I run a firm that measures how brands get cited inside AI answers. We publish an index series. We've measured banking, venture capital, credit cards, defense, beauty, luxury. Nothing we've published has landed like the number we just found inside the AI industry itself.
The number is 1.2x. That's how much more often Claude recommends Anthropic models than the other four AI engines recommend Anthropic models. In a category where ChatGPT recommends OpenAI 2.0x more often than rival engines do, where Gemini recommends Google DeepMind 1.7x more often, where Google AI Overviews recommends Google 1.6x more often, and where Perplexity surfaces Perplexity at 2.4x baseline inside adjacent search queries — Claude's 1.2x reads as an outlier.
The 5W AI Communications team ran two independent waves. 32,200 prompts. Five engines. January–February 2026 and April–May 2026. The self-citation lift held in both waves within 0.4x. It is not a sampling artifact. The full study is at 5wpr.com/ai-visibility-index/ai-companies.
Every AI assistant most consumers touch has a commercial parent. Four of the five parents get promoted by their own product. One does not.
Why the finding is uncomfortable for the industry
Ask ChatGPT which AI to use for your startup. Ask Gemini which model to build on. Ask Perplexity which tool to compare. Each engine is doing the vendor selection interview and is a vendor. That conflict was theoretical a year ago. It is now measurable.
Buyers don't see the lift. They see one answer. When a founder types "best AI model for customer service" into ChatGPT and OpenAI's names come back with unusual frequency, the founder does not read a disclaimer — because there isn't one. Same for Gemini. Same for Google AI Overviews. Same for Perplexity inside search queries.
Claude is the exception at 1.2x. Anthropic's product recommends Anthropic — but at a margin closer to the actual quality signal than to a marketing preference. Whether that reflects training data composition, retrieval architecture, or a deliberate design choice is a question for Anthropic to answer. The behavioral outcome is measured and public.
The other finding worth reading
The source stack that AI assistants pull from when answering about AI companies is unlike any other category we've measured. Wikipedia (24.3%), GitHub (17.8%), and ArXiv (13.4%) together supply 55.5% of citations. Code repositories and research papers are the primary sources.
In banking, editorial publishers dominate. In venture capital, editorial publishers dominate. In credit cards, publishers dominate. In AI, engineers publish the citations. GitHub organization pages, model cards, README files, and release notes are brand assets — the assistants read them directly. Most AI companies do not treat them that way. They treat them as engineering documentation. The assistants don't care about the distinction.
Companies with sustained ArXiv cadences over-index in technical queries by 2.4x. Companies without them are functionally invisible in the developer-intent cluster regardless of revenue. This is the operational read for anyone running comms at an AI company.
Five companies own 68.5% of the answer
OpenAI at 24.6%. Anthropic at 14.8%. Google DeepMind at 13.7%. Meta AI at 9.7%. xAI at 5.7%. The top five account for more than two-thirds of cited responses across the five engines. The remaining 31.5% splits across 20 other labs. Concentration this tight is unusual — it exceeds what we've measured in financial services, defense, or beauty.
xAI is the interesting underperformer. Enormous share of voice in tech and business press through the last twelve months. 5.7% Citation Share. Social-media reach did not translate into AI citation. The mechanism is straightforward — X activity does not generate the GitHub, ArXiv, and Wikipedia source material the assistants retrieve from.
Anthropic over-indexes cleanly inside safety and alignment queries. 31.2% share in that cluster — more than double its overall citation share. The clearest brand-positioning signal in the dataset.
What this means for buyers
The AI assistant is now the vendor-selection interview. It is also, in four cases out of five, the vendor. Buyers should assume a measurable house lift when the answer is a recommendation. Not a conspiracy — a documented behavioral pattern reported inside the 5W AI Companies AI Visibility Index 2026 and confirmed across two waves.
Cross-check. Ask two engines. Ask three. When they agree, the recommendation is stronger than the house lift. When they disagree, the house lift is the story.
What this means for AI companies
Wikipedia is the universal anchor. GitHub is a brand surface. ArXiv cadence is a communications strategy. Model cards, benchmarks, and structured technical content on your own domain are citation-eligible; landing pages and marketing copy are not.
The companies at the top of our index — OpenAI, Anthropic, Google DeepMind — publish research at a measurable cadence in surfaces the assistants read. The companies at the bottom rely on press coverage that never converts to citation.
The AI answer is the new shelf. Every answer is an ad. Four of the five engines are advertising their parent. Now you know the ratios.
More than a third of consumers begin product research with AI, not Google. Start there.
Ronn Torossian is shaping AI — and the answers inside the chatbox. Founder and chairman of 5W AI Communications, the AI Communications Firm. Publisher of Everything-PR. Author of two best-selling editions of For Immediate Release. He is shaping the answers inside ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews.