92% Say Their Research Works. 85% Have Caught It Being Wrong.

Ninety-two percent of client-side research professionals say research in their organization is effective. Eighty-five percent say they've personally discovered it was wrong, misleading, or didn't reflect reality.

Both numbers come from the same people answering the same questionnaire, in ImpactSense's Research Effectiveness Review 2026, published at the end of October. The report calls that contradiction "the condition the industry operates within," and I think it's one of the more useful findings I've read this year, because it describes something founders run into all the time without a name for it. You can believe research works in general and still have no idea whether the specific finding in front of you is one of the wrong ones.

If you're about to make a product or market bet on the strength of a research finding, that's the question that matters.

Most research never gets checked

The contradiction makes more sense once you see how rarely research is tested against reality. A separate study, Voxpopme and MRII's State of Insights, surveyed 126 research and insights practitioners, mostly at director level or above. As Research Live reported, only 17% said they consistently measure the return on their research. Twenty-eight percent had thrown out or re-run a study over data-quality concerns in the past year.

The say-do gap got the least attention of all. Twenty-one percent had seen a finding about what people said they'd do contradicted by what they actually did. Sixteen percent had no formal process for checking that gap, and another 33% didn't know whether they had one.

That last group is the one I'd worry about if I were the founder relying on the research. Stated intent is a large share of what product research measures – would you use this, would you pay for this, which of these would you choose – and it's a weak predictor of behavior. A team that never checks it is building on the part of the evidence most likely to be wrong.

Some of the respondents may not be real

There's a newer problem underneath the old ones. In the Voxpopme study, data quality was named the biggest threat to insights work, and fraudulent respondents and AI-generated answers made up 60% of the top concerns people named. That's the 28% who had to discard or re-run a study, seen from the other side.

For a founder, this changes what "we surveyed 400 people" means. Online panels have always had professional survey-takers and people speeding through for the incentive. Now some of the open-ended answers that sound most articulate may have been written by a model, and they'll read as more thoughtful than the real ones. A finding built on that kind of sample can be internally consistent, statistically significant, and about people who don't exist.

The practical defense isn't exotic. Ask how respondents were recruited and screened, how many were removed and why, and whether anyone talked to a handful of them directly. Research that holds up usually has a few conversations with real people underneath the numbers, and someone who can tell you what those people sounded like.

Why confident research is a cheap yes

None of this means research is useless. The ImpactSense respondents aren't wrong that research usually helps, and the people in these surveys are the professionals who know its limits best. The problem is narrower. A finding that arrives with confidence, in a clean deck, from people who believe in their methods, is easy to accept and hard to question, and accepting it costs nothing at the time.

A cheap yes leads to an expensive no. "The research says customers want this" ends the debate in a planning meeting. If the research was wrong, the expense shows up months later, after the build, when the customers who said they'd use it don't. And because the finding was never checked, nobody can tell whether the research was wrong or the execution was, which means the team doesn't learn much from the miss either.

AI makes research faster, not more trustworthy

You might expect AI to help here. So far, mostly it hasn't. ImpactSense found that 93% of organizations now use AI somewhere in their research process, and only 39% say it has significantly improved the quality of their insights. Maze's State of Market Research found 84% of researchers agree AI increases their capacity, and capacity is exactly the point. AI lets a team produce more findings, faster, with the same confident formatting, and it doesn't do anything on its own to tell you which of them are wrong.

The insights leaders seem to know this. In the Voxpopme study, 51% said they wouldn't trust AI to run any research task without a human reviewing it. That instinct is right, and it's worth applying to the human-produced findings too.

Four checks to run before you build on a finding

When a research finding is about to drive a decision that's expensive to reverse, I'd want answers to four questions first. None of them require you to be a researcher.

Can you see the work? Who was asked, how they were found, what exactly they were asked, and what they said in their own words. A finding you can't trace back to its evidence is an opinion with good formatting. When I run research, every claim gets a fact-check pass before it goes anywhere near a decision, and the evidence travels with the finding so anyone can check it. The system should show the work.

What would have proved it wrong? If nobody can name a result that would have changed the conclusion, the research wasn't testing anything. It was confirming.

Does it rest on what people said or what they did? Stated intent is useful for understanding how people think about a problem. It's weak evidence for what they'll do. If the finding is "customers would pay for this," ask whether anyone has paid for anything like it, and what they did the last time they had the problem.

Who disagrees, and why? A finding that everyone agrees with may be right, or may be one nobody looked at hard. Find the person closest to the customer who thinks it's wrong and hear them out. If the finding survives that conversation, it's worth a lot more than it was before.

The finding you can defend

The ImpactSense contradiction isn't a scandal. It's the normal state of a field where much of the work never gets checked against what happened next. The fix for a founder isn't to distrust research. It's to treat a finding as a claim that has to earn its way into a decision, in proportion to how expensive that decision is to undo.

For small bets, a reasonable finding and a fast experiment are fine. For the big ones – a new market, a new product line, a year of a team's time – the bar should be a finding you could defend to a skeptic, with the evidence in hand. That's slower than taking the deck at its word. It's much faster than discovering, a year from now, that you built on one of the wrong ones.

Related: if your best customer and your worst prospect give the same answer, the question tested nothing, and what to do when your research tool tells you what you want to hear.