How to run an AI visibility audit for PR clients in 2026

A note from Maya

Maya Bennett is PRmatic's editorial persona, used to give this publication a consistent practitioner voice. She is not a real PRmatic employee or a claim of personal employment experience.

I would be wary of any AI visibility audit that starts with a score and only later explains what was measured. A tidy number is useful for a dashboard, but it can hide the choices that matter: which prompts were tested, which market, which version of an answer, whether a citation counted more than a passing mention, and whether the brand description was actually correct.

For PR teams, the better audit is one you can defend in a client meeting. It should show the evidence behind the conclusion and make the next communications decision clearer.

What an AI visibility audit should answer

An AI visibility audit is a baseline diagnostic, not continuous monitoring. Its purpose is to answer a fixed set of questions about how a brand is represented across AI and generative-search surfaces at a specific point in time.

That makes it different from GEO strategy, which is about improving visibility over time, and different from AI media monitoring, which is about watching ongoing changes in coverage, mentions and narratives. The audit tells you where you stand and why; the strategy decides what to do; monitoring tells you what changes afterwards.

• Does the brand appear for important category, problem, comparison and trust questions?
• Is it merely mentioned, or actually recommended or used as an example?
• How is the brand described, and are the facts current and correct?
• Which owned and third-party sources are cited or repeatedly used around the answer?
• Which competitors, alternatives or category authorities appear instead?
• Which gaps can PR, content, search or product teams realistically address?

Step 1: define the business question before the prompt list

Start with the client decision the audit needs to support. A brand-awareness audit, a category-entry audit and a reputation-risk audit need different prompts. If the team starts by generating 100 questions without a decision framework, the result becomes a large spreadsheet with no priority logic.

For an agency client, I would usually build prompt clusters around how a real stakeholder moves from discovery to evaluation. Keep the language natural. Do not write prompts purely to force the client's brand into the answer.

• Category discovery: who are the credible providers, experts, products or organizations in this space?
• Problem discovery: what solutions or approaches are recommended for the problem the client addresses?
• Comparison: how do leading approaches, brands or options differ?
• Trust and proof: which organizations, experts or sources are considered authoritative?
• Reputation and risk: what concerns, controversies or limitations are associated with the category or brand?
• Branded accuracy: what does the system say when asked directly about the client?

Step 2: freeze the audit method so the baseline is reproducible

AI answers can vary with phrasing, timing, location, product mode and model updates. That does not make measurement useless; it means the method has to be explicit. Record the exact prompt, platform or search feature, market, language, date and whether web search was used.

Where the same prompt is repeated, keep the number of runs consistent. If you change the prompt set every month, a rising visibility score may simply reflect an easier test. A stable core prompt set can be supplemented with a smaller experimental set for emerging topics.

• Keep a fixed core set of high-value prompts for period-to-period comparison.
• Record the full answer, not only whether the client appeared.
• Capture visible citations or source links when the surface provides them.
• Separate web-connected answers from model-memory-only answers where that distinction exists.
• Version the methodology whenever you materially change prompts, platforms or scoring rules.

Step 3: score the evidence in layers, not with one mystery number

A composite score can help executives scan the result, but the underlying measures should remain visible. I would keep at least six separate signals so a client can see whether the problem is absence, weak recommendation, inaccurate description or poor source support.

The difference matters. A brand mentioned frequently but described incorrectly has a different communications problem from a brand that is accurate when named but absent from non-branded discovery prompts.

• Mention rate: percentage of tested answers that name the brand.
• Recommendation rate: percentage that actively recommend or shortlist the brand where a recommendation is relevant.
• Citation rate: percentage of tested answers that visibly cite an owned or earned source connected to the brand.
• Factual accuracy: share of material brand facts that are current, supportable and correctly attributed.
• Narrative consistency: whether positioning, category, proof points and key facts are represented consistently.
• Source diversity: whether the information environment depends on one page or is supported by several credible sources.

Step 4: use first-party search evidence before inventing proxy metrics

The measurement landscape improved materially in 2026. Google Search Console now provides a dedicated view for visibility in generative AI features in Search, including impressions, pages, countries, devices and time trends. Bing Webmaster Tools' AI Performance report shows total citations, cited pages and grounding queries across supported AI experiences.

These reports do not answer every PR question, but they are stronger than guessing from third-party scores when the underlying platform can provide direct evidence. Build the audit around first-party data where it exists, then add controlled prompt testing for the questions those platform reports cannot answer.

• Google generative-AI impressions show visibility within Google's AI search features, not necessarily clicks or recommendations.
• Bing AI citation counts show that a page was used as a source, not that the brand was ranked first or endorsed.
• Grounding queries can reveal the phrases associated with cited content and help connect source evidence to editorial opportunities.
• Platform data and prompt testing should be reported as different evidence types rather than blended into an unexplained score.

Step 5: map which sources appear around the brand

For PR, the source map is often the most actionable part of the audit. Capture the owned pages, journalism, trade publications, public records, research, reviews and other third-party sources that recur when the brand or category appears.

A 2025 large-scale research study comparing AI search with traditional web search found a strong preference for third-party earned sources in the AI-search systems it tested. I would not turn that finding into a universal ranking rule, because platforms and retrieval systems differ. I would use it as a reason to examine whether credible independent coverage is part of the information environment around the client.

• Which sources are repeatedly cited for the category, even when the client is absent?
• Which outlets or documents appear when competitors are recommended?
• Are important facts supported only by the client's own site, or corroborated independently?
• Do newer, more accurate sources exist but fail to appear?
• Are old articles or outdated descriptions still shaping the answer?

Step 6: separate AI visibility from website traffic

A visibility audit should not treat referral traffic as a substitute for answer visibility. Many AI answers satisfy the user without a click. Conversely, a referral from an AI assistant may be one highly qualified prospect even if the brand has weak visibility across the wider prompt set.

For client reporting, preserve the raw traffic totals and create a qualified prospect view separately. Exclude known internal or test traffic from the prospect KPI, and keep AI-assistant referrals in their own channel so the agency can see whether that traffic becomes commercially meaningful without pretending it represents the whole AI visibility picture.

• Raw traffic: the complete analytics record kept for auditability.
• Qualified prospect traffic: external human traffic after removing known internal or test activity.
• AI referral traffic: visits referred by AI assistants, reported separately rather than deleted.
• AI-answer visibility: mentions, recommendations, citations and generative-search impressions, measured independently of traffic.

Step 7: convert findings into an action hierarchy

The output of the audit should be a prioritized worklist, not a pile of screenshots. I would rank actions by business importance, evidence strength, reversibility and who owns the fix.

Correct factual errors before chasing visibility. Then address important non-branded prompt gaps, missing source evidence, weak or inconsistent positioning, outdated owned content and technical discoverability problems. PR should own the communications and earned-source parts of that plan while coordinating with search, content and web teams where the fix lives elsewhere.

• Priority 1: material factual errors, outdated claims or reputation risks.
• Priority 2: absence from high-intent prompts tied to the client's commercial or stakeholder goals.
• Priority 3: weak independent evidence for claims the brand needs the market to trust.
• Priority 4: inconsistent descriptions across owned pages, executive profiles and authoritative third-party sources.
• Priority 5: content structure, freshness, indexing or technical issues that reduce discoverability.
• Priority 6: lower-value exploratory prompts and experimental opportunities.

Common AI visibility audit mistakes

The biggest audit errors usually come from methodology, not tooling. A visually impressive dashboard can still produce a weak baseline if the prompts were cherry-picked or the scoring model cannot be explained.

• Testing only branded prompts and then claiming strong category visibility.
• Running each prompt once and treating one answer as a stable market fact.
• Changing prompts between periods without resetting the baseline.
• Treating a visible citation as proof of endorsement, ranking or business impact.
• Combining mentions, citations, sentiment and recommendation into one score without showing the components.
• Assuming schema, crawler files or one technical change can manufacture authority that is not supported by useful content and credible sources.
• Claiming that one PR placement caused an AI answer without a controlled before-and-after method.
• Ignoring factual accuracy because the headline visibility score improved.

How to measure the audit over time

Repeat the core audit on a regular cadence only if the method is stable enough for comparison. Monthly can work for active campaigns or fast-moving categories; quarterly may be more sensible for slower B2B or corporate-reputation programs.

I would report movement in the component metrics and explain what changed in the evidence environment. The useful narrative is not 'visibility rose 12 points.' It is 'the brand began appearing in three high-intent comparison prompts, two earned sources started recurring in cited answers, and one outdated category description disappeared.'

• Core prompt mention and recommendation rates over time.
• Number and diversity of cited owned and earned sources.
• Generative-search impressions and cited-page trends from first-party platform reports.
• Factual-error count and time to correction.
• Narrative consistency across the most important brand attributes.
• Qualified prospect and AI-referral traffic as downstream signals, reported separately.
• Campaign actions completed against the prioritized audit findings.

The PRmatic approach

PRmatic treats AI visibility evidence as part of the wider campaign record rather than as a separate vanity dashboard. The agency still decides which questions matter, which sources are credible and which findings are actionable.

Pitch Studio can use approved client facts and source material when shaping angles. Media Discovery can preserve the evidence behind relevant journalists and outlets. Results Monitoring can bring coverage, replies, search and visibility evidence back into the campaign record so the team can review outcomes alongside the work that produced them.

Through Compound Knowledge and the Visibility Intelligence layer, those observations can inform later angle, source and targeting decisions while client contexts remain distinct. PRmatic is not presented as a magic universal AI-ranking score; the goal is to connect evidence, workflow and learning so the agency can make a better next decision.

Turn an AI visibility audit into a campaign brief

If an audit reveals that a client is missing from the questions that matter, the next step should not be 'do more GEO.' Translate the gap into a communications problem: which claim lacks credible proof, which narrative is inconsistent, which audience question is unanswered, or which authoritative source environment is missing?

PRmatic is preparing early access for agencies that want to carry that evidence from research into campaign planning, media discovery, outreach and results without rebuilding the context at every stage.

Sources & further reading

• Google Search Central — Introducing Search Generative AI performance reports in Search Console — June 2026 announcement of dedicated generative-AI visibility reporting in Search Console, rolled out worldwide by August 31, 2026.
• Google Search Central — A new resource for optimizing for generative AI in Google Search — Official guidance emphasizing useful original content and established SEO foundations rather than special GEO shortcuts.
• Microsoft Bing Webmaster Blog — Introducing AI Performance in Bing Webmaster Tools Public Preview — First-party reporting on AI citations, cited pages and grounding queries across supported Microsoft AI experiences.
• OpenAI Help Center — Searching the web with ChatGPT — Explains how ChatGPT search presents citations and source links and cautions that sources can be incomplete, outdated or incorrect.
• Chen et al. — Generative Engine Optimization: How to Dominate AI Search — Large-scale 2025 comparative study of AI search and traditional web search, including source-mix findings and earned-media patterns.

Related PR guides

• AI for PR agencies: practical use cases and best practices
• AI media monitoring for PR agencies: what to track in 2026
• PR workflow automation: what to automate and what to keep human