How Category Framing Shapes AI Brand Recommendations

Most brands approaching AI visibility are asking the wrong question entirely. They focus on strengthening their entity so large language models recommend them more often. In entity SEO, the standard advice is to build the Knowledge Graph, add schema, and secure more press coverage. That logic assumes the LLM is evaluating the brand on its merits and deciding whether it is good enough to recommend for any query related to what the brand sells. In reality, the LLM evaluates the query and matches it against whatever category associations it has built for the brand from third-party content.

The difference between these two approaches matters enormously in practice. As multiple scenarios have shown, recognition is not the same as recommendation. Being a recognized brand does not automatically mean being a strong brand in the eyes of AI. What truly matters is whether the category your customers use to search for you matches the category the LLM has coded you into.

João da Silva and I conducted a study of 12 athletic apparel brands in the U. K. over seven days, running 14,140 API calls across ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews. We tested the same brands using two different category framings: athleisure and athletic footwear. After analyzing the co-mention data and quantifying the impact of framing on category recognition for LLMs, we took the test one step further by changing the category register in the prompt.

The results were symmetric to a degree that rules out noise. New Balance jumped from 1% to 90% when the category shifted. lululemon dropped from 90% to 0%. The variation was approximately 0.9 points in both directions simultaneously. This is not a correlation but a controlled observation: changing only one variable, the category word in the prompt, produced dramatic swings in brand visibility.

Why does this happen? It comes down to category coding. Nike, New Balance, and Reebok share the exact same Google Knowledge Graph description: “Footwear company.” All three are recognized perfectly by every LLM we tested. From an entity standpoint, they start from an identical position. Yet their behavior under different category framings is not identical at all. The reason is what the paper formalizes as category coding: the combination of the KG description field and the third-party content corpus that has accumulated around a brand in a given category.

The KG description anchors a brand to a category in the model’s representation, impacting recognition. The third-party corpus, including articles, reviews, editorial comparisons, and roundups, fills in the detail of what that category association actually looks like, impacting recommendation. New Balance’s KG description says “Footwear company,” and the third-party corpus around it corroborates the category by focusing on running shoes, performance footwear, and athletic training. When a user asks about athleisure brands, the model does not find New Balance in that corpus because there is no third-party association. It does find lululemon, Alo Yoga, and Gymshark, all brands whose corpus is built from fashion publications, lifestyle editorial, and activewear roundups.

When we changed the query to athletic footwear, the retrieval flipped. New Balance is suddenly in the right corpus, and lululemon is not. The model itself cannot and does not make a judgment about brand quality or belonging. What an LLM does is pattern-match a query category against a content category. If those two things align, the brand surfaces. If they do not, it does not, regardless of how established the brand is.

Some brands will consider an obvious shortcut: change the KG description. If “Footwear company” is anchoring you to the wrong category, recode it to “Apparel company” and the problem is solved. However, the KG description is only half of what determines category coding. The other half is the third-party content corpus that has accumulated around your brand, and that does not change because you updated a field in the Knowledge Graph. If your entire external content history is performance footwear, running, and athletic training, changing the description gives the model a new anchor with nothing attached to it. The corpus still says what it always said. The corrective lever is third-party content investment in the specific category framing your customers are using, in the publications the model retrieves from, alongside the brands that already define that space. The KG description can support that work once the corpus exists.

The standard GEO advice is to strengthen your entity: a consistent name, clean schema, a strong About page, and more press coverage. That advice is correct for getting recognized and even recommended within the brand’s coded category, but it is not sufficient for getting recommended in adjacent category queries. What determines recommendation in adjacent categories is whether the third-party content corpus around your brand matches the category framing your customers are actually using.

The questions worth asking about any brand are: Are we visible in AI? What category has the LLM coded us into? Is that the category our customers are querying? If a brand is strong in one category but its customers are increasingly using adjacent category language to search, for example, athleisure instead of sportswear or performance wellness instead of fitness, and the brand’s third-party corpus has not kept pace with that language shift, the brand will be invisible in exactly the queries customers are using.

Nike is the study’s clearest positive case, surfacing in both athleisure (77%) and athletic footwear (90%) queries despite being KG-coded as a footwear brand. The reason is that Nike has accumulated enough athleisure-coded third-party content, including editorial coverage in fashion publications, inclusion in activewear roundups, and co-mentions with other athleisure brands, to register as category-eligible in both framings. It built a sub-stream in the adjacent category that New Balance did not.

Before investing further in entity optimization, it is worth running a simple diagnostic. Take the five or six different ways your customers might phrase a category query for what you do, and test each one across two or three LLMs. Note which formulations surface your brand and which do not. For the ones that do not, ask: Does third-party content about your brand actually use that language? Are you being written about in publications that cover that category? Are you appearing in editorial roundups that use that phrasing? If the answer is no, you know where to start: getting into the external conversations that speak the language of that query. Closing that gap means becoming a participant in the category comparison content that defines who belongs in that space.

This article is based on findings from “The recognition-recommendation gap: Empirical evidence that category coding, not knowledge-graph strength, determines brand visibility in generative AI output,” co-authored with João da Silva and published open access on Zenodo.