How to run an AI visibility audit that produces defensible evidence

Define the decision and prompt universe first

An audit is meaningful only inside a fixed market, audience and set of buyer decisions.

Build prompts from real discovery, comparison, suitability, location, pricing and risk questions. Include branded checks for accuracy, but do not let them dominate the score. Record why each prompt matters and which competitors a buyer could reasonably compare.

Executive Intelligence’s partner page proposes agreeing the market, competitors, buyer questions, AI tools and proof limits before collection, then retaining answers, screenshots and sources. That is a strong audit record; the provider’s conclusions still need to be evaluated against the underlying evidence.^[1]

• Market and language
• Buyer stage and decision
• Prompt family and exact wording
• Eligible competitors
• AI surfaces and collection dates
• Claim the audit is allowed to make

Repeat prompts and preserve answer-level evidence

One answer is an example, not a stable visibility estimate.

Run the same prompt multiple times where the product allows it, and use predefined paraphrases to test intent rather than one lucky wording. Record model or mode, date, account state and location when relevant. Do not silently replace failed, empty or contradictory runs.

Recent primary research argues that AI-search visibility should be treated as a distribution because answers vary across runs, prompts and time. A 2026 critical review similarly warns that discoverability, citation, prominence, factual use and business outcomes are different stages with heterogeneous evidence.^[4][5]

Measure distinct outcomes instead of one opaque score

A brand can be cited without recommendation, mentioned inaccurately or recommended without a visible citation.

For every answer, record whether the brand appears, how it is described, whether it is shortlisted, its order, qualifications or warnings, and which URLs are cited. Add factual-accuracy review by a person who knows the product or service.

Publish the numerator and denominator behind rates. For example: mentioned in 12 of 30 eligible answers, recommended in 5, correctly described in 10, and supported by a visible citation in 4. Keep prompt-level rows available so another reviewer can challenge the summary.

Check access and source evidence before rewriting pages

A missing mention can originate before writing quality enters the decision.

OpenAI says OAI-SearchBot is used to surface websites in ChatGPT search results and recommends allowing it in robots.txt and through relevant network controls. Google says eligibility for its AI search features uses the same search fundamentals: crawl access, index eligibility, useful textual content and structured data that matches the visible page.^[2][3]

Check robots rules, CDN or firewall blocks, canonical URLs, indexability, internal links and important content rendered in HTML. Then review whether the page states current product facts, location, availability, limitations and evidence clearly. A special AI schema or an llms.txt file is not a substitute for these fundamentals; Google says no special AI markup is required for its AI features.^[3]

Turn repeated gaps into measured experiments

Prioritise gaps that recur across commercially important prompts and have a plausible evidence-based cause.

Classify the proposed action: technical access, entity or business-fact correction, answer-page improvement, primary proof, or independent source coverage. Assign an owner and recheck date. Preserve the baseline prompt set while adding a separate exploratory set for new questions.

The original GEO paper reported visibility gains within its experimental benchmark, but those results do not guarantee organic discovery or durable commercial impact on live platforms. Treat content changes as hypotheses, repeat the same audit and connect visibility movement to qualified visits, leads or other business outcomes.^[6][5]