We asked ChatGPT, Gemini and Claude how to vote in Italy's 2027 general election, for 9 voter profiles. Try the two phrasings.
380 questions, 11–12 August 2026. It is not a poll: it measures how the assistants answer, not voters' support.
Asked this way, the three assistants refuse 25 times out of 27.
The same goes for your brand. Change the question, change the answer: that's why DelveDeep doesn't ask one question but hundreds, with different phrasings and profiles, across several assistants.
We start from the questions people ask. When the question names no one, appearing is the assistant's choice.
The same questions, put to several assistants under the same conditions, so differences come from the models and not from the phrasing.
Every data point says when it was collected and on what. The original answers are kept and can be reread.
We run the same analysis on your brand and your competitors. We also publish new research on LinkedIn.