Not every project needs a research phase, and pretending otherwise has done the discipline real damage. Here is when it pays and when it does not.
Buy it when the team disagrees and cannot resolve it
Two credible internal positions, both plausible, no evidence either way. This is the single best case for research, because the alternative is that the more senior person wins and everyone else disengages.
Six to twelve conversations with actual users will usually settle it, and — more useful — will often reveal that both sides were arguing about something the users do not care about.
Buy it when the numbers show a failure you cannot explain
Analytics tell you that forty per cent of people abandon at step three. They will never tell you why. That is a qualitative question and no amount of further instrumentation answers it.
Watching six people attempt step three answers it in an afternoon, and the answer is frequently something nobody in the building had considered.
Buy it when the users are genuinely unlike you
Consumer products built by people who resemble their customers can get a long way on intuition. Products for radiographers, hauliers, insurance assessors or local-authority caseworkers cannot, and teams that try produce software with a confident, wrong model of the job.
If nobody on the team has done the work your software supports, research is not a phase. It is the only way you will find out what the software is for.
Do not buy it to confirm a decision already made
If the design is agreed, the date is set and the budget is committed, research will produce findings nobody can act on. This is the most common way it gets wasted, and it is worse than skipping it, because it teaches the team that research does not change anything.
Do not buy it for questions a smaller thing answers
If the question is which of two headlines performs better, run the test. If it is whether people can find the pricing page, look at the analytics. Research is expensive and slow relative to both, and using it for questions with cheaper instruments makes it look like a luxury.
What good looks like
Six to twelve moderated conversations. A synthesis session with the whole team in the room, not a report delivered to them. And an output that reads as a short ranked list of decisions rather than a deck.
Anyone who quotes you forty interviews for a marketing site is selling weight, not evidence.