Two people ask an AI engine the same practical question, which brand should I buy, and get different brands back. The only difference between them is that one said "athleisure" and the other said "athletic footwear." That one-word swap is the entire variable in a study Search Engine Land wrote up on July 21, 2026, by researchers Maryanna Franco and João da Silva, published open access on Zenodo (DOI 10.5281/zenodo.20331344). They put 12 athletic apparel brands through ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews for seven days: 14,140 API runs, two category framings, nothing else moved.
lululemon was recommended in 90% of answers under "athleisure" queries and in 0% of answers under "athletic footwear." New Balance ran the opposite way, 1% to 90%. Of the twelve brands, Nike was the only one the engines recommended at scale under both framings, at 77% and 90%. The brands' knowledge-graph profiles were identical throughout. The same entity data produced opposite answers, decided by which category word the question happened to carry.
Entity-SEO budgets are generally set on the premise that knowledge-graph strength is what earns an AI recommendation. This study gives that premise something checkable to be tested against, which is worth doing before the next tranche of spend goes out this month.
Sources: Search Engine Land, "How category framing changes which brands AI recommends," July 21, 2026; Franco & da Silva, "The recognition-recommendation gap," Zenodo, 2026 (DOI 10.5281/zenodo.20331344).
Credibility: Medium. The method and the sample are disclosed, and the DOI resolves to an open-access record, so the 14,140-run design is checkable rather than asserted. What keeps it off high: this is a single study, not yet independently replicated or peer-reviewed in a journal; its scope is one vertical, UK athletic apparel; and one outlet has covered it so far, Search Engine Land, under Franco's own byline, which makes the researcher and the reporter the same party, with SEL's named editors in between.
Fact bullets:
- Published July 21, 2026 via Search Engine Land; underlying paper DOI 10.5281/zenodo.20331344, publicly checkable.
- Sample: 12 athletic apparel brands, 5 AI engines (ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews), 14,140 API runs over 7 days, 2 category framings tested (athleisure vs. athletic footwear).
- New Balance: 1% recommendation rate under "athleisure" queries vs. 90% under "athletic footwear" queries (+89 points).
- lululemon: 90% under "athleisure" vs. 0% under "athletic footwear" (-90 points). Nike: 77% vs. 90% (+13 points), the only brand recommended at scale under both framings.
Stay ahead of AI search
The changes that move your rankings and AI citations — and the exact move to make — before they cost you. Free, three times a week.
For teams running an AI-visibility or entity-SEO program: the diagnostic question changes shape. The old one was "is our knowledge graph complete." The live one is "which category vocabulary has the model matched us against." A brand can be described accurately and completely in an engine's entity data and still never surface, because the buyer is asking about a category the model has coded it out of. An entity audit will not show you that, because nothing in the entity data is wrong.
Keep the study's size in view before anyone reaches for the budget: a single vertical inside a single market, run across seven days, with nobody having replicated it yet. What it justifies is an afternoon. Finding out whether your own brand carries the problem costs a few dozen queries, and the answer comes back one phrasing at a time.
Recommended actions:
1. Write down the five or six category phrasings your customers actually use for your product, then run each one through ChatGPT, Gemini and Perplexity this week and log the recommendation rate per phrasing.
2. Ask your content team to audit which category words your existing material already uses, on-site and off, including reviews, roundups and comparison articles. If the corpus talks "athleisure" while buyers ask about "footwear," that gap is what closes with language, not with markup.
3. Read the result against your AI-visibility tracker's per-prompt breakdown rather than its aggregate share-of-voice number. An aggregate can sit flat while one category framing collapses underneath it, and the per-prompt view is what tells you which one did.
The signal that this is your problem, and not just a study about sportswear, is a per-prompt breakdown where the same brand is near-absent under one category phrasing and everywhere under another.
Search Engine Land: "How category framing changes which brands AI recommends," by Maryanna Franco, edited by Angel Niñofranco, reviewed by Danny Goodwin, published July 21, 2026. https://searchengineland.com/category-framing-brands-ai-recommends-482715
Franco, M. & da Silva, J.: "The recognition-recommendation gap: Empirical evidence that category coding, not knowledge-graph strength, determines brand visibility in generative AI output," Zenodo, 2026. DOI: https://doi.org/10.5281/zenodo.20331344