Evidence should challenge your assumptions, not confirm them

Most product teams don't skip research. They skip the version of research that could actually change their mind.

It's a subtle failure mode. You have a direction. You talk to a few customers, run a survey, look at some usage data, and somehow the evidence always seems to support the plan you were already leaning toward. That's not because the plan was right. It's because the questions were shaped by the answer you wanted.

The tell

If you can predict what your research is going to say before you do it, it isn't research. It's confirmation. Real discovery should have some chance of sending you back to the drawing board. If it never does, the process is decorative.

What I try to do instead

Before starting discovery on something I already have a hypothesis about, I try to write down the version of the evidence that would make me abandon the idea. Not a strawman, but the strongest, most plausible way I could be wrong. Then I go looking for that evidence specifically, not just evidence that agrees with me.

This doesn't mean every hypothesis needs to survive contact with reality intact. Most don't, in some way. The point isn't to be right on the first try. It's to find out you're wrong while it's still cheap to change course.

A recent example

During discovery for an analytics product, early conversations produced a list of requested reports and features. I changed the questions. Instead of asking only what users wanted added, I asked what decisions they were trying to make, how they made those decisions today, where the process slowed down, and what happened when the information could not be trusted. That shift revealed the real value of the product: reducing manual reporting work and helping teams reach decisions with greater confidence.

Where this breaks down

AI-assisted research makes this both easier and more dangerous. It's now trivial to generate a synthesis that sounds authoritative and happens to agree with your priors, because the model is pattern-matching on what a plausible-sounding answer looks like, not independently verifying it against reality. The discipline of designing discovery that can prove you wrong matters more, not less, when synthesis is cheap and confirmation is one prompt away.

Evidence that only ever agrees with you isn't evidence. It's decoration on a decision you'd already made.