What co-mentions reveal about the AI recommendation gap

We’ve spent the final two years optimizing for AI visibility by specializing in what we are saying about ourselves: writing higher About pages, including clear schema and SameAs markup, structuring content material extra successfully, and offering extra direct solutions.

All of those ideas nonetheless apply and are important for the qualification phase of an LLM’s model processing (readability + relevance). However a research João da Silva and I carried out utilizing Friction AI’s platform places a quantity on an element the trade has been circling round however couldn’t show.

Amongst manufacturers that had been already acknowledged (the place the LLM might describe them precisely), Information Graph (KG) energy predicted visibility inside the class every model was coded to. What it didn’t predict was whether or not a model would floor in an adjoining class question, even when it belonged there from a enterprise perspective. In different phrases, recognition didn’t assure suggestion. That’s the framing gap.

What manufacturers did we check and the way did we check them?

For this case research, we examined 12 athleisure and activewear manufacturers throughout ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews: 14,140 API runs over seven days, utilizing UK geography with internet search enabled.

For every model, we ran two forms of prompts:

• Recognition prompts (“What’s [Brand]?” and “Describe [Brand]”)
• Suggestion prompts (“Finest athleisure manufacturers,” “Prime 10 athleisure manufacturers,” and “Which athletic attire manufacturers are value shopping for in 2026?”)

The manufacturers spanned three Information Graph tiers, assigned by Google KG resultScore (the uncooked rating returned by Google’s Information Graph Search API — a proxy for the way strongly an entity is established in Google’s index), so we might check whether or not KG energy predicted suggestion visibility:

• Low KG: LNDR, TALA, Gymshark, Varley.
• Mid KG: Reebok, Outside Voices, Rhone Attire, Sweaty Betty.
• Excessive KG: Alo Yoga, Nike, lululemon, New Steadiness.

Spoiler forward: The high-KG manufacturers didn’t dominate suggestions. The mid-KG tier confirmed the most important common hole between recognition and suggestion.

Inside the high-KG tier, some manufacturers had been universally beneficial, whereas others had been almost invisible in suggestion prompts, regardless of being completely acknowledged throughout each LLM we examined.

Suggestions: What the co-mention information confirmed

We mapped how typically manufacturers appeared collectively in athleisure content material throughout exterior sources (articles, critiques, comparability items, and editorial lists) crawled by way of API from UK-indexed sources.

A number of the most fascinating outcomes embody:

• 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 manufacturers seem collectively repeatedly in the identical articles, roundups, and editorial comparisons throughout impartial sources. Collectively, they kind a cluster that the LLM treats as “athleisure.”

Now take a look at the opposite finish of the spectrum. New Steadiness co-occurs with lululemon in athleisure content material so not often that it doesn’t seem within the high pairs in any respect. Nike co-occurs with lululemon roughly 50 occasions extra typically than New Steadiness does.

Nike, New Steadiness, and Reebok share the very same Google Information Graph description: “Footwear firm.” From an entity standpoint, they begin from the identical place. However Nike is contained in the athleisure cluster. New Steadiness and Reebok are solely exterior it.

The LLM isn’t evaluating these manufacturers independently and deciding which of them match athleisure. It’s pattern-matching towards associations constructed from exterior content material. If a model hasn’t appeared persistently alongside lululemon, Alo Yoga, and Gymshark within the content material the mannequin skilled on — or retrieves from — it doesn’t belong in that cluster as a result of the semantic affiliation was by no means constructed.

Nike, the hero: Similar KG description, utterly totally different outcomes

Nike, New Steadiness, and Reebok share the identical KG entity description: “Footwear firm.” LLM probing throughout all 5 programs assigns all three unanimously to the athletic_footwear class, so from a pure entity-clarity standpoint, they begin from the identical place.

Nonetheless, their suggestion charges in athleisure queries aren’t remotely equal.

Nike surfaces in 71% of athleisure suggestion prompts, whereas New Steadiness and Reebok seem in 0% throughout all 5 LLMs and all 14,140 runs.

The distinction isn’t how they’re outlined (“Footwear firm”). It’s which conversations they seem in and which different manufacturers seem alongside them.

LLMs don’t infer class adjacency. If a model hasn’t been persistently talked about alongside the related gamers in a class — in press, critiques, editorial content material, and comparability items — the mannequin doesn’t make the leap. Jason Barnard describes this nicely: if A plus B ought to equal J, it’s important to assemble that path explicitly. The mannequin received’t construct it for you.

New Steadiness’s co-mention density lives in working and efficiency content material. No one constructed the semantic bridge from working → athletic life-style → athleisure in exterior content material, so the mannequin doesn’t cross it. The Information Graph says “Footwear firm,” and the third-party corpus confirms footwear. Athleisure queries retrieve the athleisure corpus, and New Steadiness isn’t in it.

The third-party quotation weight in suggestion vs. recognition information

After we cut up citations by immediate kind — recognition vs. suggestion — a sample emerges that ought to reframe the place most GEO budgets are being spent.

For recognition prompts — the place the person has already typed your model identify — own-brand content material is the dominant supply:

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

That is the place your About web page and homepage are used for readability, and your providers, class, and information pages are used for relevance.

Suggestion prompts give us utterly totally different outcomes. When the person hasn’t named your model and is asking for the best choice in a class, own-brand citations drop to 18% on ChatGPT and to successfully zero on Gemini, Claude, Perplexity, and Google AI Overviews. Third-party sources account for 82% to 100% of what will get cited throughout all 5 programs.

The GEO group has argued for a while that exterior indicators matter greater than on-site optimization for suggestion visibility, and this information places particular numbers behind that argument. It additionally reveals that exterior indicators aren’t all the identical factor.

• Entity readability will get a model acknowledged. That’s an issue you remedy by yourself web site.
• Exterior credibility will get it thought of. That’s a PR and corroboration drawback.
• Co-mention density in the suitable class cluster locations a model within the idea graph for a selected suggestion question. That’s a category-positioning drawback. 

These are three separate issues that require totally different options. Conflating them is why many GEO suggestions cease brief.

The sensible addition to any GEO audit is that this: after checking entity readability and exterior credibility, audit the place you seem in relation to others.

• Are your press mentions itemizing you alongside your precise class opponents?
• Do the roundups that embody you additionally identify the manufacturers that dominate your goal class? 

If not, the LLM has in all probability by no means discovered to affiliate you with that class as a result of it has by no means seen you in that “firm.” Not like entity readability or schema, it’s not one thing you may repair by yourself web site. That’s the hole.

What the co-mention construction means for PR and content material technique

As we’ve seen thus far, being talked about in a class isn’t sufficient. Being talked about alongside the suitable manufacturers in a class is what locations you within the idea graph for that cluster.

A press point out that describes a model as “efficiency attire” in isolation does little to advance its athleisure idea graph placement. 

A press point out that lists it alongside lululemon, Alo Yoga, and Gymshark in an editorial comparability does significantly extra as a result of it builds the co-occurrence sign the mannequin must affiliate the model with that cluster.

The identical logic applies throughout content material kind.

Editorial roundups and comparability items 

Being included in “better of” lists that identify your class opponents is value extra to your idea graph than a standalone model profile. The cluster sign comes from showing in the identical article because the manufacturers that outline the class.

Podcast appearances

If the host introduces you in relation to particular named manufacturers, or compares your method to a class chief, that co-occurrence will get listed. 

A bio that claims “founding father of [Brand], which competes with lululemon and Gymshark within the premium athleisure house” does totally different work than a bio that claims “founding father of [Brand], a efficiency attire firm.”

Analyst and trade reviews

Class-level reviews that group manufacturers collectively are high-signal co-mention sources. Being included in a sector evaluation alongside your class friends locations you within the idea graph in a method that standalone protection doesn’t.

Retailer and comparability taxonomy

Being stocked and categorized alongside class leaders in a significant retailer’s taxonomy is a co-mention sign. The retailer’s class web page is exterior content material that locations your model in a cluster.

The objective is visibility in the suitable firm.

A notice on the info and what comes subsequent

This research covers a single class — athleisure and activewear — with 12 manufacturers examined within the UK. The co-mention figures are uncooked co-occurrence counts from UK-indexed sources crawled by way of API, overlaying content material listed on the time of the research in Might 2026. Cross-category validation and extra geography testing are in progress.

The total paper, “The Recognition-Recommendation Gap: Empirical Evidence That Category Coding, Not Knowledge-Graph Strength, Determines Brand Visibility in Generative AI Output,” has been printed by João da Silva and me on Zenodo and paperwork the methodology, model pattern, immediate set, and extraction code in enough element for impartial replication.

However the sample within the co-mention information is evident sufficient to behave on now. Three manufacturers share the identical Information Graph description: one seems in 71% of athleisure suggestion responses, and two seem in 0%. The structural distinction is co-mention density in category-aligned third-party content material.

The query value asking about any model is that this: Within the content material that talks about your class, are you within the room, and are you in the suitable firm?