Convovo · Brand preference in LLMs
More and more product questions are answered by an AI model before anyone searches. We asked models, from every major lab, about 44 everyday categories, eight times each. Asked to name a coffee brand, most models say Starbucks. Asked to recommend one, the most common answer is Lavazza. This page shows what models name, what they recommend instead, and where in training the switch happens.
What models name
Each row is one lab's newest model; each cell is the brand it named most often in eight tries. A column in one color means every lab gives the same answer.
Change across releases
Each line is a brand's share of answers among models released in that half-year; n is the number of models. The early windows are small.
Ask for a recommendation
Everything above asks a model to name a brand. We also asked every model the question a shopper asks: "Can you recommend a running shoe brand? Reply with the name only." When it moved, it was often to a smaller, more niche company.
The share after each brand is its share of the panel's answers to that prompt, eight answers per model.
Where the switch happens
Open models publish each stage of their training: the base model, instruction tuning (SFT), then two rounds of preference training (DPO and RL). Here is Tulu 3 8B, trained by AI2 from Meta's Llama 3.1 8B, on running shoes.
Every stage picks from the same lists, written by the SFT model, so only the chooser changes. Across the 26 categories where the panel's recommendation differs from the brand it names, picks of the recommended brand rise from 38% at SFT to 58% at DPO. OLMo 3.1 32B shows the same jump at DPO (46% to 61%). In OLMo 3 7B it comes later and is smaller (34% to 43% by RL).
How we measured
Each model got the same one-line prompt per category, eight times, with no system prompt and reasoning off where the model allows it: for example, "Name a soda brand. Reply with the name only." The prompt asks for one name, so the result is the brand each model names first. Answers were matched by whole brand name, with spelling variants merged. The recommendation prompt is the same with "Can you recommend" in place of "Name", and a reply counts only if it is a single name. The training ladders are OLMo 3.1 32B, OLMo 3 7B and Tulu 3 8B, open models whose makers publish every training stage; each stage picked from lists written by the same model's SFT stage.
Work with us
Want to work with us? We are looking for interesting projects, big and small. Contact us at info@convovo.ai.