New arXiv study reveals AI assistants favor top brands but L2 challengers convert best

A new arXiv paper titled "Prominence-Stratified Failure Modes in Retrieval-Augmented Commercial Recommendation" presents a massive audit of how AI assistants like ChatGPT and Claude act as recommendation engines for brands. The researchers conducted ~37,000 production runs across four model configurations and 215 commercially-framed prompts spanning 19 sectors. They evaluated 533 brands from a reference catalog stratified into five prominence tiers (L1 category leaders to L5 regional players) based on external authority lists. The study found that AI assistants do not simply return links like search engines — they directly nominate brands, making marketing to AI a fundamentally different challenge.

The failure modes vary sharply by brand tier. L1 brands appear in nearly every relevant retrieval but only convert to recommendations 25-41% of the time, meaning visibility alone isn't enough — differentiation is the real leverage. L2 challengers have the highest conversion rates (37-52%) but suffer from persona-mediated substitution on Anthropic models, where the AI swaps them for alternatives based on user persona. L3 mid-market brands mark the inflection point: aggregate coverage drops to 88%, conversion to 34-40%, and persona effects peak. Most alarmingly, 48-52% of L4 specialists and L5 regional brands never surface in any of the 37,000 runs — catastrophic invisibility. The authors conclude that no uniform optimization recipe works; brands must tailor their AI marketing investment to their position on the prominence ladder.