The strongest brand entity in the study got zero recommendations — because the question used the wrong category word

The source, and what to discount

The primary source is *Beyond Knowledge-Graph Strength: How Category Coding Drives Brand Visibility in Generative AI*, by Maryanna Franco and Joao da Silva, version 1.0 dated 19 May 2026, deposited on Zenodo under CC-BY 4.0 and written up by Search Engine Land on 21 July 2026. Two caveats belong in the first paragraph, not a footnote: the paper states on its own cover page that it is not peer reviewed, and its authors are affiliated with an AI-visibility vendor — as are we. The design, however, is the kind that survives that discount: a single manipulated variable, a fixed brand set, a fixed geography, five engines, and a run count large enough that the extremes are not noise.

What changed, and what didn't

Twelve brands were put to the five engines under two category framings — "athleisure" and "sportswear" — with everything else held constant. Recognition (does the engine know the brand when you name it?) stayed near-universal. Recommendation (does the engine bring the brand up unprompted in a category question?) did not:

The rank correlation between Knowledge Graph strength and surfacing within a brand's own coded category was +0.43 — real but modest. The correlation between Knowledge Graph strength and the recognition-to-recommendation gap across categories was −0.10: effectively nothing. Authority buys you presence in the category the engines have already filed you under. It buys you nothing in the one next door.

The mechanism: your own website is not the input

The paper's explanation is what it calls category coding: the Knowledge Graph short-description field plus the third-party content corpus that has accumulated around the brand within a given category determine which category the model files the entity under. The supporting measurement is the one worth pinning to the wall — across all ten brands with citation data, the brand's own domain accounted for under 2% of cited sources in recommendation responses, with a maximum of 1.96% (Reebok). lululemon: 53 own-brand citations against 4,076 third-party. Alo Yoga: 2 against 3,446. New Balance, in the category it does not surface in, had 156 third-party citations and zero of its own.

This matches what we measure locally — with one honest difference

We have measured the same two gaps first-party, on local businesses rather than apparel brands. On recognition versus recommendation: across 168 local businesses in 41 US metros, the engines named the business in 98.8% of answers when asked about it by name, and in 19.8% when asked the buyer's question — a 5x gap, with 58% of those businesses appearing in none of their four buyer-intent answers. That is the recognition–recommendation gap, in a different vertical, with a different method, landing in the same place.

On the corpus: across 511 ChatGPT web-search answers about 172 local businesses, the business's own website was cited in 28% of answers — the engine built the other 72% from review aggregators and "best-of" directories, averaging 5.5 sources per answer. Note the difference before you stack the numbers: the Zenodo paper reports own-domain share of all citations (<2%), we report the share of answers containing at least one own-domain citation (28%). Different denominators, same conclusion — the page you control is a rounding error in the evidence the engine actually reads.

Three things this changes for anyone selling or buying AI visibility

1. 1The category word in the prompt is a variable, and most audits hold it fixed by accident. If a check tests "best sportswear brands" and your buyers ask about "athleisure," the result is not wrong — it is answering a different question. Ask both, and expect them to disagree.
2. 2Entity/authority work is necessary but not sufficient. A +0.43 within-category correlation and a −0.10 cross-category one means schema, Wikipedia and brand-search volume defend the ground you hold. They do not take new ground.
3. 3The lever is the third-party corpus, not the site. Under 2% of citations from own domains says the work is getting written about in the category you want to be recommended in — directories, roundups, review platforms, community threads — which is ongoing work, not a one-time site fix.

The replicable part of all this is cheap to check for a single business: put the buyer's question — in the category word your buyers actually use — to more than one engine and record which names come back. That is what our free AI visibility check does.