Most brands approaching AI visibility inquire the incorrect question: How do we get stronger arsenic an entity truthful that LLMs urge america more?
In entity SEO, we thin to say, “Build the Knowledge Graph, adhd schema, and get much press.” But that logic assumes the LLM is evaluating the marque and deciding whether it’s bully capable to urge for immoderate query related to what the marque sells. The LLM evaluates the query and matches it against immoderate class associations it has built for the marque from third-party content.
The quality matters enormously successful practice.
As we’ve seen successful aggregate scenarios, recognition isn’t the aforesaid arsenic recommendation. So being a recognized marque isn’t synonymous pinch being a beardown brand.
What matters is whether the class your customers are utilizing to hunt for you matches the class the LLM has coded you into.
What the information showed
João da Silva and I conducted a study of 12 diversion apparel brands successful the U.K. complete 7 days, pinch 14,140 API runs crossed ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews. We tested the aforesaid brands utilizing 2 different class framings: athleisure and diversion footwear.
After looking astatine the results derived from co-mentions and putting numbers connected the effect of framing connected class nickname for LLMs, we took the trial 1 measurement further and changed the class registry successful the prompt.
The results were symmetric to a grade that rules retired noise:
| New Balance | 64,235 | 1% | 90% | +89 | Jumped (footwear-coded) |
| Nike | 25,996 | 77% | 90% | +13 | Small displacement (footwear-coded pinch beardown athleisure co-mentions) |
| Alo Yoga | 3,062 | 63% | 0% | -63 | Dropped (athleisure-coded) |
| lululemon | 810 | 90% | 0% | -90 | Dropped (athleisure-coded) |
| Sweaty Betty | 751 | 9% | 0% | -9 | Stable |
| Reebok | 665 | 1% | 20% | +19 | Small shift |
| Outdoor Voices | 455 | 26% | 0% | -26 | Small shift |
| Rhone Apparel | 400 | 5% | 0% | -5 | Stable |
| Varley | 381 | 6% | 0% | -6 | Stable |
| TALA | 356 | 5% | 0% | -5 | Stable |
| Gymshark | 277 | 37% | 0% | -37 | Dropped (athleisure-coded) |
| LNDR | 2 | 0% | 0% | 0 | Stable |
Notes:
- New Balance goes from 1% to 90%.
- lululemon goes from 90% to 0%.
The variety is astir 0.9 points successful some directions simultaneously.
We’re not talking astir relationship here, but a controlled observation: what happens erstwhile we alteration only 1 adaptable — the class connection successful the prompt.
See wherever your marque appears successful AI search, wherever competitors are winning, and what it takes to go the reply AI recommends.
Why this happens: Category coding
Nike, New Balance, and Reebok stock the nonstop aforesaid Google Knowledge Graph (KG) description: “Footwear company,” truthful each 3 are recognized perfectly by each LLM we tested. From an entity standpoint (recognition), they commencement from an identical position. However, their behaviour nether different class framings isn’t identical astatine all.
The logic is what the insubstantial formalizes arsenic class coding: the operation of the KG explanation section and the third-party contented corpus that has accumulated astir a marque successful a fixed category.
The KG explanation anchors a marque to a class successful the model’s practice (impacts recognition).
The third-party corpus — articles, reviews, editorial comparisons, and roundups — fills successful the item of what that class relation really looks for illustration (impacts recommendation).
Looking astatine the illustration from New Balance
New Balance’s KG explanation says “Footwear company,” and the third-party corpus that has accumulated astir it corroborates the class by focusing connected topics related to moving shoes, capacity footwear, and diversion training.
When a personification asks astir athleisure brands, the exemplary doesn’t find New Balance successful that corpus because location is nary third-party association. But it does find lululemon, Alo Yoga, and Gymshark: each brands whose corpus is built from manner publications, manner editorial, and activewear roundups.
When we changed the query to diversion footwear, the retrieval flipped: New Balance is abruptly successful the correct corpus, and lululemon is not.
The exemplary itself can’t and isn’t making a judgement astir marque value aliases belonging. What an LLM does is pattern-match a query class against a contented category. If those 2 things align, the marque surfaces. If they don’t, it doesn’t, sloppy of really established the marque is.
So, tin you conscionable recode your KG description?
Some brands reference this will see the evident shortcut: Change the KG description. If “Footwear company” is anchoring you to the incorrect category, recode it to “Apparel company,” and the problem is solved.
However, the KG explanation is only half of what determines class coding. The different half is the third-party contented corpus that has accumulated astir your brand, and that doesn’t alteration because you updated a section successful the Knowledge Graph. If your full outer contented history is capacity footwear, running, and diversion training, changing the explanation gives the exemplary a caller anchor pinch thing attached to it. The corpus still says what it ever said.
The corrective lever is third-party contented finance successful the circumstantial class framing your customers are using: successful the publications the exemplary retrieves from, alongside the brands that already specify that space. The KG explanation tin support that activity erstwhile the corpus exists.
Get the newsletter hunt marketers trust on.
See terms.
What this intends for your GEO strategy
The modular GEO proposal is to fortify your entity: a accordant name, cleanable schema, a beardown About page, and much property coverage. That proposal is correct for getting recognized and moreover recommended wrong the brand’s coded category, but it isn’t capable for getting recommended successful adjacent class queries.
What determines proposal successful adjacent categories is whether the third-party contented corpus astir your marque matches the class framing your customers are really using.
The questions worthy asking astir immoderate marque are:
- Are we visible successful AI?
- What class has the LLM coded america into?
- Is that the class our customers are querying?
If a marque is beardown successful 1 category, but its customers are progressively utilizing adjacent class connection to hunt (for example, athleisure alternatively of sportswear, aliases capacity wellness alternatively of fitness), and the brand’s third-party corpus hasn’t kept gait pinch that connection shift, the marque will beryllium invisible successful precisely the queries customers are using.
Nike is the study’s clearest affirmative case, surfacing successful some athleisure (77%) and diversion footwear (90%) queries, contempt being KG-coded arsenic a footwear brand.
The logic is that Nike has accumulated capable athleisure-coded third-party content, including editorial sum successful manner publications, inclusion successful activewear roundups, and co-mentions pinch different athleisure brands, to registry arsenic category-eligible successful some framings. It built a sub-stream successful the adjacent class that New Balance didn’t.
Track your visibility crossed AI search, uncover missed opportunities, and turn your beingness wherever customers are asking questions.
What is the audit mobility everyone should beryllium asking?
Before investing further successful entity optimization, it’s worthy moving a elemental diagnostic: Take the 5 aliases six different ways your customers mightiness building a class query for what you do, and trial each 1 crossed 2 aliases 3 LLMs. Note which formulations aboveground your marque and which don’t.
For the ones that don’t, the questions to inquire are:
- Does third-party contented astir your marque really usage that language?
- Are you being written astir successful publications that screen that category?
- Are you appearing successful editorial roundups that usage that phrasing?
If the reply is no, you cognize wherever to start: getting into the outer conversations that speak the connection of that query.
Closing that spread intends becoming a subordinate successful the class comparison contented that defines who belongs successful that space.
This article is based connected findings from “The recognition-recommendation gap: Empirical grounds that class coding, not knowledge-graph strength, determines marque visibility successful generative AI output,” co-authored pinch João da Silva and published unfastened entree connected Zenodo.
English (US) · 
Indonesian (ID) ·