A digital respondent is a model of one kind of person, asked a question. Not a chatbot playing a role. A model trained on what people like that actually said and did, so that when you ask it whether it would pay forty dollars, its answer comes from the same place a real answer would.
The useful property is not that any one answer is right. It is that a thousand of them spread the way a thousand people would. Ask a room whether they would switch banks for a point of interest and you get a shape: some yes, more maybe, a stubborn no. A good digital respondent panel gives you that shape. An assistant model gives you forty polite yeses.
What they are good for
- Ranking. Ten concepts, which three deserve real fieldwork.
- Direction. Does this price land closer to the ceiling or the floor.
- Segments. Does the message that works for the switcher also work for the loyalist.
- Speed. Asking at six in the evening and reading at eight in the morning.
What they are not
They are not a reason to stop interviewing people. The respondents are only as good as the human data behind them, and that data goes stale. Every real interview you run teaches them something new, which is why we run both, in the same study, and show you the interviews behind every number.
They are also not an oracle for the thing nobody has done before. A model of a population predicts well inside what that population has lived through. A genuinely new product is where you want the real conversation first, and the respondents second.
How to tell a good one
Ask it the same question forty times and look at the spread. Ask it something with a known answer from your last wave and check the number. Ask it why, and see whether the reasons sound like your customers or like a press release. We publish our own results on the open evaluations page, including the ones we lose.