I have spent decades watching industries describe themselves more generously than the facts allow. The AI industry just got caught doing it in writing — by its own machines.

My firm, 5W, just published the AI Companies AI Visibility Index 2026 — the first public two-wave benchmark of how AI assistants describe the AI industry itself. We ran 32,200 prompts across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews, in two separate waves, four months apart. We only published what held in both.

Here is what held.

When you ask an AI assistant to recommend a model, it favors the one built by the company that owns it. ChatGPT recommended OpenAI's models 2.0x more often than rival engines did. Gemini favored Google DeepMind by 1.7x. Google AI Overviews favored Google by 1.6x.

The counter-argument is obvious: maybe those models really are the best, and the engines are simply right. Maybe. But that argument collapses against the one engine that breaks the pattern. Claude recommended Anthropic's models just 1.2x more often — the lowest self-citation lift we measured. If self-preference were purely a function of quality, every engine would land in roughly the same place. They don't. The spread is the tell.

This is not a scandal. It is a structural reality of how retrieval and training data entangle inside a production AI system. But it is a reality that every founder, every investor, and every journalist using these tools to "research the AI landscape" needs to understand. The engine you ask is not a neutral narrator. It has a parent. And in recommendation queries, it acts like one.

There is a second finding that matters more for anyone building a company in this space.

In every other industry we have benchmarked — banking, venture capital, credit cards — editorial publishers dominate the sources AI cites. In the AI industry, they don't. GitHub and ArXiv together supplied 31.2% of all citations — code repositories and research papers, functioning as primary source material.

Sit with that. For an AI company, your GitHub organization page is brand collateral. Your model cards are brand collateral. Your ArXiv cadence is a communications strategy. The companies winning Citation Share — OpenAI, Anthropic, Google DeepMind — all publish into those surfaces at a measurable rhythm. The companies with real revenue but no publishing cadence under-index regardless of how much money they have raised.

Most communications teams in this industry are still optimizing press releases and landing pages. The AI is reading the code.

The structural shift here is not subtle. The buyer — the founder, the developer, the LP, the policymaker — increasingly starts their research inside an AI engine, not a search engine. The engine returns an answer, not a list of links. And the answer is shaped by who published into the surfaces the model retrieves from — and which model you happened to ask.

You can shape that answer, or you can inherit it. Build the citation infrastructure before the crisis — not during it.

The full Index is at 5wpr.com/ai-visibility-index/ai-companies.

Ronn Torossian is Founder and Chairman of 5W, the AI Communications Firm, and publisher of Everything-PR.