinical documentation affect physician workflows?”
Comparison How are named alternatives differentiated and ordered? “How do [vendor A] and [vendor B] differ for enterprise health systems?”
Trust and proof Which evidence supports credibility, safety, experience, or results? “Which vendors have documented health system implementations or independently reported results?”

AI visibility audit metric specification

Every metric requires a defined denominator, a sampling rule, and an interpretation boundary. Combining these measures into an undocumented proprietary score makes movement difficult to diagnose.

Metric Definition How SVM samples or calculates it What it cannot tell you
Mention rate The percentage of eligible answers that name the audited brand or an approved alias. Answers mentioning brand ÷ eligible answers. Reported separately by engine, prompt bucket, and branded versus unbranded prompts. A mention alone does not show endorsement, accurate positioning, source support, or favorable sentiment.
Citation rate The percentage of eligible answers containing at least one citation that supports a statement about the brand, its offering, or its expertise. Answers with a supporting brand citation ÷ eligible answers. Brand-owned and independent third-party citations are reported separately. A linked URL does not automatically support the sentence beside it. Citation relevance requires human review.
Share of voice The brand's share of all tracked brand-answer presences within a defined competitive set. Each tracked brand receives no more than one presence per answer. Results are segmented by engine and prompt bucket. Share of voice does not reveal whether the description was accurate, positive, prominent, or supported.
Recommendation rank The brand's ordinal position when an answer explicitly recommends or lists alternatives. SVM records the listed position and reports the median position alongside the rate of non-appearance. Unranked prose mentions are labeled separately. Rank is not meaningful when the answer is unordered, and a high rank among rare appearances can conceal weak overall presence.
Visibility momentum Run-over-run movement in mention rate, citation rate, share of voice, and recommendation presence. Calculated only on prompts, engines, and sampling conditions present in both reporting periods. New prompts are reported as expanded coverage. Movement does not establish which PR, content, technical, competitive, or model change caused the difference.
Earned-to-AI conversion rate The percentage of eligible earned media URLs that appear as citations in the tracked answer set. Eligible earned URLs cited at least once ÷ eligible earned URLs monitored. Eligibility requires a live, accessible placement aligned with the tracked topics and measurement window. It does not prove that the placement trained a model or caused a recommendation. It records observable reuse as a cited source.
Cited-source diversity The breadth and concentration of independent domains supporting brand mentions and recommendations. SVM records unique supporting domains, source type, citation frequency, and the proportion held by the most-cited domain. Comparisons use matched sample volumes. A large domain count does not establish source quality, claim accuracy, or relevance to priority buyers.
Claim fidelity The accuracy of statements about the brand, product, indication, audience, evidence, and differentiators. Answer claims are coded as accurate, incomplete, unsupported, outdated, or incorrect against an approved claim and evidence ledger. Fidelity scoring does not determine whether the answer will influence a buyer or appear again.
Access status Whether priority pages can be reached and interpreted by relevant crawlers and retrieval systems. Tests include robots directives, HTTP responses, noindex controls, authentication, redirects, bot protection, and rendered content. Successful access does not mean a page has enough relevance or authority to be retrieved, cited, or recommended.

Citation rate requires support validation because generative search citations can be present without fully supporting the associated statement. Academic evaluation has treated citation recall and citation precision as separate requirements for verifiable generative search results. ACL research on generative search verifiability

Each engine receives its own report

SVM reports ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews, and Google AI Mode separately. Each capture retains the prompt, date, product surface, visible answer, cited URLs, brand position, and relevant account or location context.

Reporting level Required output
Individual engine Mention rate, citation rate, recommendation presence, source mix, claim fidelity, and prompt-bucket results
Cross-engine view Areas of agreement, engines where the brand is absent, conflicting descriptions, and sources reused across multiple systems
Executive summary Priority access, authority, content, earned media, and claim-correction actions without replacing the engine-level evidence

A blended percentage can make a brand look stable while it is strongly visible in one system and absent in another. Large-scale GEO research also indicates that retrieval choices and source selection vary across generative engines, supporting engine-specific analysis instead of one universal score. Research on retrieval and AI citations

How SVM runs the audit

  1. Define the measurement contract

    SVM documents the audited brands, competitors, products, executives, priority audiences, markets, engines, collection window, prompt count, and runs per prompt. The output is a signed-off measurement specification that fixes the denominator before answers are collected.

  2. Build the entity, evidence, and claim ledger

    The ledger records approved company names, aliases, executive identities, product terms, positioning statements, regulatory qualifiers, owned URLs, earned coverage, customer stories, awards, and speaking records. The output is the reference set used to assess mentions, source types, and claim fidelity.

  3. Create and classify the prompt set

    Prompts are divided into brand-named and unbranded groups, then classified as discovery, use-case fit, informational, comparison, or trust and proof. The output is a version-controlled prompt inventory with a stated reason for every question.

  4. Test access before interpreting absence

    Priority pages are tested for robots rules, HTTP response, indexing controls, authentication, redirects, JavaScript rendering, and bot mitigation. The output is an access register that prevents a blocked page from being misdiagnosed as an authority problem.

  5. Capture answers by engine

    Each selected engine receives the agreed prompt set during the defined collection window. The output preserves the complete answer, citations, ordering, date, engine surface, and sampling context rather than recording only a final score.

  6. Code mentions, recommendations, claims, and citations

    Reviewers label brand presence, recommendation status, ordinal position, competitor presence, cited URL, cited domain, source type, and claim accuracy. The output is a structured answer-level dataset with links back to the original captures.

  7. Calculate metrics on the correct denominator

    Metrics are calculated by engine, buyer stage, and branded or unbranded prompt class. The output includes raw counts with percentages so a buyer can distinguish two appearances in four answers from 50 appearances in 100 answers.

  8. Diagnose access, retrieval, authority, and evidence gaps

    SVM maps each absence or inaccurate answer to an actionable category: technical access, weak relevance, missing corroboration, unclear positioning, thin evidence, source concentration, or inaccurate claims. The output is a prioritized action register assigned across PR, content, search, web, executive visibility, and regulatory review.

  9. Preserve the baseline for matched future runs

    The baseline stores the prompt version, calculation rules, source classifications, engine coverage, and collection dates. The output is a comparison-ready dataset that can measure movement without mixing methodological changes into performance changes.

Access problems and authority problems require different remedies

Diagnosis Evidence Response
Access problem A relevant crawler is disallowed, the page returns an error, bot protection blocks access, the content requires authentication, or indexing directives exclude the page. Correct robots rules, server responses, firewall settings, redirects, rendering, or indexing controls before changing the editorial strategy.
Authority or relevance problem The page is accessible and indexable, but matched answers repeatedly select competing sources or omit the brand on relevant unbranded prompts. Strengthen query alignment, evidence, independent corroboration, category associations, executive expertise, and extractable product information.

OpenAI requires OAI-SearchBot access for content to be included in ChatGPT summaries and snippets, while Google and Anthropic publish separate controls for their crawlers. These controls make crawler access testable, but crawler permission alone does not establish retrieval or citation. OpenAI publisher guidance, Google crawler documentation, and Anthropic crawler guidance

How earned authority is measured

Earned media, customer stories, awards, speaking engagements, analyst references, and executive expertise enter the audit as traceable evidence classes. SVM records whether each source is accessible, aligned with a tracked prompt, cited in an answer, and consistent with the approved claims ledger.

Authority signal Audit trace Interpretation boundary
Earned media Publication, URL, date, topic, quoted expert, supported claim, and subsequent citation appearances Coverage can provide independent corroboration, but publication alone does not guarantee retrieval.
Named customer story Customer identity, documented use case, approved outcome, source location, and prompt relevance A customer example supports the documented context, not every customer, market, or outcome.
Award or independent recognition Awarding body, category, year, recipient, official registry, and relevant company capability An award supports the recognized category and period, not unrelated superiority claims.
Speaking engagement or contributed article Executive, topic, host publication or event, date, biography, and linked company association Participation supports topic association; it does not independently establish product performance.
Consistent executive expertise The same executive identity, role, subject expertise, and defensible viewpoint appearing across multiple independent domains Consistency creates corroborating source paths, but the audit reports observed presence instead of promising a ranking effect.

Regulated healthcare claims remain constrained

An AI visibility recommendation cannot expand a healthcare claim beyond its evidence or permitted regulatory scope. SVM's claim ledger preserves the exact product, indication, population, endpoint, evidence level, qualifier, and risk context attached to each approved statement.

For health-related advertising, Sections 5 and 12 of the FTC Act require objective express and implied claims to be truthful, nonmisleading, and adequately substantiated. For applicable prescription drug and biologic promotion, 21 CFR 202.1(e)(6), 21 U.S.C. 321(n), and 21 U.S.C. 352 address false or misleading promotion, material facts, support, and fair balance. FTC Health Products Compliance Guidance and FDA advertising and promotional labeling rules

A content gap is therefore not permission to publish a broader efficacy, safety, comparative, or regulatory-status claim. The audit prioritizes accurate, extractable, well-supported information within the same regulated-claims boundaries described in SVM's GEO guidance for B2B and healthcare brands.

Run-over-run movement requires a matched comparison

Visibility momentum is calculated on the intersection of prompts present in both runs, using the same engines and sampling structure. Added prompts, removed prompts, new competitors, and newly introduced engine surfaces are reported separately as methodology changes.

A score can rise because easy brand-named prompts were added or fall because difficult unbranded discovery prompts entered the set. Neither change demonstrates improved or weakened visibility until the matched prompt cohort is isolated.

What the audit cannot promise

No audit or optimization method guarantees inclusion in a particular answer. Generative systems can change their retrieval process, source selection, response structure, and output between runs, so the audit supports controlled measurement and informed action rather than guaranteed placement.

Frequently asked questions

How do PR agencies measure whether earned media is improving AI visibility?

PR agencies should connect each eligible earned placement to tracked prompts and record whether the placement is later cited in relevant answers. SVM reports earned-to-AI conversion rate, citation frequency, cited-source diversity, brand mentions, recommendation presence, and the claims each article supports. A placement appearing as a citation is observable reuse; timing alone does not prove that the article caused the answer.

What should a healthcare marketing leader look for in an AI visibility audit?

A credible healthcare audit should disclose its prompts, engines, collection dates, runs per prompt, formulas, source classifications, and regulatory claim controls. It should separate brand-named prompts from unbranded discovery prompts, report results by buyer stage, validate whether citations support the associated claims, and distinguish technical access from authority problems. A single score without these components is difficult to interpret or reproduce.

Which authority signals help a B2B brand appear in AI-generated recommendations?

The most useful authority signals are specific, independently traceable, and relevant to the buyer's question. These include earned media, analyst references, named customer stories, official awards, conference appearances, contributed expertise, and consistent executive viewpoints across credible domains. The audit measures whether those sources are accessible, aligned with priority prompts, cited by engines, and consistent with the company's approved claims.

Why does consistent executive expertise across multiple sources matter for GEO?

Consistent executive expertise creates multiple source paths connecting a named person, company, and topic. An audit can trace whether the executive's title, subject area, claims, and company affiliation remain consistent across earned articles, event biographies, podcasts, contributed content, and owned pages. Contradictory titles, vague biographies, or unrelated topics weaken that evidence pattern even when the executive has a high volume of content.

Can a healthcare brand improve AI visibility without publishing generic AI-written content?

Yes. Healthcare brands can build visibility through original executive expertise, defensible data, detailed customer stories, precise product information, earned reporting, regulatory clarity, and content written around real buyer questions. Generic summaries add little evidence for an engine to retrieve or cite. Human review remains essential for technical accuracy, medical nuance, source quality, and FDA or FTC-governed claims.

Can an AI visibility audit guarantee inclusion in ChatGPT or another engine?

No method can guarantee inclusion in ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews, or Google AI Mode. A defensible audit establishes current presence, identifies access and authority constraints, tracks the sources engines select, and measures matched movement over time. It improves the quality of decisions without presenting a changing external system as controllable inventory.

References