Victor CariouCo-founderWhat the study found
A single word in your prompt can change which brand an LLM recommends.
Researchers at Cornell Tech asked about a hundred LLMs to name a brand in 42 categories, and most of the time they gave the market leader. When they asked them to recommend one and pick instead, the models often switched. Trello became Asana, Chrome became Firefox, Nike became Brooks.
In total, asking for a recommendation instead of a name changes the winning brand in 16 categories out of 42. When the model recommends a few brands and then has to pick one, it's 26.
So why does the model switch?
It comes from how it's trained.
First, it learns from a huge amount of web content, which is where it learns which brands exist. The leaders dominate at that stage since they're the most written about.
Then, it's trained on pairs of answers where one was judged better, and learns to answer like the better one. That's the step where the recommended brands gain the most ground.
Yet that data barely mentions these brands: the recommended brand comes up 3 times in 260,000 examples. What the model learns there is what an expert answer looks like. So when it recommends something, it picks the brand that best fits the need among those it already knows, which is often a specialist.
How they measured it
The author ran 102 models from all the big labs through their API, without web search, on 42 categories from soda to cloud providers. Each question was asked three ways (name, choose, recommend), plus a two-step version where the model recommends a few brands and then picks one. To see when the switch happens, he also tested open models at each stage of their training.
Our critical read
| Sample | 102 models, 42 categories, 8 runs each. | |
|---|---|---|
| Freshness | Run in October 2026, on current models. | |
| Reproducibility | Method is clear, but no data or code shared. | |
| Independence | Academic, nothing to sell. | |
| SaaS fit | One-line prompts, mostly consumer brands. |
It's a solid study, with a big panel and the same pattern across models. A few limits though.
The prompts are one line long, with no context and a name-only answer. A real B2B question comes with a team size, a budget and a stack, and we don't know yet if the switch holds there.
Most categories are consumer ones. Project management software is the only real B2B SaaS category in the published results, so it still has to be confirmed on SaaS queries.
And the explanation on training is a strong hypothesis rather than a proof. It's tested on a few open models, and ChatGPT, Claude or Gemini may not work the same way. It's also a preprint, by a single author.
Applied to B2B SaaS
Track the whole spectrum of prompts
For each topic, you need prompts from list prompts like "best CRMs for startups" to choice prompts like "which CRM would you pick". That's how you get a real view of your visibility, and it also shows you how LLMs see you: as the market leader or as the expert.
Own a specific topic
As a consequence, most markets still have room for a brand that clearly owns a specific topic. Product marketing is key here.
Explain why you're the best pick
The model has learned to prefer answers that look well thought out. So when it pulls sources to answer a recommendation prompt, a page explaining why you're the best pick for a specific need fits that kind of answer much better than a list where you're one name among ten.
If you want to know where you stand, run the same test on your own category: ask ChatGPT, Claude and Gemini to name a brand, then to recommend one, then to pick.