What people say. What people do. The gap that eats launches.

In short: The say-do gap is the difference between what customers say they'll do and what they actually do. Stated intent predicts real purchase behavior only about 20 to 30% of the time, which is why most research points teams confidently in the wrong direction. Survey-based tools (Qualtrics, Attest, Zappi and most "AI research" platforms) inherit the gap by design; behavioral methods are built to close it. Synthetic personas are the fastest way to get an answer, but you can only trust one if you can measure how correlated it is to real behavior. Heatseeker runs synthetic personas and live in-market experiments in one system, so the answer you get in minutes is grounded in what buyers actually do, around 95% correlation, not 20%.

Ask a room full of customers whether they'd buy the sustainable option. Almost every hand goes up. Then watch the checkout data: they buy the cheaper one. Ask if a feature is a must-have (must-have!), then check the logs: nobody touches it. Rate this concept out of ten: a nine. Real-world response to the ad: a scroll.

This is the say-do gap, and it is the single most expensive thing in marketing research. Not because people lie. Because we keep asking them to predict their own behavior in a comfy, hypothetical, no-stakes moment, and humans are spectacularly bad at that.

So why does everyone keep asking?

Because asking is easy, and it feels like data. You get numbers, charts, a percentage with a decimal point. It looks like certainty.

The trouble is what those numbers are made of. Stated preference (what people say when you ask) lines up with revealed preference (what they do when it's real) only about one time in four or five. So you can run a flawless survey, beautifully sampled, and still be wrong four times out of five. Fast and fluent doesn't save you. It just gets you to the wrong call sooner, with a nicer deck.

Most "insights" tools are just a survey wearing a lab coat

Here's the part the category doesn't advertise. Strip the AI language off most research tools and underneath is a survey. They recruit a panel, ask it questions, and, in a move of real comedic courage, validate the answers against more questions.

The tell is the method, not the marketing. So it's worth being blunt about who measures saying and who measures doing:

None of these are bad tools. A survey is a fine way to learn what people think. It's just a poor way to predict what they'll do, and predicting what they'll do is the whole job when a launch budget is on the line.

Synthetic personas are faster, and that's exactly the trap

The category is racing toward synthetic: ask an AI persona instead of a panel, get an answer in minutes instead of weeks. The speed is real and it's not going away.

But speed creates a trust problem the moment the stakes are real. A synthetic model will answer any question with total confidence, whether or not it's right. And a confident guess is more dangerous than a shrug, because you'll actually act on it. No CMO can carry a black-box answer into a boardroom and bet a launch on it. Without trust, there is no synthetic research in the enterprise.

The fix isn't to avoid synthetics. It's to measure them. The only way to trust a synthetic is to know how correlated it is to real behavior: calibrate it against a live in-market experiment, and report how closely it tracks what buyers actually did, with a confidence label attached instead of a shrug dressed as certainty. A synthetic you can validate is an instrument. One you can't is a horoscope.

That calibration question, how do you know your AI is right, is its own subject. We wrote it up here: how do you know your AI is right?

So, how do you actually close the gap?

Start from the decision you're trying to de-risk, and pick the method that sits closest to real behavior with real stakes:

• Deciding what to launch, which message leads, whether demand is real, before you spend. You want behavior plus validation: a live experiment, and a synthetic you can trust because it's calibrated against one.
• Need category purchase data at scale? That's NielsenIQ's world, just remember it reads what already sold, not what will.
• Working a qualitative, in-context question? dscout and in-field ethnography watch behavior directly, at small scale.
• Need a fast read on what people say? A survey is fine, just don't mistake stated intent for what they'll do.

The one rule under all of it: the closer your method sits to real behavior with real stakes, the smaller your say-do gap, and the more your research predicts what customers actually do.

This isn't theory. A global delivery platform used exactly this approach, grounding predictions in what real buyers did, not what a model assumed, to validate a new go-to-market in 14 days and cut media waste along the way. Not by trusting a model. By checking it against reality.

That's what Heatseeker is for: ask anything, get an answer in minutes, grounded in live experiments with your real buyers, behavior, not opinion.