Most marketers building for AI visibility are solving the wrong problem. They're focused on brand authority — more backlinks, stronger entity signals, richer schema markup — under the assumption that if an LLM knows your brand well enough, it will recommend you. New research suggests that assumption is fundamentally broken.
The actual mechanism is simpler and more actionable: AI models don't evaluate your brand and decide if it's good enough. They pattern-match your brand's content category against the category implied by the user's query. If those two things align, you surface. If they don't, you're invisible — regardless of how well-known you are.
The Experiment That Proves Category Coding Is Real
Research published on Search Engine Land by Maryanna Franco and João da Silva makes this concrete. They ran 14,140 API queries across ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews, testing 12 U.K. athletic apparel brands under two different category framings: athleisure and athletic footwear. Only one variable changed — the category word in the prompt.
The results weren't subtle. lululemon went from a 90% recommendation rate under "athleisure" queries to 0% under "athletic footwear" queries. New Balance did the opposite — jumping from 1% to 90% when the framing shifted to footwear. Nike, with strong third-party content in both categories, held relatively stable across both.
What makes this finding particularly sharp is that Nike, New Balance, and Reebok share the exact same Google Knowledge Graph description: "Footwear company." From a pure entity-recognition standpoint, they start from an identical position. Yet their recommendation behavior diverges dramatically under different framings. The Knowledge Graph description anchors brand recognition — but it doesn't drive recommendation. That job belongs to the third-party content corpus: the articles, reviews, editorial comparisons, and roundups that have accumulated around a brand and associated it with a specific category in the model's training data.
lululemon's corpus is dominated by fashion publications, lifestyle editorial, and activewear roundups — athleisure content. When a query invokes the footwear category, there's no matching signal for the model to retrieve. The brand effectively doesn't exist in that context.
What This Means for Content Strategy and GEO
For performance marketers running AI-assisted lead-gen or content programs, this research reframes the entire optimization problem. Generative Engine Optimization (GEO) isn't primarily about entity strength — it's about category alignment.
Here's the practical implication: the language you use across your landing pages, ad copy, content briefs, and PR pitches is actively shaping which category an LLM associates you with. If there's a mismatch between the category your customers use when they search and the category your content corpus signals to AI models, you will lose recommendation share — even if your brand is well-known and well-structured.
Consider a B2B SaaS example. A marketing automation platform that has built most of its content around "email marketing" may be systematically absent from AI recommendations when prospects search for "revenue operations tools" or "go-to-market automation" — even if the product handles those use cases perfectly. The tool is athleisure; the query is footwear. The model never makes the connection because the third-party corpus never made it either.
This is why prompt-aware content strategy matters. Every piece of content you publish — owned or earned — is a data point that trains the model's category association for your brand. A product page that uses the same category language your buyers use in their prompts isn't just good copywriting. It's GEO infrastructure.
Actionable Takeaways for Performance Marketers
- Audit your category language. Pull your top landing pages, ad copy, and cornerstone content. What category terms are you repeatedly using? Now map those against the actual queries your ICP uses when searching for solutions — in AI tools, not just traditional search. Identify gaps.
- Run your own category framing test. Query ChatGPT, Claude, or Perplexity for your product category using 5–10 different phrasings. Note when your brand appears and when it doesn't. The queries where you're absent are your content gaps.
- Target the third-party corpus deliberately. AI recommendations are heavily influenced by what others write about you. Prioritize earned coverage in publications, roundups, and comparison pieces that use the category language your buyers use. A review that calls you an "AI-powered revenue automation platform" is more valuable than one that calls you a "marketing tool" if your buyers are searching the former.
- Reframe content briefs around query categories, not just keywords. When briefing writers or AI content tools, specify the category framing you're optimizing for — not just the topic. This shapes the language density around specific category signals throughout the piece.
- Don't assume product breadth protects you. Nike's stability across both framings comes from years of deliberately building corpus in both categories. Most brands don't have that. If you sell into multiple use cases or buyer personas, each one likely needs its own category-coded content strategy.
- Watch for category drift in AI outputs. If competitors are generating content that associates your category with their brand specifically, they're actively building the corpus signal that will displace you in AI recommendations. Monitor co-mentions and category associations in AI tools the same way you'd monitor SERP rankings.
The New Optimization Surface
The brands that will win AI recommendation share aren't necessarily the ones with the biggest budgets or the strongest domain authority. They're the ones whose content — owned, earned, and syndicated — most precisely mirrors the language their buyers use when prompting AI tools.
Recognition was the game in traditional SEO. Category alignment is the game in generative search. The marketers who internalize this shift now — and build content infrastructure that speaks the exact category language of their buyers — will compound a structural advantage that's extremely difficult to reverse once it's established in a model's learned associations. The window to build that corpus intentionally is open. It won't stay that way.
