Category Framing Significantly Impacts AI Brand Recommendations

The way a brand is framed within a category significantly influences its recommendation rate by large language models (LLMs), according to a recent study. The research, conducted by João da Silva and an unnamed collaborator, challenges the conventional SEO approach of solely focusing on building a brand's knowledge graph and schema. Instead, it highlights that LLMs primarily match user queries to category associations they have built for a brand from third-party content, rather than evaluating the brand's inherent quality.

The study analyzed 12 athletic apparel brands in the U.K. over seven days, performing 14,140 API calls across platforms including ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews. Researchers tested the same brands using two distinct category framings: "athleisure" and "athletic footwear." The results demonstrated a clear impact of this framing on category recognition for LLMs. For instance, New Balance, when framed as "athletic footwear," achieved a 90% footwear rate and a 1% athleisure rate, a significant shift of +89% towards footwear coding. Conversely, Alo Yoga, coded as "athleisure," saw its footwear rate drop to 0% from 63% in the athleisure category, a decrease of -63%.

Similar dramatic shifts were observed for other brands. Lululemon, also coded as athleisure, experienced a -90% drop in its footwear rate from 90% to 0%. Gymshark dropped by -37% when framed within the athleisure category. Nike showed a smaller shift of +13%, indicating it was footwear-coded but with strong athleisure co-mentions. The data suggests that a brand's recognition is not synonymous with its strength; rather, its recommendation depends on the alignment between the category customers use in their searches and the category the LLM has assigned to the brand.

This distinction is crucial for brands aiming for AI visibility. The study's findings indicate that focusing on how a brand is perceived and categorized within specific search contexts is more effective than solely enhancing its entity-level data. The symmetrical nature of the results across different brands and framings suggests that these outcomes are not due to random noise but represent a consistent pattern in how LLMs process and recommend brands based on category associations.