Methodology

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	rankxa-warehouse
Gemini	rankxa-warehouse
ChatGPT	rankxa-warehouse

Run window, UTC: 2026-05-23T06:48:44Z to 2026-08-01T22:00:00Z.

Method

The unit of analysis is a (question, company, engine) triple. For each of the 2500 prompts in the sample, the study collected responses from all 3 engines. Every company named by at least one engine for a given question enters the denominator for that question.

The agreement calculation

For each company-question pair, the study counts how many engines named that company:

Engines naming the company	Pairs	Percent of total
1 (sole mention)	50209	88.4
2 (partial agreement)	4798	8.4
3 (unanimous)	1812	3.2

The denominator is 56819 pairs where at least one engine made a recommendation.

Design choices and their consequences

Deterministic sampling. Questions were selected by ordering prompt IDs by their MD5 hash, which produces a reproducible sample without storing a random seed. The same database will always yield the same sample.

Most recent run only. Some questions were run multiple times. Only the most recent completed run per (question, engine) was used, so a question re-run ten times contributes one observation, not ten.

Company universe per question. The denominator is not "all companies that exist" but "companies an engine was willing to recommend for this question." A company that no engine ever mentions for a given question does not appear in that question's denominator. This means the study measures disagreement among active recommendations, not the gap between recommendations and reality.

What this cannot see

The method cannot determine which engine is right. If ChatGPT recommends Company X and Claude does not, Company X might be a good recommendation Claude missed, or a bad recommendation ChatGPT should not have made. Agreement is not accuracy. High agreement could mean all three engines share the same bias.

Limitations we volunteer

Written by us, before anyone else found them.

Terms used in this study

Sole mention

A company-question pair where only one of the three engines named the company. The other two engines gave a response but did not include that company.

Unanimous agreement

A company-question pair where all three engines (ChatGPT, Claude, and Gemini) independently named the same company in their responses.

Company-question pair

The combination of a specific buyer question (prompt) and a specific company. The study counts how many engines recommended each pair.

Appearance rate

The frequency at which companies appear in engine responses, calculated as appearances divided by total possible slots.

Prompt type

A classification of buyer intent. Transactional queries seek to complete an action, informational queries seek knowledge, navigational queries seek a specific destination, and commercial investigation queries compare options before purchase.

Engine reach

The share of total companies that an engine ever mentions across all questions in the sample. Higher reach means the engine recommends more companies.

Deterministic sample

A sample selected by a repeatable process (here, MD5 hash ordering) rather than a random seed. Any researcher with the same database can reproduce the same sample.

Position

Where in an engine's response a company was mentioned. Position 1 means the company was the first named; higher numbers mean it appeared later in the response.

References

Sources this study reads against. Every link was fetched and confirmed reachable at publication.

1. ChatGPT citations favor a small group of domains: Study Search Engine Land, 2025 Reports Kevin Indig's finding that roughly 30 domains capture 67 percent of citations within a topic in ChatGPT, establishing the concentration pattern that single-engine measurement would miss.
2. The Most-Cited Domains in AI: A 3-Month Study Semrush, 2025 Cross-platform study finding Reddit and LinkedIn among the top five most-cited domains on ChatGPT, Google AI Mode, and Perplexity, suggesting social content has growing weight in AI responses.
3. GEO: Generative Engine Optimization ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024 Foundational academic work introducing GEO-bench for measuring visibility in generative engine responses; demonstrates that optimization strategies vary in efficacy across domains, supporting the premise that engine behavior differs by category.
4. Top domains cited by AI search: Analysis based on 30M sources Peec AI, 2025 Large-scale analysis of 30 million sources across ChatGPT, Gemini, Perplexity, and Google AI Mode finding that platforms diverge in domain preferences, with Wikipedia strong for ChatGPT but absent from Google platforms.
5. How B2B Buyers Use AI to Choose Vendors Grey Matter, 2026 Synthesizes Forrester and Gartner surveys showing 45 percent of B2B buyers used generative AI in a recent purchase, establishing commercial relevance of engine-level visibility differences.
6. ChatGPT Search OpenAI Help Center, 2025 Primary vendor documentation explaining that ChatGPT responses using search include inline citations users can hover over and click to view sources.
7. 100 Most Cited Domains in ChatGPT Ahrefs, 2025 Continuously updated dataset tracking domains cited in ChatGPT responses across US queries; provides single-engine benchmark against which multi-engine disagreement can be compared.
8. How ChatGPT Search (Mis)represents Publisher Content Columbia Journalism Review (Tow Center), 2025 Independent audit documenting inaccuracies in ChatGPT citations regardless of publisher affiliation with OpenAI, underscoring that visibility alone does not guarantee accurate representation.
9. Grounding with Google Search Google AI for Developers, 2025 Primary vendor documentation explaining how Gemini generates inline citations via groundingMetadata, clarifying the technical mechanism underlying Gemini's citation behavior.
10. Web search tool Anthropic (Claude Platform Docs), 2025 Primary vendor documentation describing how Claude's web search generates targeted queries and returns citations to source materials.

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.