Blog · AI visibility audit
Your First AI Visibility Audit: What to Measure and How
Dharini Shah · September 8, 2026
An AI visibility audit measures how AI assistants such as ChatGPT, Gemini, Claude, Perplexity, Copilot and Google AI Mode describe, cite and recommend your brand for the questions your buyers ask. A useful first audit measures five things: visibility rate, citation rate, recommendation share, accuracy of how you are described and the sources the models rely on. It also checks whether AI crawlers can reach your site at all.
The audit becomes your baseline. Every later decision, from which pages to rewrite to which publications to pitch, depends on knowing where you stand and why. This guide walks through the audit step by step and explains what each measurement tells you.
What an AI visibility audit is not
It is not a screenshot of one ChatGPT answer. AI answers vary between runs. SparkToro and Gumshoe analyzed 2,961 prompt runs and found less than a 1 in 100 chance that two responses from ChatGPT or Google's AI would list the same brands. An audit must sample repeatedly and report percentages.
It is also not a traditional SEO audit, although it overlaps with one. Rankings and technical health matter because many AI systems retrieve from the web, but an AI audit adds prompt-level testing, off-site source analysis and brand description accuracy.
Step 1: Define the scope
Decide three things before you run anything.
Models. Cover the assistants your buyers use. For most B2B brands, that means at least ChatGPT, Google AI Overviews or AI Mode, Perplexity and Gemini. Add Claude and Copilot if your audience is technical or Microsoft-centric.
Competitors. Include four to eight brands: direct competitors, plus any the models themselves suggest when asked for alternatives to you.
Markets. If you sell in several countries or languages, audit the most important one first. Answers can differ by location and language.
Step 2: Build the prompt set
Create 20 to 50 prompts grouped into clusters:
| Cluster | Example prompt for a hypothetical invoicing tool |
|---|---|
| Brand | "What is [Brand] and who is it for?" |
| Category discovery | "What are the best invoicing tools for freelancers in the UK?" |
| Comparison | "[Brand] vs [Competitor] for a small agency" |
| Alternatives | "Cheaper alternatives to [Competitor]" |
| Use case | "Which invoicing software handles multi-currency and VAT?" |
| Pricing | "How much does invoicing software cost for a five-person team?" |
| Trust | "Is [Brand] secure and GDPR compliant?" |
Use real buyer language from sales calls, support tickets and community threads. Include the constraints buyers mention, because constraints change which brands get recommended.
Step 3: Run the prompts repeatedly
Run each prompt several times per model, ideally in fresh sessions without personalization or memory influencing the result. Aggregate runs within each cluster to get useful sample sizes. As a rough guide from standard sampling math, a 40% visibility rate measured across 100 answers carries a 95% margin of error of about 10 percentage points. Fewer answers mean wider uncertainty.
Record for every answer: which brands appear, which is recommended, the reasons given, every cited URL and any factual claims about your brand.
Step 4: Measure the five core metrics
1. Visibility rate
The percentage of answers that mention your brand. This shows whether AI systems associate you with the topic.
2. Citation rate
The percentage of answers that cite at least one of your URLs. This shows whether your content is used as evidence.
3. Recommendation share
The percentage of relevant answers in which your brand is actively recommended for the buyer's need, not just listed. Classify each mention as recommended, listed or qualified with caveats.
4. Accuracy
Compare every factual claim about your brand with reality: category, audience, features, pricing, integrations, locations and leadership. Log each error and where it seems to come from.
5. Source mix
List the domains cited in answers for each cluster and model. This shows which sources shape your category. Public research suggests sources differ by platform: Profound's analysis of 680 million citations found Wikipedia led for ChatGPT, while Reddit led for Perplexity and Google AI Overviews. Your category may differ, which is why you check.
Step 5: Audit technical access
Check whether AI systems can retrieve your site.
- robots.txt: confirm you are not blocking AI search crawlers such as OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot and Googlebot. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
- CDN and firewall rules: bot protection can silently challenge legitimate crawlers.
- Rendering: confirm important content appears in the HTML response, not only after client-side JavaScript runs.
- Server logs: look for AI crawler requests on priority pages and for errors returned to them.
- Indexing: Google's AI features rely on Google's index, so check Search Console coverage for priority pages.
Step 6: Audit off-site brand information
List the third-party pages that describe you: review platforms, directories, marketplaces, partner pages, Wikipedia or Wikidata where applicable, and comparison articles. Note outdated or conflicting facts. These pages often explain inaccuracies found in Step 4.
Step 7: Compare with competitors
For each cluster and model, compare your visibility rate and recommendation share with competitors. Note which sources support competitor recommendations. A competitor cited through three recent comparison articles has a different advantage than one recommended because of strong review volume.
Step 8: Turn findings into priorities
Classify each finding by where it sits in the chain from access to recommendation:
| Finding type | Example | Typical fix |
|---|---|---|
| Access | CDN blocks PerplexityBot | Adjust bot rules |
| Accuracy | Assistants quote last year's pricing | Update pricing page and third-party profiles |
| Content gap | No page answers a high-value use-case prompt | Publish answer-first page |
| Source gap | Competitor dominates cited comparison articles | Outreach, reviews, original data |
| Positioning | You are listed but not recommended | Clarify ideal customer and add proof |
Fix access and accuracy issues first, because they limit everything else. Then prioritize content and source gaps by business value.
What to include in the audit report
- Scope: models, competitors, prompts and dates.
- Methodology: runs per prompt and how answers were classified.
- Baseline metrics by model and cluster.
- Accuracy errors and their likely sources.
- Source mix for key clusters.
- Technical access findings.
- A prioritized action list with owners.
Keep the prompt set and method stable so you can re-run the audit and compare results.
Common mistakes
Testing only branded prompts. Brand questions almost always mention you. Category and comparison prompts reveal real competitive position.
Ignoring personalization. Logged-in sessions with memory can skew results. Use clean sessions for audits.
Reporting AI rankings. Order changes between runs. Report percentages.
Skipping the source analysis. Without it, you know that you are losing but not why.
A hypothetical example
A hypothetical cybersecurity training company runs its first audit across four models and 40 prompts. Visibility for brand prompts is near 100%, but only 9% for category prompts. Two assistants describe it as "for enterprises only," although it launched a small-business plan a year ago. The source analysis shows that description comes from an old analyst summary and an outdated review profile. Its robots.txt also blocks Claude-SearchBot through a broad rule written to stop scrapers. The priority list starts with the crawler rule and the outdated profile, before any new content.
How Bob Builds AI helps
Bob Builds AI's Visibility Monitoring automates repeated sampling across AI models and reports visibility rate, citation rate, competitor positioning, sentiment and the sources behind each answer. Agent Analytics helps teams see how AI crawlers access their sites, and the AI search readiness resources cover technical foundations.
FAQ
What is an AI visibility audit?
An AI visibility audit measures how AI assistants describe, cite and recommend your brand for the questions buyers ask. It covers visibility rate, citation rate, recommendation share, accuracy of brand descriptions and the sources models rely on, along with technical checks of AI crawler access.
How many prompts should an AI visibility audit include?
Most first audits use 20 to 50 prompts grouped into clusters such as brand, discovery, comparison, alternatives, use case, pricing and trust. Each prompt should be run multiple times per model, because AI answers vary between runs.
Which AI models should I audit?
Audit the assistants your buyers use. For most businesses, that includes ChatGPT, Google AI Overviews or AI Mode, Perplexity and Gemini. Add Claude and Microsoft Copilot where your audience is technical or uses Microsoft tools heavily.
How do I check if AI crawlers can access my site?
Review robots.txt for AI search crawlers such as OAI-SearchBot, Claude-SearchBot and PerplexityBot, check CDN and firewall settings for bot challenges, confirm key content is in the initial HTML and look for AI crawler requests in server logs.
How often should I run an AI visibility audit?
Run a full audit quarterly and track priority prompt clusters weekly or monthly. Re-audit after major changes such as a rebrand, pricing change, product launch or significant AI model update.
What should I fix first after an AI visibility audit?
Fix access and accuracy issues first. If crawlers are blocked or AI systems describe you incorrectly, new content will not help much. Then address content gaps and third-party source gaps in order of business value.
Can I do an AI visibility audit manually?
Yes, for a small prompt set. Manual audits require clean sessions, repeated runs and careful recording of mentions, citations and sources. As the prompt set and number of models grow, automated monitoring makes repeated sampling practical.
Conclusion
A good first AI visibility audit answers three questions: how often AI systems mention and recommend you, whether what they say is accurate, and which sources and technical factors explain the result. Repeated sampling, source analysis and an access check turn a vague sense of AI presence into a prioritized plan.
Start with a narrow scope, one market, four models and 30 prompts, and document the method so you can repeat it. The first audit sets the baseline for everything that follows. Bob Builds AI can help automate the sampling and keep the audit running continuously.