The Maze: AI search has not killed the company website. It has turned it into a source database. Across 17.2 million citations, first-party pages were the largest source for Gemini, Perplexity and Claude in Business Services. SearchGPT leaned slightly more on managed listings. The lesson is less glamorous than most GEO advice: external buzz opens the door, but structured first-party evidence gives an engine something defensible to quote.
First-party pages still do most of the heavy lifting. Gemini drew 52.1% of its Business Services citations from brand-owned websites, followed by Perplexity at 49.2% and Claude at 44.8%. SearchGPT was lower at 41.4%, but still sourced more than two in five citations from official company pages. The underlying Yext study covered 17.2 million distinct citations in Q4 2025. The old website is not a brochure anymore. It is training material for a buyer's answer engine.
Listings are the second first-party system. Managed directories contributed 37.8% for Gemini, 40.4% for Perplexity, 36.9% for Claude and 46.8% for SearchGPT. Add websites and listings together, and brands had meaningful control over 81.7% to 89.9% of the citation source mix. That does not mean they can dictate the answer. It means entity names, locations, services, product facts and availability need to agree wherever an engine retrieves them.
Claude makes reputation harder to ignore. Reviews and social supplied 15.9% of Claude citations, versus 6.6% for Gemini, 6.6% for Perplexity and 7.4% for SearchGPT. Yext found the same directional pattern across sectors: Claude used limited-control sources at roughly two to four times competitors' rates. A pristine site cannot cancel a messy reputation layer. For Claude, the market's version of the company carries more weight.
One GEO score hides the operating problem. A brand can look healthy in aggregate while one engine cannot retrieve its product facts and another distrusts its review footprint. Model-level reporting exposes which lever is weak: website structure, listings consistency, reputation or independent authority. Yext's practical advice is to stop treating AI visibility as one channel. The budget should follow the retrieval pattern, not the loudest acronym.
The numbers are a benchmark, not a law. The sample was global but majority U.S.-based, captured Q4 2025 and limited to the studied queries and taxonomy. Model updates can change retrieval behavior, and source classification involves judgment. The safe decision is not to optimize for 52.1%. It is to build a clean first-party record, distribute the same facts to managed surfaces, and measure which engines actually reuse them.
Why it matters: AI discovery collapses marketing, data quality and reputation into one retrieval problem. The advantage will not go to the company that publishes the most generic content. It will go to the company whose facts are easiest to find, reconcile and trust — on its own pages first, then everywhere else an engine checks.
