Most SEO audits end where the real work begins. We spend years obsessing over Google’s blue links, only to realize that the industry has shifted toward Retrieval-Augmented Generation (RAG) and direct answer engines. LLM citation tracking If your brand isn’t being cited in a ChatGPT response or surfacing in a Perplexity discovery, your traditional search rankings are becoming vanity metrics.
So, where do you start? How do you measure something as elusive as "AI visibility"? You start with a baseline. And no, one or two searches aren't enough. To get statistically significant data, you need between 20-50 prompts. Here is why that number matters and how to execute the audit without wasting time.
Why is 20-50 prompts the magic baseline number?
When I conduct an audit for a SaaS client, I see teams test three or four brand-related keywords and call it a day. That is statistically insignificant. You need a mix of head terms, long-tail problem-solution queries, and competitor comparison queries to map your visibility.
If you only test your brand name, you aren't measuring visibility; you’re measuring vanity. To understand how an LLM perceives your entity, you must test its knowledge of your solutions, your competitive positioning, and your specific use cases. Using 20-50 prompts allows you to segment your testing into three distinct buckets:
- Knowledge-based queries: "What is [Company Name]?" Comparative queries: "[Your Product] vs. [Competitor Product]" Problem-solution queries: "How to solve [X problem] using [Your Category]?"
If you aren’t appearing in at least 60% of these iterations, you have an entity optimization problem, not a keyword problem.
How does AI visibility differ from traditional SEO?
Traditional SEO is about ranking URLs. AI visibility is about entity recognition and authority. When you look at your GA4 data, you might see "Direct" or "Organic" traffic spikes, but those don't tell you *why* an AI picked your content over a competitor's. In the era of RAG, the LLM isn't just looking for keywords; it’s looking for a verified Knowledge Graph entry.
If you want to move the needle, you need to stop thinking about meta descriptions and start thinking about context. AI models build answers based on what they find in their training data and what they retrieve via live web lookups. If your site structure is messy or your schema is broken, the AI simply ignores you. Always remember: What would I screenshot to prove this changed? If you can't point to a specific citation or a change in the model’s summary of your brand, you haven't actually moved the needle.
What role does Schema.org play in AI discovery?
Broken schema is the silent killer of AI visibility. Many developers think that because their code passes a basic validator, it’s fine. It isn't. I constantly find schema that is syntactically correct but contextually bankrupt. Specifically, you need to focus on @id linking. By defining your entities (Organization, Person, Product, PerplexityBot crawler Service) and linking them via unique identifiers, you are building the map that an AI needs to follow.
Before you run your prompt audit, run your core pages through the Google Rich Results Test. If you aren't outputting valid JSON-LD that defines exactly what your brand *is*—not just what it sells—you are making it impossible for the model to associate your content with the correct entity.
Audit Component Traditional SEO Focus AI Visibility Focus Content Keyword Density Entity Context & Factuality Authority Backlink Count Co-occurrence & Citation Frequency Schema SERP Features (Rich Snippets) Entity Disambiguation (@id links)How do you perform a manual audit without burning hours?
I get asked about manual audit time constantly. If you are doing this manually, you should expect to spend 4-6 hours for a solid baseline audit across 50 prompts. If you spend longer than that, you are overthinking the data entry. Use a simple spreadsheet to track:
The Prompt (The exact query used). The Source (Which LLM—ChatGPT, Perplexity, etc.). The Citation (Did you get a link? Did you get a mention?). The Quality (Was the summary accurate?).Tools like FAII.ai are becoming essential here, as they provide automated ways to monitor how your brand appears in LLM responses. For larger organizations, I often recommend a mix of manual prompt testing to "feel" the model's logic and automated monitoring to track changes over time. Companies like Four Dots have been doing great work in the space of entity-based marketing, proving that when you focus on the entity, the visibility follows naturally.
Are you actually tracking AI referral traffic in GA4?
This is where most people fail. GA4 is built for clicks. AI is built for consumption. When a user reads a summary in an AI answer, they might not click your link. Does that mean the visibility check failed? Absolutely not. That is "Brand Awareness" in an LLM context.
However, you should set up specific UTM parameters for any content syndication you control and watch your "Referral" traffic segment closely. Keep a list of bots you block in your robots.txt file—but ensure you aren't blocking the scrapers that feed the LLMs you want to appear in. If you block everything, you forfeit your seat at the table.
What is the checklist for a successful baseline?
Before you commit to your 20-50 prompt audit, ensure your house is in order. Do not skip these steps, or your baseline will be useless:
- Verify @id Linking: Use the Google Rich Results Test to ensure your organization is explicitly defined. Audit Your RAG Readiness: Is your content clear, factual, and devoid of the jargon that models often filter out? Check Your Knowledge Graph: Search your brand on Google and Bing. If you don't have a Knowledge Panel, your AI visibility will always be lower than your competitors who do. Document the Baseline: Capture screenshots. Again: What would I screenshot to prove this changed? If the AI describes your product incorrectly, screenshot the response and use it to update your site's landing page content to clear up the confusion.
Final thoughts on AI visibility
Stop chasing algorithms and start chasing entity dominance. The 20-50 prompt baseline is your starting line. If you find that the AI consistently misidentifies your product or ignores your brand in favor of a competitor, the fix isn't "more content." The fix is better entity structure, cleaner schema, and a more focused effort on ensuring your brand is cited as the authority for the specific problems you solve.
Don't be the brand that complains about AI traffic drops while ignoring the fact that they aren't even showing up in the summary. Get the baseline, map the entities, and document the change.