We Tested 40 AI Prompts: This Brand Was Invisible in 30
A 23K-session local business gets 95% engagement from ChatGPT — but zero AI discovery visibility. We cross-referenced GA4, GSC, and 40 real LLM prompts to find out why.
The bottom line: We ran 40 real prompts through ChatGPT and Gemini for a multi-location service business with 23,000 monthly sessions. When users mentioned the brand by name, AI recommended it 100% of the time with positive sentiment. When they asked generic discovery questions — "best clinic in Madrid," "where should I go in Getafe" — the brand didn't appear in a single response. GA4 and Google Search Console data tell the exact same story. This is what invisible AI discovery looks like, and it's fixable.
Methodology
This case study uses real data from a client who gave us access to their GA4 and Google Search Console accounts. We tested 40 prompts across ChatGPT and Gemini, covering 4 categories: discovery (10), comparison (10), use case (10), and problem-solving (10). All numbers are from a 30-day window in March 2026. The business is anonymized but the data is exact.
1 The Gap: 0.3% vs 16% AI Traffic Share
We manage several websites that are optimized for LLM visibility. When we compared their AI traffic share to the client's site, the difference was stark.
The client's site gets 60x more total traffic than our smallest site — but roughly the same number of AI-sourced visits. Their 23,000 monthly sessions come primarily from paid social (30%), organic search (30%), and direct (23%). AI contributes just 64 sessions. Our much smaller sites, built with LLM visibility in mind, pull 62-86 AI sessions each.
Volume wasn't the problem. Discoverability was.
2 The Test: 40 Prompts, 2 Providers, 4 Categories
We used our AI visibility scanner to test how ChatGPT and Gemini respond to prompts a potential customer would actually type. We organized them into four categories:
| Category | Example Prompt | What It Tests |
|---|---|---|
| Discovery | "Best [service] in Madrid with advanced technology" | Can AI find you without your name? |
| Comparison | "[Brand] vs [Competitor A] vs [Competitor B]" | Does AI know your strengths? |
| Use case | "I'm in [city] looking for [specific need]" | Does AI recommend you for real scenarios? |
| Problem-solving | "I have [condition], where should I go?" | Does AI trust your expertise? |
The results were binary. Not gradual — binary.
When the Brand Was Named (Comparison)
- 10 out of 10 prompts — mentioned every time
- Sentiment: 70% positive, 30% neutral, 0% negative
- Strengths cited: clinical versatility, medical-grade technology, subscription model, effective for resistant cases
- Only weakness: "higher base price per session"
When the Brand Was Not Named (Discovery, Use Case, Problem-Solving)
- 0 out of 30 prompts — never mentioned
- Competitors recommended instead: Dorsia, Centros Ideal, Láserum, Hedonai
- Even in cities where the client has physical locations, AI didn't know they existed
- Both ChatGPT and Gemini showed the same blind spot
AI knows the brand and likes it. But it only recommends what it can find through structured, extractable content — and this site wasn't providing that.
3 GA4 Confirms: ChatGPT Traffic Is the Highest Quality Channel
The 64 monthly sessions from ChatGPT aren't just a curiosity metric. They're the best traffic this business gets.
| Source | Sessions | Engagement Rate | Pages/Session |
|---|---|---|---|
| ChatGPT | 63 | 95.2% | 2.7 |
| Google Organic | 6,677 | 68.3% | 2.6 |
| Direct | 5,313 | 61.8% | 2.3 |
| Facebook Paid | 6,797 | 54.3% | 2.8 |
| Site Average | 23,007 | 61.9% | 2.7 |
Engagement rate in GA4 measures the percentage of sessions that lasted more than 10 seconds, had a conversion event, or included 2+ page views. It's the clearest signal of whether a visitor actually did something meaningful.
ChatGPT visitors engage at 95% — 33 percentage points above the site average and 27 points above Google organic. These aren't casual browsers. They arrive with intent, because an AI told them this business was worth visiting.
The problem: all 63 sessions come from users who already knew the brand name. The comparison prompts drive this traffic. The discovery prompts — the ones that would bring new customers — send zero.
4 GSC Confirms: 40,000 Impressions, 0.5% CTR
Google Search Console told the same story from a different angle. The client's blog content ranks for high-volume informational queries — but converts almost none of them into clicks.
Three blog posts alone account for over 62,000 monthly impressions at positions 4-8 — solidly on page one — but collectively generate fewer than 500 clicks. The branded queries ("business name + location") sit at position 1 with healthy CTR. The non-branded queries are where the gap lives.
This mirrors the LLM finding exactly: strong brand recognition, weak generic discoverability. Google and AI models are telling us the same thing through different signals.
5 What Our Scanner Found: Score 71/100 (Grade C)
The LLMGeoKit scanner evaluates websites on how well LLMs can extract, understand, and cite their content. This site scored 71/100 — a C grade. Here's where it lost points:
| Category | Score | Issue |
|---|---|---|
| Metadata | 15/15 ✓ | Title, description, OG tags all present |
| Content Structure | 20/20 ✓ | Proper heading hierarchy, semantic HTML |
| Robots & Crawling | 10/10 ✓ | Sitemap, robots.txt, allows crawlers |
| Schema Markup | 10/15 | Has JSON-LD but no FAQ schema |
| Citations & Attribution | 12/15 | No author attribution (meta or JSON-LD) |
| Extractability | 4/15 | No FAQ sections, no data tables, no definition lists |
| llms.txt | 0/10 | File doesn't exist |
The traditional SEO boxes are all checked — metadata, structure, crawling. The failures are all in LLM-specific areas: extractability, structured FAQ content, author authority signals, and llms.txt.
This explains the binary result: AI models can read the site fine when pointed to it (comparisons work), but can't discover it through structured extraction (discovery fails). The content exists but isn't formatted for how LLMs find and recommend businesses.
6 The Specific Fixes (Not Theory — From the Data)
Based on the scanner results and cross-referencing with what worked on our own LLM-optimized sites, here are the concrete changes that would move this business from invisible to recommended:
1. Add llms.txt (Impact: High, Effort: 30 minutes)
An llms.txt file tells AI models what your business does, where you operate, and what you're an authority on. Think of it as a robots.txt for LLMs. This site has nothing — AI models are guessing based on scraped content. With a clear llms.txt listing services, locations, and expertise, discovery prompts have something concrete to match against.
2. Add FAQ Schema to Key Pages (Impact: High, Effort: 2-3 hours)
The site has 6 accordion/dropdown elements but no FAQPage schema. LLMs heavily prioritize FAQ-structured content for extraction — it's the easiest format for an AI to parse into a recommendation. Adding schema markup to existing FAQ content gives LLMs the structured signal they need.
3. Add Author Attribution (Impact: Medium, Effort: 1 hour)
No <meta name="author"> or JSON-LD author field exists. When an AI decides whether to cite a source as authoritative — especially for health and service businesses — author signals matter. Adding the medical director's name and credentials to the schema builds the trust signal LLMs use to decide who to recommend.
4. Create Geo-Specific Landing Pages (Impact: High, Effort: 1-2 days)
The business has 10+ locations across multiple cities and regions. But the LLM test showed zero mentions in city-specific discovery prompts — even in cities where they have a physical presence. Competitors with dedicated city pages are winning those prompts. Each location needs a structured page with local schema, services, and FAQ content that AI can extract for geo-specific recommendations.
Total estimated effort: 1-2 days of implementation. These aren't content rewrites or marketing campaigns. They're structural changes to how existing content is formatted and annotated.
7 What This Means for Your Business
The shift from traditional search to AI-assisted discovery isn't coming — it's already here. This case study shows what happens to a successful, well-trafficked business that hasn't adapted:
Where They Stand
- 23,000 sessions/month (healthy business)
- Strong brand — AI knows and likes them
- 64 AI sessions with 95% engagement
- But 100% of AI traffic is brand-dependent
- Zero AI discovery for new customers
What's at Stake
- Competitors appear in AI answers they don't
- As AI traffic grows, the gap compounds
- Blog has 62K monthly impressions going unconverted
- 1-2 days of structural fixes could change the equation
- Every month without changes is traffic to competitors
The data is clear: if you're only visible when people already know your name, you're leaving the fastest-growing discovery channel entirely to your competitors.
Want to know where your business stands? Our free AI visibility scan tests your website against the same criteria used in this case study. In under a minute, you'll see your score, your gaps, and the specific fixes that matter for your business.