Why you get a different answer when you ask an assistant yourself

The short answer

Almost certainly not word for word, and that is expected rather than a fault. These models are non-deterministic: ask the same question twice and the wording, the order of brands, even which competitors appear can change. Replies also shift with the account asking, the country it is asked from, and the day. That is precisely why a check asks 12 buyer questions across eleven assistants rather than spot-checking one: a single answer is noise, the pattern across roughly a hundred is the signal. Every answer we report is stored verbatim with the date it was given, so you can see exactly what was said and when, and each check is a measurement of that moment, not a permanent verdict. Two more things worth knowing: we ask each assistant through its provider's official API with web search switched on, not through the consumer app, and the two can differ in wording and occasionally in the brands named (DeepSeek is asked in its standard non-thinking mode, the one its app uses unless a user switches DeepThink on); and because every question is asked once per assistant, the check prints a margin of error beside its headline share, and a tracked brand's page shows how consistent each answer has been across its recent checks.

Reading results with that in mind

• Look at the share of answers and the range printed beside it, not a single cell.
• Look at consistency across checks: a tracked brand's page shows how often each question led and how often each assistant named you first, across recent checks.
• Every answer is stored word for word with its date, so you can always show exactly what was said and when.