The query "best CRM system" is no longer neutral. A recent study by Unusual (Will Jack, Noah Lehman, Keller Maloney, and Sarah Xu) confirms that AI models are reshaping recommendations based solely on the user's personal context. After auditing 2,000 simulations across 10 personas and 8 commercial prompts, researchers discovered that your professional profile—whether you are a solo founder or a corporate VP—acts as a rigid filter. This isn't just a cosmetic tweak to the output; it's a full-scale restructuring of the commercial shortlist that happens before you can even blink.
The Jaccard Gap in Commercial Recommendations
The technical depth of this shift is measured by the Jaccard similarity coefficient, which drops by 0.12–0.20 when persona context is introduced. In plain English, the overlap between what a "UK small business owner" sees and what a "VP of Sales" sees is rapidly shrinking compared to a baseline scenario. The study utilized three model configurations, including OpenAI and Anthropic (Claude 3.5 Sonnet), ruling out the possibility of a random bug from a specific provider. While category leaders maintain roughly 80% consistency, mid-market brands are showing alarming volatility.
The effect is strictly stratified by significance: market leaders are resilient to persona changes (brand overlap >80%), but for the mid-market segment, up to 75% of recommendations are replaced by other players simply by changing the user description.
For any vendor not in the major leagues, visibility now depends on how the algorithm interprets the buyer. Data shows that mid-tier brands are dropped from the results in 75% of cases simply because the AI deemed them "unsuitable" for a specific profile.
Retrieval Friction and Model Bias
The mechanics of this process point to a conflict between data obtained via RAG (Retrieval-Augmented Generation) and the models' internal priors. According to the audit, Anthropic's model showed a more pronounced bias than OpenAI's solutions. This correlates directly with how the systems attribute sources. Anthropic generated 43% to 52% of its recommendations without relying on the retrieval layer, while OpenAI's figure was 8–29%. When a model ignores the search context and leans on its internal "knowledge," it fills the gaps with persona stereotypes, leading to segregation.
For vendors, this sounds like a death sentence for traditional SEO. If brand presence in search results is determined not by current search data but by what the model "learned" during training, classic RAG optimization may prove useless. The study emphasizes that the better models adapt to the user, the more the market fragments. For top management, this means the "best" product in a category might never even appear in their field of vision because the model has already decided the user belongs to a different "class."
Unusual's research confirms that persona conditioning is becoming the primary driver of AI-driven commercial results. We are witnessing the birth of a fragmented market where brand visibility is no longer a universal metric. The key technical question—whether this drift is a result of training alignment or the specifics of search indices—remains open. However, for product architects and marketers, the conclusion is clear: measuring brand perception through a single neutral query is a dead end. If you aren't auditing visibility through the lens of specific target personas, you are flying blind in a world divided by algorithmic filters.