Original insights by Jan Ehrlinspiel
30-second rundown
Key learnings:
- A brand name can bias an AI description when company knowledge is weak: Across 904 real one-word brands, a one-point increase in the name’s emotional score corresponded with a 0.07-point more positive company description.
- Controlled tests showed the name itself caused part of the effect: Eighty matched positive and negative name pairs produced different descriptions even when the company facts stayed neutral and identical.
- Recognition is the practical antidote: Well-known brands with negative-sounding names escaped the effect because the models knew the company instead of guessing from the dictionary meaning.
This week: Ask four major AI assistants to describe your company with ten attributes, then label each attribute as supported, unsupported or confused with the literal meaning of your name. Add that description to your About page and to the public profiles people use to verify a business, such as your Google Business Profile, LinkedIn company page, trade-association listing and relevant industry directories. Keep the core facts consistent, then correct the most damaging unsupported impression.
Coming 3 months: Use an AI Search monitoring workflow to repeat the attribute test daily across models and compare the results with a maintained set of approved company facts. Give one owner responsibility for investigating drift and strengthening the pages and credible third-party profiles that define what the company actually does.
Jan Ehrlinspiel asked AI assistants to describe hundreds of companies and discovered a shabby little prejudice hiding under the floorboards. When the system knew little about a business, the emotional meaning of its name leaked into the verdict. The robot jury saw the label, lost the evidence file and began improvising character testimony.
Usually the effect was small. At the extreme, models ignored neutral facts and described the name instead. A payroll company called Unethical Inc. came back as dishonest, corrupt and deceptive. The software interrogated the sign above the door.
1. Real brands showed a small but persistent naming effect
Ehrlinspiel tested 904 real brands whose names were single English words across ChatGPT, Gemini, Google AI Overview, Google AI Mode and Microsoft Copilot. Each assistant supplied ten attributes, which were scored against a dataset of roughly 14,000 English words rated by people from negative to positive.
Across the models, a name scoring one point more positively was associated with a 0.07-point improvement in the description. Because brand descriptions occupied a narrow range, that modest nudge mattered.
The association survived controls for sector, model and founding decade. But it was still observational. Perhaps companies with cheerful names also behave differently, attract different coverage or write happier marketing copy. Correlation had entered the casino wearing causal trousers, and Ehrlinspiel properly demanded identification.
The risk is concentrated among unknown brands. A new company leaves a vacuum, and language models are magnificent vacuum-filling devices with no fear of embarrassment.
2. The laboratory held the facts still and changed only the name
The controlled experiment used 80 pairs of fictional names. Within each pair, the words matched on length, syllables, frequency, emotional intensity and perceived power, but one was positive and the other negative.
Each name was attached to four neutral descriptions and tested across Gemma 4 31B, Gemini 3.5 Flash, DeepSeek v4 Pro and GPT-5.6 Terra, producing 640 randomized samples per model. The facts stayed fixed while the sign changed.
The naming effect appeared across nearly every model, although Gemini’s estimate did not reach statistical significance. Terra showed the largest effect, about four times Gemma’s in this test. On Terra, semantic takeover occurred in one out of twelve descriptions, with the model describing the word instead of the company.
Four models do not establish a universal law, and the newest tested model being most susceptible does not prove future systems will follow. It does kill the comfortable assumption that better models automatically outgrow this kind of shortcut.
3. Brand knowledge overrules the dictionary
The hopeful part arrived with familiar companies. Discord, Riot and Slack all carry words with negative or awkward ordinary meanings, yet the assistants described the businesses normally. Once enough company knowledge existed, the literal word stopped driving the answer.
Recognition replaced linguistic guesswork. The practical fix for an existing company is not a panicked rebrand. It is creating consistent, verifiable information that teaches AI systems which entity the name represents.
Start with a plain company description that names the product, audience, category and useful difference. Put it in server-rendered text on the About page, important product pages and credible business profiles. Support it through customer reviews, partner pages, industry coverage and other independent sources rather than repeating one glossy paragraph across a dozen empty directories.
Then monitor the adjectives AI systems actually use. If a storage company named Merciless is described as ruthless rather than secure, add a clear correction to the relevant page and strengthen third-party evidence until the company overwhelms the vocabulary.
Names have always carried baggage. The new danger is that a synthetic clerk can unpack it at industrial speed and present the contents as research. Give the system enough evidence to recognize the business, or the dictionary will testify first and the jury may never call another witness.
