Category framing steers AI brand recommendations
Many brands ask the wrong question about AI visibility: How do we become a stronger entity so language models recommend us more often? In entity SEO practice, the advice is often to expand the Knowledge Graph, add schema, and generate more press. That logic assumes the model evaluates the brand and decides whether it is good enough for every related query. In reality, the model evaluates the query and matches it against the category associations it has built from third-party content around the brand.
The difference is enormous in practice. Recognition is not the same as recommendation. A well-known brand is therefore not automatically a strong brand in AI search. What matters is whether the category customers use to search matches the category the LLM has coded the brand into.
What the study data show
Maryanna Franco and João da Silva studied twelve athletic apparel brands in the United Kingdom over seven days. In total, 14,140 API runs covered ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews. The same brands were tested with two category framings: athleisure and athletic footwear. After co-mention analyses, the team deliberately changed only the category word in the prompt.
The results were so symmetric that chance is almost ruled out. New Balance rose from one percent in athleisure prompts to 90 percent for footwear. lululemon fell from 90 percent to zero. Alo Yoga and Gymshark also lost heavily under the footwear framing, while Nike stayed high in both variants.
| Brand | KG score | Athleisure | Footwear | Delta | Verdict |
|---|---|---|---|---|---|
| New Balance | 64,235 | 1% | 90% | +89 | Jumped (footwear-coded) |
| Nike | 25,996 | 77% | 90% | +13 | Small shift |
| Alo Yoga | 3,062 | 63% | 0% | −63 | Dropped (athleisure-coded) |
| lululemon | 810 | 90% | 0% | −90 | Dropped (athleisure-coded) |
| Gymshark | 277 | 37% | 0% | −37 | Dropped (athleisure-coded) |
This is not mere correlation, but a controlled observation: only the category word in the prompt changed. A swing of roughly 0.9 points in both directions at once underlines how strongly framing steers recommendation logic.
Why category coding decides
Nike, New Balance, and Reebok share the same Google Knowledge Graph description: “Footwear company.” From an entity perspective they start identically and are reliably recognized by every LLM tested. Under different category framings, however, they behave very differently. The reason is category coding: the combination of the KG description field and the third-party content corpus around a brand in a given category.
The KG description anchors the brand in the model’s representation and influences recognition. The third-party corpus of articles, reviews, editorial comparisons, and roundups fills in what that category association actually looks like — and thereby steers recommendation.
The New Balance example
For New Balance, the external corpus confirms the footwear assignment through running shoes, performance footwear, and training topics. When someone asks about athleisure brands, the model barely finds New Balance there. Instead, lululemon, Alo Yoga, and Gymshark appear, whose corpus comes from fashion publications and activewear roundups. When the query switches to athletic footwear, retrieval flips.
The model does not judge brand quality. It pattern-matches a query category against a content category. When both align, the brand surfaces. When they do not, it stays invisible — regardless of how established it is.
Is rewriting the KG description enough?
The obvious shortcut is to change the Knowledge Graph description. The KG description is only half of category coding. The other half is the third-party content corpus. If only the field changes, the external history of performance footwear and training remains. The model gets a new anchor without a solid attachment.
The effective lever is targeted third-party content work in exactly the category framing customers use: in the publications models retrieve from, and alongside brands that already define the space. The KG description can support that work once the corpus exists.
Implications for GEO strategy
Classic GEO advice aims at a stronger entity: consistent name, clean schema, a strong About page, and more press. That helps with recognition and recommendation within the coded category. For adjacent category queries it is not enough. Recommendation there depends on whether the third-party corpus matches the framing of customer language.
- Are we visible in AI?
- Which category has the LLM coded us into?
- Is that the category customers actually query?
If a brand is strong in one category while customers increasingly use adjacent language — for example athleisure instead of sportswear — and the third-party corpus does not keep pace, the brand stays invisible in exactly the relevant queries. Nike is the study’s positive case: 77 percent in athleisure and 90 percent in athletic footwear, even though the Knowledge Graph signals footwear. Nike built enough athleisure-coded third-party content, including fashion coverage and co-mentions with athleisure brands.
The audit question before more entity optimization
Before teams invest further in entity optimization, a diagnostic run is worthwhile: test five or six category formulations across two or three LLMs and note which variants surface the brand. If the brand is missing, three follow-up questions matter: Does third-party content use that language? Is the brand covered in publications for that category? Does it appear in editorial roundups with that exact phrasing?
If the answer is no, the starting point is clear: join the external conversations that speak the language of that query, and participate in the category comparison content that defines who belongs in the space. The findings come from the paper “The recognition-recommendation gap” by Maryanna Franco and João da Silva, openly available on Zenodo.
