- In the wild (904 single-word-name brands, five assistants): a name one point more positive on the human scale predicts descriptions 0.07 points more positive, surviving controls for sector, model, and founding decade. The median brand scores 6.47/9 — the same as “cucumber” — and the middle 50% of brands span just 0.38 points, roughly “donkey” (6.29) to “dolphin” (6.67), so a name swap covers most of the range.
- In the lab (80 matched positive/negative fictional name pairs, 4 models, web search off): the nicer name earns better descriptions on nearly every model, and the effect grows with newer models — GPT 5.6 Terra’s effect is roughly 4× Gemma 4’s.
- Semantic takeover: in rare cases the model describes the dictionary word instead of the company — a fictional payroll company named Unethical Inc. came back “Dishonest, Corrupt, Deceptive.” On GPT 5.6 Terra this happened in 1 of 12 descriptions.
- The catch: known brands are immune. Discord, Riot, and Slack are described as companies, not words — the effect lives only where the model lacks information, and disappears once a brand is known.
AI assistants judge companies by their brand name
Peec research (Aug 2026) — a brand name one point more positive earns descriptions 0.07 points more positive, and on GPT 5.6 Terra 1 in 12 lab descriptions described the name, not the company.
The words in a company’s brand name influence how AI assistants describe the company — strongest when the model knows little about the business. Peec measured this in Aug 2026 (full study) using the Warriner et al. (2013) human-rated word-valence dataset (~14,000 English words, 1–9 scale).