Key takeaways 

• Visibility in AI search is earned through verifiable authority signals, not simulated metrics. 
• “AI SEO” dashboards don’t show real users’ prompt inputs; they infer them. 
• For all brands, developing a clear prompt query set comes first before using tools for reporting.  

A week ago, we kicked off an experiment. We took a brand-new company with zero backlinks, zero authority (what you’d call a blank slate), and set out to see if we could become visible, cited, and trusted inside AI search engines such as ChatGPT, Gemini, and Perplexity. 

Phase one was pre-content, pre-site; map buyer phrasing, audit who already shows up, and decide where our offer makes sense. We call this product-to-AI search alignment: the intersection of buyer prompting, the sources the AI trusts, and what we can build to connect the two. This is not SEO, not “what keywords are popular,” but what buyers ask in natural language, which sources the engines trust, and what we can publish to connect the two. 

But very quickly, we hit a wall. 

The AI SEO Boom and the Data Mirage 

AI-SEO tools are having their gold-rush moment. The global market is estimated to hit US $5.93 billion by 2035. And adoption is already high: 86 % of SEO professionals say they use AI in their strategy. But here’s the irony: many of those platforms still think like Google. 

On paper, it looks like evolution. In practice, we discovered a mirage. 

Here’s why: the leading AI search engines like OpenAI’s ChatGPT, Google’s Gemini, and Perplexity do not share raw prompt data with any third-party tool. Privacy, IP protection, and model security make that data inaccessible. So, when an AI SEO platform says it “tracks real prompts,” what it often means is: It infers prompts based on seed keywords, semantic expansion, inferred buyer intent, and LLM-created question variants. Those prompts are synthetic proxies, not real usage data. 

Note: Google’s “Fan-Out Queries” Don’t Change This; and Here’s Why 

Google’s new fan-out query behavior does not make users’ real AI prompts “visible.” Fan-out queries occur entirely within the model: a single user question is split into multiple internal subqueries to retrieve entities, facts, and sources before generating an answer. These sub-queries are created, executed, and discarded by Google in real time. They are not logged, exposed, or shared with publishers, SEOs, or third-party tools. Privacy constraints, model security, and intellectual-property protections still prevent access to raw prompts or retrieval paths. As a result, any tool claiming to “track AI prompts” is still inferring likely query patterns, not observing real user inputs or AI decision logic. 

The Result: AI Tools Can’t See What Buyers Actually Type 

If you’re a B2B founder or executive, it’s important to know you won’t get true volume metrics of when your buyer types in “best enterprise CRM 2025” into AI. Visibility tools don’t track actual “what someone typed in” volume (otherwise called prompts).  

As a result, AI visibility tools measure presence (whether and how often a brand appears in generated answers), not volume (how often a prompt is actually typed).  

Any “trend” or “visibility score” is therefore directional, not a reflection of real user prompt frequency. 

What’s Left to Fix This? 

You’re left either relying on large tools to auto-generate the prompts (which we don’t recommend), or manually defining and testing your own prompts. Most AI visibility dashboards blend tool-generated prompts with ad-hoc queries you can add yourself. 

The Part Most Dashboards Skip: A Credible Prompt Set  

Ultimately, AI visibility measurement means nothing if the inputs aren’t spot on. It’s important to lock a benchmarked prompt set first, then plan to measure. Our team started by making a specific, focused query universe. Using the exact factors and logical reasoning that research scientists at Google, Open AI, and academic LLM evaluation labs design benchmark datasets, we sampled buyer phrasing with a prompt-exploration tool, then ran manual tests in ChatGPT, Gemini, and Perplexity. 

Reverse Engineering Buyer Prompts 

In our analyst’s research, he found there is no specific benchmark number for choosing the exact number of prompts to test, rather than the number we need to focus on clusters to cover, and that eventually gives us the right number. 

Our prompt dataset must cover all buyer intentions, all category definitions, all types of AI reasoning, all types of brand recall, all industries, all problem types, and all competitive contexts.  

From there, we log every text: exact wording, timestamp, engine version, full answer, citations, competitors, and our in-answer position. We plan to keep prompts fixed for the whole 90-day sprint. 

NF-150: The No Fluff AI Visibility Benchmark Dataset 

Here, we detail each prompt cluster and why they are important. 

Cluster 1: Branded Direct Prompts 
These test whether AI systems can correctly recognize and describe our entity, who we are, what we do, the industries we serve, and the services we offer. This cluster should target 100% recall. 

Clusters 2–3: Semi-Branded & Category Prompts 
These measure how often our brand is recommended when the buyer does not explicitly name us. This is the same prompt space AI visibility tools use to benchmark the share of recommendations in non-branded queries. 

Clusters 4–5: Problem & Comparison Prompts 
These mirror real buyer decision language: problems to solve, alternatives, and “best of” evaluations. They surface our true competitive set and reveal how we are positioned within AI-generated lists and comparisons. 

Cluster 6: Advanced Semantic Prompts 
These go beyond basic visibility and recommendation. They test whether the model understands the playbook we sell, making this the strongest signal of deep authority. 

Conclusion: How AI SEO/GEO Tools Can Be Used to the Best Advantage 

Are tools useless? No. You just need to engineer the inputs (prompts) first and use the tools correctly. Here’s how we now categorize common “AI-SEO” tooling. 

• Prompt exploration. Helpful for surfacing the language buyers actually use. Treat volumes as directional, NOT accurate as we have mentioned earlier in this blog. The value is in the phrasing of your fixed prompt set. 
• AI-answer benchmarking. Useful once you appear sometimes. These tools help track presence, citations, and competitor frequency inside answers. 
• SEO suites with AI panes. Good for content ops and traditional search. Their AI sections can hint at where engines summarize brands, but they won’t replace manual prompt testing for a new name. 
• Competitive AI visibility tools. Good for snapshots when you already have some entity recognition. Early on, you’ll still rely on manual runs. 

FAQs 

Do AI-SEO tools track real prompts? 
No. ChatGPT, Gemini, and Perplexity don’t expose user prompt logs. Most tools infer prompts from keywords, scraped answers, and modeled demand. Treat their metrics as directional presence, not true prompt volume. 

What do these dashboards actually measure? 
Presence inside AI answers: whether you appear, how often, sometimes with citation and competitor counts. Useful for spotting patterns once you already show up, not for estimating real demand. 

So what’s the right way to read a “visibility score”? 
As a trend, not a forecast. Look for answer screenshots, citations, engine versions, and geography. If a vendor can’t show those, don’t use the score in planning. 

Does Google’s fan-out query behavior change anything? 
No. Fan-out happens inside the model. Those sub-queries aren’t exposed to you or vendors. You still won’t see real prompts. 

How do I test my brand’s visibility across AI platforms? 
Run a fixed prompt benchmark weekly across ChatGPT, Gemini, and Perplexity. Capture full answers, timestamps, and engine versions. Score presence, rank, citations, and competitor frequency. 

How do I define “low AI visibility”? 
You appear in fewer than half of relevant unbranded answers, branded answers are wrong or thin, citations are weak, or you lose to the same competitors across multiple runs and regions. 

Why do engines hallucinate about my company? 
Signals are thin or conflicting. Fix the public record first: precise About and Service pages, consistent schema, aligned off-site profiles, and credible third-party citations. Re-test with branded prompts until stable. 

What should early-stage brands do first? 
Treat “prompt data” as market research. Publish clear answers, earn citations where engines already look, and only then lean on dashboards to monitor presence.