How the category structure changes is what types of AI it recommends

Many products approaching the realization of AI are asking the wrong question: How do we become strong as a business so that LLMs recommend us more?

In the SEO business, we often say, “Build a Knowledge Graph, add a schema, and get more traffic.” But that assumption assumes that the LLM evaluates the brand and decides whether it is good enough to recommend in any question related to what the product is selling. LLM evaluates the query and matches it with any organizations that have created a product from third-party content.

The difference is very important in practice.

As we have seen in many cases, recognition is not the same as praise. So being a well-known brand is not synonymous with being a strong brand.

What’s important is that the category your customers use to search for you matches the category LLM has coded for you.

That’s what the data showed

João da Silva and I conducted a survey of 12 athletic clothing stores in the UK during seven days, with 14,140 API runs on ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews. We examined similar brands using a two-category framework: running and athletic shoes.

After looking at the results from co-referencing and quantifying the impact of framing on LLMs’ category recognition, we took the test further and changed the category register immediately.

The results are symmetrical with respect to the noise level:

Notes:

• New Balance from 1% to 90%.
• lululemon from 90% to 0%.

The difference is about 0.9 points in both directions at the same time.

We are not talking about correlation here, but a controlled view: what happens if we change only one variable — the class name in the notification.

Why this happens: Category coding

Nike, New Balance, and Reebok share the exact same Google Knowledge Graph (KG) definition: “Shoe company,” so all three are well recognized by all the LLMs we tested. From a business perspective (recognition), they start from the same place. However, their behavior under different hierarchical frameworks is not the same at all.

The reason is what the paper officially does as a category code: a combination of the KG description field and a corpus of third-party content piled up around the product in a particular category.

The definition of KG concentrates the product in the stage of model representation (visualization of effects).

A third-party corpus – articles, reviews, editorial comparisons, and roundups – fills in the details of what that organization actually looks like (recommendation of implications).

We look at an example from New Balance

KG’s definition of New Balance is “Shoe company,” and the third-party corpus it has accumulated confirms the category by focusing on topics related to running shoes, performance shoes, and athletic training.

When a user asks about athletic products, the model does not find New Balance in that chorus because there is no third-party association. But it finds lululemon, Alo Yoga, and Gymshark: all brands whose corpus is built through fashion publications, lifestyle editorials, and activewear roundups.

When we changed the question to athletic shoes, the return changed: New Balance was suddenly in the right corpus, and lululemon was not.

The model itself cannot and does not make a judgment about the quality of the product or its ownership. What LLM does is match the query class pattern against the content class. When those two things go together, a product appears. If they don’t, they don’t, no matter how established the brand is.

So, can you just rewrite your description of KG?

Some brands reading this will consider the obvious shortcut: Change the definition of KG. If “Shoe Company” attaches you to the wrong category, retype “Clothes Company,” and the problem is solved.

However, the definition of KG is only part of what determines the coding of a category. The other part is the third-party content corpus that has been added to your product, and that doesn’t change just because you updated a field in the Knowledge Graph. If your entire history of outdoor content is work shoes, running, and athletic training, changing the description gives the model a new anchor without anything attached to it. The corpus still says what it has always said.

The adjustment lever is the investment of third-party content in a specific category that your customers use in the frame: in the publication that the model downloads from, next to the products that already define that space. A KG definition can support that task once the corpus is available.

What does this mean for your GEO strategy

General GEO advice is to strengthen your business: a consistent name, a clean schema, a strong About page, and more media coverage. That advice is good enough to be seen and even recommended within the product coding section, but not enough to get recommended in the adjacent section questions.

What determines a recommendation in the adjacent categories is whether the corpus of third-party content around your product matches the category your customers actually use.

Questions to ask about any product are:

• Are we visible to AI?
• What category does the LLM put us in?
• Is that a category our customers are asking about?

If a brand is strong in one category, but its customers are increasingly using a side language to search (for example, sports instead of sportswear, or performance instead of fitness), and the third-party corpus does not match that language change, the product will not appear in the queries used by customers.

Nike is the clearest subject in the study, appearing in both athletics (77%) and athletic footwear (90%) questions, despite being coded KG as a footwear brand.

The reason is that Nike has accumulated enough third-party content with sports code, including editorial coverage in fashion magazines, inclusion in activewear, and collaborations with other athletic brands, to register as category-worthy in both categories. It created a sub-stream in the near future that New Balance didn’t.

What research question should everyone be asking?

Before investing further in business development, it is worth using a simple diagnosis: Take five or six different ways that your customers might say a segment question about what you do, and test each of them in two or three LLMs. Note which forms appear in your product and which do not.

For those who don’t, the questions to ask are:

• Does third-party content about your product use that language?
• Are you writing about it in books covering that category?
• Do you appear in editorial groups that use those phrases?

If the answer is no, you know where to start: tapping into foreign conversations that speak the language of that question.

Bridging that gap means participating in the comparative content of the category that defines who belongs to that gap.

This article is based on the findings of “The recognition recommendation gap: Empirical evidence that category encoding, not knowledge graph strength, determines product visibility in generative AI production.,” co-authored with João da Silva and published open access on Zenodo.