Methodology
AI Utah 100 AI-Visibility Study
Engines and exact versions
A study that does not name the model version it ran against is not reproducible, because the answer changes when the model does.
| Engine | Model version |
|---|---|
| Claude | claude-sonnet-5 |
| Gemini | gemini-2.5-pro |
| ChatGPT | gpt-5.5-2026-04-23 |
| Google AI Overview | Google AI Overview |
Run window, UTC: 2026-07-31T00:20:28.878843+00:00 to 2026-07-31T01:00:08.619947+00:00.
Method
The study began with the AI Utah 100 list published by AI Utah and Codebase. From that list, we removed mega-cap companies and universities whose brands would appear in AI answers regardless of any optimisation effort. The exclusion list is published as a visible column in the dataset, not applied as a silent filter. After exclusions, 47 companies remained.
Prompt design
For each company, we wrote six buyer-intent prompts based on the company's own website description of what it sells. Four prompts were generic (e.g., "best CMMS maintenance management software"), and two were Utah-qualified (e.g., "best CMMS maintenance management software in Utah"). No prompt ever contained a company name. The test is whether AI engines recommend you when someone shops your category, not whether the model has heard of you.
This produced 282 distinct prompts across the cohort.
Engines tested
Each prompt was sent to four engines:
| Engine | Model Version |
|---|---|
| ChatGPT | gpt-5.5-2026-04-23 |
| Claude | claude-sonnet-5 |
| Gemini | gemini-2.5-pro |
| Google AI Overview | Live web panel |
All queries used web search enabled because an ungrounded model tests memory rather than the product a buyer actually uses.
Extraction and matching
Many companies in the cohort have names that are common English words: Grain, Seer, Pattern, Artifact, Latitude, Domo, Phi, Jaunt. To avoid false positives, a fixed extraction model converted every answer into an ordered list of the vendors it recommended. Matching ran against that structured list rather than raw text. Two lookalike false positives were caught in manual review.
The run window was 31 July 2026, producing 1296 total queries. After removing cases where no AI Overview was shown (18) and extraction failures (5), 1107 usable answers remained.
Limitations we volunteer
Written by us, before anyone else found them.
- Single pass. Run-to-run variance is not characterised.
- 18 queries returned no Google AI Overview panel. Excluded from that engine's denominator rather than counted as absences.
- Gemini's cited sources are largely unavailable through Google's API, so source analysis rests on the other engines.
Terms used in this study
- AI visibility
- Whether and where a company appears in AI-generated responses to non-branded buyer-intent queries in its category.
- Buyer-intent prompt
- A search query phrased as a prospective buyer would phrase it, asking for recommendations in a product or service category without naming specific vendors.
- Grounded model
- An AI model that retrieves live web content when generating responses, as opposed to relying solely on training data. All engines in this study used web search.
- Generic prompt
- A category search query without geographic qualification, such as 'best CMMS maintenance management software.'
- Utah-qualified prompt
- A category search query that specifies Utah, such as 'best CMMS maintenance management software in Utah.'
- Mention position
- The rank at which a company appears in an AI-generated recommendation list. Position one is first; position five is fifth.
- Displacement
- When a company does not appear in AI responses for its category, the competitors that appear instead are said to displace it.
- Third-party source
- A website other than a company's own domain that mentions the company, such as analyst reports, review aggregators, Wikipedia, or professional networks.
- Extraction model
- An automated process that converts unstructured AI responses into structured lists of recommended vendors for consistent matching and analysis.
- AI Overview
- Google's AI-generated summary panel that appears at the top of some search results, synthesising information from web sources.
References
Sources this study reads against. Every link was fetched and confirmed reachable at publication.
- GEO: Generative Engine Optimization arXiv / ACM KDD 2024, 2024 Foundational academic paper formalizing generative engines and introducing GEO-bench; demonstrates that content optimization can boost visibility by up to 40 percent in generative engine responses.
- 2X Survey Finds 96% of B2B Companies Are Invisible in AI Discovery Demand Gen Report, 2025 Industry benchmark measuring AI visibility across the B2B buyer journey; contextualizes the scale of the invisibility problem for brands not optimized for generative retrieval.
- DerivateX Study Finds B2B SaaS Companies Are Invisible to AI-Assisted Buyers Demand Gen Report, 2025 Directly comparable methodology: tested 50 B2B SaaS companies across four AI platforms with 1,400 buyer-intent prompts; found Claude mentions 88 percent of tested brands versus 100 percent for ChatGPT and Gemini.
- ChatGPT Search OpenAI Help Center, 2025 Vendor documentation confirming ChatGPT responses using search include inline citations; establishes that ChatGPT's brand recommendations are web-grounded rather than purely parametric.
- The Most-Cited Domains in AI: A 3-Month Study Semrush, 2025 Large-scale citation study tracking more than 230,000 prompts across three LLMs over 13 weeks; establishes baseline domain preferences for AI citation behavior.
- AI Overviews Wikipedia, 2026 Reference entry noting Google AI Overviews appear on more than 48 percent of total Google Search queries as of March 2026, establishing the feature's reach in search.
- 100 Most Cited Domains in ChatGPT Ahrefs, 2025 Domain-level citation analysis using Brand Radar tool; found Reddit captures 16.7 percent of ChatGPT citations, with Wikipedia, Amazon, Forbes, and Business Insider also prominent.
- AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information Profound, 2025 Comparative analysis of citation patterns across platforms; reports Wikipedia is ChatGPT's most cited source at 7.8 percent of total citations, while Reddit leads for Google AI Overviews at 2.2 percent.
- Web search tool Anthropic (Claude Platform Docs), 2025 Official documentation confirming Claude's web search returns citations from search results, establishing the mechanism by which Claude surfaces brand mentions.
Data
The complete row-level dataset is published open and ungated under CC BY 4.0. Every number in this study can be recomputed from it.
Want to know how AI answers describe you?
We run the same measurement on your category. Fifteen minutes with founder Omar Jenblat, your own numbers, no deck.