--- title: "Pseudonymize brands in AI prompts" description: "SurveyTester can replace your dictionary brands with stable fake names before a prompt reaches a public AI vendor, and put the real names back in the answer." url: https://www.surveytester.com/academy/en/company-admin-settings/ai-brand-pseudonymization/ language: en type: article audiences: ["SurveyTester Administrators", "Team Management"] published: 2026-10-31 updated: 2026-10-31 tags: ["AI", "privacy", "confidentiality", "brands", "pseudonymization", "anonymization", "company dictionary", "AI engines", "OpenAI", "Mistral"] --- # Pseudonymize brands in AI prompts _Company Admin Settings · SurveyTester Academy_ Surveys under test are often confidential: they name the client's brands, the competitors and the products that are about to launch. When an AI feature sends a page or a questionnaire to a public AI vendor such as OpenAI or Mistral, those names travel with it. **With "Pseudonymize brands" switched on for an AI engine, SurveyTester replaces every brand from your Company Dictionary with a stable fake name before the prompt leaves, and puts the real name back into the answer.** The vendor sees "Velorin" instead of "Nivea"; you see "Nivea" in the result. ## Why brands are the sensitive part Most survey content is generic: "How often do you buy shampoo?" is no secret. What makes a survey confidential is *who* it is about. A list of brands tells anyone who reads it which client ordered the study, which competitors they watch and, in a concept test, which product has not been announced yet. Removing the brands removes most of what an outsider could learn from a prompt, while the AI still understands the task: a fake brand name is still a brand name, so testcase generation, survey answering and questionnaire analysis work as before. ## Why this is hard without SurveyTester - **Prompts are built automatically.** Nobody types them, so nobody can leave the brand names out by hand. - **One survey, many calls.** A single test run can send dozens of pages to the AI. Each one would need the same replacements, consistently. - **The answer has to fit the survey again.** When the AI picks "Velorin" as the answer, the bot must click "Nivea" on the page. A one-way filter would break the features it is meant to protect. ## How it works 1. **The brands come from your Company Dictionary.** Every entry of the type **Brand** gets a fake name the first time it is needed. Entries of the type *Corrected word* are not replaced. 2. **The fake name is stored and never changes.** A brand keeps its pseudonym across all projects and over time, so the AI sees a consistent picture, and repeated prompts stay comparable. The same word always gets the same fake name, and two brands never share one. 3. **The fake names look like brands.** They are pronounceable and follow the brand's spelling pattern: "Coca-Cola" becomes a capitalised word, an acronym like "BMW" stays three capital letters. 4. **The answer gets the real name back,** in the spelling the survey used. If the page writes "Coca-cola", the answer says "Coca-cola", so answers still match the categories on the page. The replacement follows a few rules so it does not change ordinary words: - Only whole words are replaced: "Axe" is replaced, "taxes" is not. - Longer entries win: "Coca-Cola Zero" is replaced as one brand, not as "Coca-Cola" plus " Zero". - A brand that is also an everyday word is only replaced when it is capitalised: "Dove" is replaced, "a dove" is not. ## Switch it on for an AI engine The setting belongs to the AI engine, because whether a vendor may see your brands depends on the vendor. A self-hosted engine in your own data centre does not need it; a public cloud vendor usually should have it. 1. Navigate to **Administration → AI Engines**. 2. Click **Edit** next to the engine, or **Add Engine** for a new one. 3. Tick **Pseudonymize brands**. 4. Save. ![Placeholder: screenshot of the Edit AI Engine dialog with the model, the reasoning effort and the ticked checkbox "Pseudonymize brands" with its help text](https://www.surveytester.com/academy/_img/company-admin-settings/ai-brand-pseudonymization/en/ai-engine-pseudonymize-brands.png "w=40rem | The setting in the engine dialog") Engines from our catalog come with a recommended setting: public cloud vendors on, self-hosted engines off. As long as you do not change the checkbox, the engine follows that recommendation. Custom engines start with the setting off. Engines that pseudonymize show the badge **Brands pseudonymized** in the engine list. ![Placeholder: screenshot of the AI Engines list with two engines; the OpenAI engine shows the green badge "Brands pseudonymized" next to its model](https://www.surveytester.com/academy/_img/company-admin-settings/ai-brand-pseudonymization/en/ai-engines-pseudonym-badge.png "w=44rem | The badge in the engine list") ## See the fake names The **Company Dictionary** shows the fake name of every brand in the column **AI Pseudonym**. It is assigned the first time an AI prompt or the dictionary list needs it. You do not set it yourself. ![Placeholder: screenshot of the Company Dictionary list with the columns Word, Language, Category, Type and AI Pseudonym; brand rows show fake names like "Velorin" in grey italics, corrected words leave the column empty](https://www.surveytester.com/academy/_img/company-admin-settings/ai-brand-pseudonymization/en/company-dictionary-pseudonym-column.png "w=44rem | The AI Pseudonym column") > [!TIP] > Add the brands of a new study to the Company Dictionary before the first AI feature runs on it: the client's > brands, the competitors and the product names of a concept test. Brands that are not in the dictionary are sent > unchanged. ## Which AI features are covered All features that send text to an AI engine: testcase generation (including screenout testcases), answering surveys with AI, questionnaire document analysis, AI text analysis and the AI support of the page parser. ## What it does not cover Pseudonymization is a strong layer of protection, not a guarantee: - **Brands missing from the dictionary** are sent as they are. Client names and project code names count too: add them as brands if they should be hidden. - **Misspelled forms** that differ from the dictionary entry ("Coca Cola" without the hyphen) are not recognised. - **Images are sent unchanged.** A prompt that includes an image, for example a stimulus with a logo, is not pseudonymized at all, because the logo would contradict the fake name in the text. - **Questions about real brand knowledge** become impossible: an AI that only sees "Velorin" cannot know what the brand stands for. ## Notes - SurveyTester's own AI request log keeps the real text, so support can still read what was asked. - Changing the checkbox affects the next AI request. Results that were generated before stay as they are.