Co-mentions Expose AI’s Recommendation Gap

We’ve spent the last two years optimizing for AI visibility by focusing on what we say about ourselves: refining About pages, adding clear schema and SameAs markup, structuring content more effectively, and providing more direct answers. These principles remain essential for the qualification phase of an LLM’s brand processing, where clarity and relevance are key. However, a study I conducted with João da Silva using Friction AI’s platform quantifies a factor the industry has been circling but couldn’t prove.

Among brands already recognized (where the LLM could accurately describe them), Knowledge Graph (KG) strength predicted visibility within their coded category. What it didn’t predict was whether a brand would surface in an adjacent category query, even if it belonged there from a business perspective. Recognition didn’t guarantee recommendation. That’s the framing gap.

We tested 12 athleisure and activewear brands across ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews, running 14,140 API calls over seven days from UK geography with web search enabled. For each brand, we used two prompt types: recognition prompts (“What is [Brand]?” and “Describe [Brand]”) and recommendation prompts (“Best athleisure brands,” “Top 10 athleisure brands,” and “Which athletic apparel brands are worth buying in 2026?”). The brands spanned three Knowledge Graph tiers, assigned by Google’s KG resultScore: low KG (LNDR, TALA, Gymshark, Varley), mid KG (Reebok, Outdoor Voices, Rhone Apparel, Sweaty Betty), and high KG (Alo Yoga, Nike, lululemon, New Balance).

Spoiler: The high-KG brands didn’t dominate recommendations. The mid-KG tier showed the largest average gap between recognition and recommendation. Within the high-KG tier, some brands were universally recommended, while others were nearly invisible in recommendation prompts, despite perfect recognition across every LLM tested.

We mapped how often brands appeared together in athleisure content across external sources like articles, reviews, comparison pieces, and editorial lists. Key co-mention pairs included lululemon + Alo Yoga (534), lululemon + Nike (482), Alo Yoga + Nike (449), and Gymshark + lululemon (264). These brands form a cluster the LLM treats as “athleisure.” Meanwhile, New Balance co-occurs with lululemon in athleisure content so rarely it doesn’t appear in the top pairs, and Nike co-occurs with lululemon roughly 50 times more often than New Balance does.

Nike, New Balance, and Reebok share the exact same Google Knowledge Graph description: “Footwear company.” From an entity standpoint, they start from the same position. But Nike is inside the athleisure cluster, while New Balance and Reebok are entirely outside it. The LLM isn’t evaluating brands independently; it’s pattern-matching against associations built from external content. If a brand hasn’t appeared consistently alongside lululemon, Alo Yoga, and Gymshark in the content the model trained on or retrieves from, it doesn’t belong in that cluster because the semantic association was never built.

Nike surfaces in 71% of athleisure recommendation prompts, while New Balance and Reebok appear in 0% across all five LLMs and all 14,140 runs. The difference isn’t how they’re defined; it’s which conversations they appear in and which other brands accompany them. LLMs don’t infer category adjacency. If a brand hasn’t been consistently mentioned alongside relevant players in press, reviews, editorial content, and comparison pieces, the model doesn’t make the leap. Jason Barnard describes this well: if A plus B should equal J, you must construct that path explicitly. The model won’t build it for you.

New Balance’s co-mention density lives in running and performance content. Nobody built the semantic bridge from running to athletic lifestyle to athleisure in external content, so the model doesn’t cross it. The Knowledge Graph says “Footwear company,” and the third-party corpus confirms footwear. Athleisure queries retrieve the athleisure corpus, and New Balance isn’t in it.

When we split citations by prompt type, a pattern emerges that should reframe where most GEO budgets are spent. For recognition prompts, where the user has already typed your brand name, own-brand content is the dominant source: ChatGPT cited it 49% of the time, Perplexity 36%, and Claude 23%. This is where your About page and homepage provide clarity, and your category and guide pages demonstrate relevance.

Recommendation prompts give completely different results. When the user hasn’t named your brand and asks for the best option in a category, own-brand citations drop to 18% on ChatGPT and effectively to zero on Gemini, Claude, Perplexity, and Google AI Overviews. Third-party sources account for 82% to 100% of what gets cited across all five systems. The GEO community has argued that external signals matter more than on-site optimization for recommendation visibility, and this data puts specific numbers behind that argument. It also shows that external signals aren’t all the same.

Entity clarity gets a brand recognized, a problem you solve on your own site. External credibility gets it considered, a PR and corroboration problem. Co-mention density in the right category cluster places a brand in the concept graph for a specific recommendation query, a category-positioning problem. These are three separate problems requiring different solutions. Conflating them is why many GEO recommendations stop short.

The practical addition to any GEO audit is this: after checking entity clarity and external credibility, audit where you appear in relation to others. Are your press mentions listing you alongside your actual category competitors? Do the roundups that include you also name the brands that dominate your target category? If not, the LLM has probably never learned to associate you with that category because it has never seen you in that “company.” Unlike entity clarity or schema, it’s not something you can fix on your own website. That’s the gap.

Being mentioned in a category isn’t enough. Being mentioned alongside the right brands in a category places you in the concept graph for that cluster. A press mention describing a brand as “performance apparel” in isolation does little to advance its athleisure concept graph placement. A press mention listing it alongside lululemon, Alo Yoga, and Gymshark in an editorial comparison does considerably more because it builds the co-occurrence signal the model needs.

The same logic applies across content types. Editorial roundups and comparison pieces that include your category competitors are worth more than standalone brand profiles. Podcast appearances where the host introduces you in relation to specific named brands or compares your approach to a category leader get indexed. Analyst and industry reports that group brands together are high-signal co-mention sources. Retailer and comparison taxonomies that stock and categorize you alongside category leaders also provide co-mention signals. The goal is visibility in the right company.

This study covers a single category, athleisure and activewear, with 12 brands tested in the UK. The co-mention figures are raw co-occurrence counts from UK-indexed sources crawled via API in May 2026. Cross-category validation and additional geography testing are in progress. The full paper, “The Recognition-Recommendation Gap: Empirical Evidence That Category Coding, Not Knowledge-Graph Strength, Determines Brand Visibility in Generative AI Output,” has been published on Zenodo, documenting methodology, brand sample, prompt set, and extraction code for independent replication.

But the pattern in the co-mention data is clear enough to act on now. Three brands share the same Knowledge Graph description: one appears in 71% of athleisure recommendation responses, and two appear in 0%. The structural difference is co-mention density in category-aligned third-party content. The question worth asking about any brand is this: In the content that talks about your category, are you in the room, and are you in the right company?