What co-mentions reveal about the AI recommendation gap

We’ve spent the past 2 years optimizing for AI visibility by focusing connected what we opportunity astir ourselves: penning amended About pages, adding clear schema and SameAs markup, structuring contented much effectively, and providing much nonstop answers.

All of these principles still use and are basal for the qualification phase of an LLM’s marque processing (clarity + relevance). But a study João da Silva and I conducted utilizing Friction AI’s level puts a number connected a facet the manufacture has been circling astir but couldn’t prove.

Among brands that were already recognized (where the LLM could picture them accurately), Knowledge Graph (KG) spot predicted visibility wrong the class each marque was coded to. What it didn’t foretell was whether a marque would aboveground successful an adjacent class query, moreover if it belonged location from a business perspective. In different words, nickname didn’t guarantee recommendation. That’s the framing gap.

What brands did we trial and really did we trial them?

For this lawsuit study, we tested 12 athleisure and activewear brands crossed ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews: 14,140 API runs complete 7 days, utilizing UK surface science pinch web hunt enabled.

For each brand, we ran 2 types of prompts:

• Recognition prompts (“What is [Brand]?” and “Describe [Brand]”)
• Recommendation prompts (“Best athleisure brands,” “Top 10 athleisure brands,” and “Which diversion apparel brands are worthy buying successful 2026?”)

The brands spanned 3 Knowledge Graph tiers, assigned by Google KG resultScore (the earthy people returned by Google’s Knowledge Graph Search API — a proxy for really powerfully an entity is established successful Google’s index), truthful we could trial whether KG spot predicted proposal visibility:

• Low KG: LNDR, TALA, Gymshark, Varley.
• Mid KG: Reebok, Outdoor Voices, Rhone Apparel, Sweaty Betty.
• High KG: Alo Yoga, Nike, lululemon, New Balance.

Spoiler ahead: The high-KG brands didn’t predominate recommendations. The mid-KG tier showed the largest mean spread betwixt nickname and recommendation.

Within the high-KG tier, immoderate brands were universally recommended, while others were astir invisible successful proposal prompts, contempt being perfectly recognized crossed each LLM we tested.

Recommendations: What the co-mention information showed

We mapped really often brands appeared together successful athleisure contented crossed outer sources (articles, reviews, comparison pieces, and editorial lists) crawled via API from UK-indexed sources.

Some of the astir absorbing results include:

• lululemon + Alo Yoga: 534 co-mentions.
• lululemon + Nike: 482 co-mentions.
• Alo Yoga + Nike: 449 co-mentions.
• Gymshark + lululemon: 264 co-mentions.
• Gymshark + Alo Yoga: 252 co-mentions.

These brands look together many times successful the aforesaid articles, roundups, and editorial comparisons crossed independent sources. Together, they shape a cluster that the LLM treats arsenic “athleisure.”

Now look astatine the different extremity of the spectrum. New Balance co-occurs pinch lululemon successful athleisure contented truthful seldom that it doesn’t look successful the apical pairs astatine all. Nike co-occurs pinch lululemon astir 50 times much often than New Balance does.

Nike, New Balance, and Reebok stock the nonstop aforesaid Google Knowledge Graph description: “Footwear company.” From an entity standpoint, they commencement from the aforesaid position. But Nike is wrong the athleisure cluster. New Balance and Reebok are wholly extracurricular it.

The LLM isn’t evaluating these brands independently and deciding which ones fresh athleisure. It’s pattern-matching against associations built from outer content. If a marque hasn’t appeared consistently alongside lululemon, Alo Yoga, and Gymshark successful the contented the exemplary trained connected — aliases retrieves from — it doesn’t beryllium successful that cluster because the semantic relation was ne'er built.

Nike, the hero: Same KG description, wholly different results

Nike, New Balance, and Reebok stock the aforesaid KG entity description: “Footwear company.” LLM probing crossed each 5 systems assigns each 3 unanimously to the athletic_footwear category, truthful from a axenic entity-clarity standpoint, they commencement from the aforesaid position.

However, their proposal rates successful athleisure queries aren’t remotely equivalent.

Nike surfaces successful 71% of athleisure proposal prompts, while New Balance and Reebok look successful 0% crossed each 5 LLMs and each 14,140 runs.

The quality isn’t really they’re defined (“Footwear company”). It’s which conversations they look successful and which different brands look alongside them.

LLMs don’t infer class adjacency. If a marque hasn’t been consistently mentioned alongside the applicable players successful a class — successful press, reviews, editorial content, and comparison pieces — the exemplary doesn’t make the leap. Jason Barnard describes this well: if A positive B should adjacent J, you person to conception that way explicitly. The exemplary won’t build it for you.

New Balance’s co-mention density lives successful moving and capacity content. Nobody built the semantic span from moving → diversion manner → athleisure successful outer content, truthful the exemplary doesn’t transverse 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 successful it.

---

The third-party citation weight successful proposal vs. nickname data

When we divided citations by punctual type — nickname vs. proposal — a shape emerges that should reframe wherever astir GEO budgets are being spent.

For nickname prompts — wherever the personification has already typed your marque sanction — own-brand contented is the ascendant source:

• ChatGPT cited own-brand contented 49% of the time.
• Perplexity: 36%.
• Claude: 23%.

This is wherever your About page and homepage are utilized for clarity, and your services, category, and guideline pages are utilized for relevance.

Recommendation prompts springiness america wholly different results. When the personification hasn’t named your marque and is asking for the champion action successful a category, own-brand citations driblet to 18% connected ChatGPT and to efficaciously zero connected Gemini, Claude, Perplexity, and Google AI Overviews. Third-party sources relationship for 82% to 100% of what gets cited crossed each 5 systems.

The GEO organization has based on for immoderate clip that outer signals matter much than on-site optimization for proposal visibility, and this information puts circumstantial numbers down that argument. It besides shows that outer signals aren’t each the aforesaid thing.

• Entity clarity gets a marque recognized. That’s a problem you lick connected your ain site.
• External credibility gets it considered. That’s a PR and corroboration problem.
• Co-mention density successful the correct class cluster places a marque successful the conception chart for a circumstantial proposal query. That’s a category-positioning problem. 

These are 3 abstracted problems that require different solutions. Conflating them is why galore GEO recommendations extremity short.

The applicable summation to immoderate GEO audit is this: aft checking entity clarity and outer credibility, audit wherever you look successful narration to others.

• Are your property mentions listing you alongside your existent class competitors?
• Do the roundups that see you besides sanction the brands that predominate your target category? 

If not, the LLM has astir apt ne'er learned to subordinate you pinch that class because it has ne'er seen you successful that “company.” Unlike entity clarity aliases schema, it’s not thing you tin hole connected your ain website. That’s the gap.

What the co-mention building intends for PR and contented strategy

As we’ve seen truthful far, being mentioned successful a class isn’t enough. Being mentioned alongside the correct brands successful a class is what places you successful the conception chart for that cluster.

A property mention that describes a marque arsenic “performance apparel” successful isolation does small to beforehand its athleisure conception chart placement. 

A property mention that lists it alongside lululemon, Alo Yoga, and Gymshark successful an editorial comparison does considerably much because it builds the co-occurrence awesome the exemplary needs to subordinate the marque pinch that cluster.

The aforesaid logic applies crossed contented type.

Editorial roundups and comparison pieces 

Being included successful “best of” lists that sanction your class competitors is worthy much to your conception chart than a standalone marque profile. The cluster awesome comes from appearing successful the aforesaid article arsenic the brands that specify the category.

Podcast appearances

If the big introduces you successful narration to circumstantial named brands, aliases compares your attack to a class leader, that co-occurrence gets indexed. 

A bio that says “founder of [Brand], which competes pinch lululemon and Gymshark successful the premium athleisure space” does different activity than a bio that says “founder of [Brand], a capacity apparel company.”

Analyst and manufacture reports

Category-level reports that group brands together are high-signal co-mention sources. Being included successful a assemblage study alongside your class peers places you successful the conception chart successful a measurement that standalone sum doesn’t.

Retailer and comparison taxonomy

Being stocked and categorized alongside class leaders successful a awesome retailer’s taxonomy is simply a co-mention signal. The retailer’s class page is outer contented that places your marque successful a cluster.

The extremity is visibility successful the correct company.

A statement connected the information and what comes next

This study covers a azygous class — athleisure and activewear — pinch 12 brands tested successful the UK. The co-mention figures are earthy co-occurrence counts from UK-indexed sources crawled via API, covering contented indexed astatine the clip of the study successful May 2026. Cross-category validation and further surface science testing are successful progress.

The afloat paper, “The Recognition-Recommendation Gap: Empirical Evidence That Category Coding, Not Knowledge-Graph Strength, Determines Brand Visibility successful Generative AI Output,” has been published by João da Silva and maine connected Zenodo and documents the methodology, marque sample, punctual set, and extraction codification successful capable item for independent replication.

But the shape successful the co-mention information is clear capable to enactment connected now. Three brands stock the aforesaid Knowledge Graph description: 1 appears successful 71% of athleisure proposal responses, and 2 look successful 0%. The structural quality is co-mention density successful category-aligned third-party content.

The mobility worthy asking astir immoderate marque is this: In the contented that talks astir your category, are you successful the room, and are you successful the correct company?