The 94% figure comes from John Gottman's own published research and gets repeated constantly, almost always without the methodology attached to it. The honest version is more interesting than the headline: Gottman's team did find that specific, observable behaviors — the "Four Horsemen" — reliably distinguish stable couples from ones heading toward separation, but the famous accuracy number describes a model built and tested under conditions that don't quite match how it usually gets described online.
Where the Number Comes From
Starting in the 1980s and 90s, Gottman and longtime collaborator Robert Levenson brought couples into a lab, wired them to measure heart rate and other physiological signals, and had them discuss a real point of conflict on camera. Trained coders then classified specific behaviors — criticism, contempt, defensiveness, withdrawal — frame by frame.
Years later, the researchers followed up with the same couples to see who had split up and who hadn't. Using the behavioral and physiological data collected earlier, they built statistical models that classified couples into "stable" or "heading toward divorce," and reported very high classification accuracy against the actual outcomes — this is where the widely quoted figure originates.
That's a genuinely striking finding. It's also a more specific and narrower claim than "this test predicts your divorce with 94% accuracy," which is roughly how it circulates in popular retellings.
What the Studies Actually Measured
The distinction that matters most is between predicting forward and classifying backward. In the earliest and most-cited version of this work, researchers already knew which couples had divorced by the time they built the classification model — the behavioral codes were being used to explain an outcome that had already happened, then tested for how well they sorted couples into the categories that were already known.
That's a legitimate and useful research method. It is a different exercise than a fortune-teller watching a fifteen-minute conversation between strangers and correctly guessing, in advance, who will still be married in a decade — which is closer to the popular imagination of what the finding means.
Gottman's group and others have since run more genuinely prospective versions of this work, following newly observed couples forward rather than only looking backward at known outcomes. Those studies generally support the core finding — specific conflict behaviors do predict relationship trajectory — but report more modest and more variable accuracy than the number that made it into headlines.
The Four Horsemen, Defined
Whatever the exact accuracy figure, the behavioral categories themselves are the part of this research that has held up best and proven most useful in practice.
| Horseman | What it looks like | Why it's corrosive |
|---|---|---|
| Criticism | Attacking a partner's character rather than addressing a specific behavior — "you always" and "you never" statements | Turns a solvable problem into a referendum on someone's worth |
| Contempt | Sarcasm, eye-rolling, mockery, or a tone of moral superiority | Communicates disgust rather than disagreement — the single horseman most strongly linked to relationship decline |
| Defensiveness | Responding to a complaint with a counter-complaint or self-justification instead of acknowledgment | Blocks the conversation from moving toward resolution, since nothing gets accepted |
| Stonewalling | Shutting down and disengaging, often after feeling physiologically overwhelmed | Removes the only mechanism a couple has for actually resolving anything — read more in the full guide to stonewalling |
These four rarely arrive alone. They tend to follow a rough sequence — criticism invites defensiveness, which under repeated conflict can curdle into contempt, which eventually produces stonewalling as the only remaining exit. Recognizing the sequence, not just the individual behaviors, is often more useful than treating them as four separate checkboxes.
Why the Number Is Contested
Several legitimate critiques have followed this research since it became widely known, and they're worth taking seriously rather than dismissing as noise around a solid finding.
- Same-sample fitting. Building a classification model on data where the outcome is already known, then reporting how accurately that model sorts the same data, tends to produce better-looking numbers than the model would achieve on genuinely new couples. This is a well-known statistical trap, and some methodologists have argued the original studies didn't fully escape it.
- Independent replication has been limited. Comparable accuracy figures from labs outside Gottman's own research group are harder to find than the popularity of the number would suggest, which is a reasonable basis for caution about how far the finding generalizes.
- Small, non-representative samples. Much of the foundational work was done with couples recruited from a specific geographic area willing to be filmed in a lab, which is not automatically representative of couples broadly.
- The popularization outran the caveats. Books, articles, and talks summarizing this work have repeatedly dropped the methodological qualifiers that were present in the original academic papers, which is how "a striking classification result under specific research conditions" became "this test predicts your divorce."
None of this means the underlying finding is wrong. The Four Horsemen framework has been useful enough that it's now a standard part of couples therapy training well beyond Gottman's own institute. What's contested specifically is the precision of the headline number, not the existence of the pattern it's attached to.
What This Means If You Recognize the Patterns
The honest, useful version of this research is closer to a risk indicator than a prophecy. Recognizing one or more of the Four Horsemen in your own relationship is worth paying attention to — it's a pattern with a documented association with relationship strain, not a countdown timer.
It's also worth being specific about which horseman is present, because they're not equally serious. Occasional defensiveness under stress is close to universal and much less concerning on its own than a consistent pattern of contempt, which shows up across the research as the most damaging of the four by a meaningful margin.
The most useful question isn't "does this number apply to me." It's "which of these four shows up in my relationship, how often, and what happens right after it does" — because what happens after, whether the moment gets repaired or left to fester, tends to matter more than whether it happened at all.
What the Research Doesn't Claim
It's worth being as clear about the limits of this work as about its findings, because both get lost in the popularized version.
The research doesn't claim that any of the Four Horsemen appearing once, or even occasionally, dooms a relationship. Every couple, including ones that go on to have long, stable marriages, shows some defensiveness or criticism under enough stress. The research is about pattern and frequency, not about the presence of a single instance.
It also doesn't claim these behaviors are unchangeable traits belonging to one type of person. The same person can show high contempt in one relationship and none in another, which suggests the behavior is at least partly relational — shaped by the specific dynamic between two people — rather than purely a fixed individual characteristic.
And it doesn't claim to account for everything that determines whether a relationship lasts. Financial stress, external life events, and compatibility on values and goals all matter too, and none of them show up in a fifteen-minute conflict conversation, however well-coded that conversation is.
How This Fits Alongside Other Relationship Research
Gottman's work is one of the most visible strands of relationship science, but it isn't the only one, and the broader field offers useful context for how much weight to put on any single framework.
Attachment research, developed separately and following a different methodology entirely, converges on some of the same territory — particularly around withdrawal and its corrosive effect on a relationship, which lines up closely with what stonewalling describes from a different angle. Personality research, including the Big Five model, has found that certain broad traits, especially emotional stability, correlate with relationship satisfaction and longevity across large, long-running studies unconnected to Gottman's lab.
None of these frameworks fully explains relationship outcomes on its own, and none should be treated as the single deciding factor. What they consistently agree on, across very different research methods, is that how a couple handles conflict and uncertainty matters more than whether conflict happens at all — which is the part of Gottman's work worth taking seriously, independent of exactly how the accuracy figure was calculated.
The Antidote for Each Horseman
- Replace criticism with a specific complaint. "I felt overlooked when the plans changed without me knowing" addresses the actual event; "you never think about me" attacks character and invites defense rather than change.
- Replace contempt with a genuine attempt at appreciation. Actively building a habit of noticing what a partner does well counteracts the drift toward disgust that contempt runs on — it doesn't erase real grievances, but it changes the emotional climate they're raised in.
- Replace defensiveness with partial acknowledgment. Accepting even one piece of a complaint as valid — "you're right that I was late, that part's on me" — before adding your own perspective changes the whole shape of the conversation.
- Replace stonewalling with a stated pause. "I'm overwhelmed, I need twenty minutes, I'll come back" keeps the mechanics of self-regulation while removing the silent, unexplained exit that does most of the damage.
None of these are complicated in concept. What makes them hard is applying them in the exact moment the nervous system most wants to do the opposite — which is why building the habit outside of conflict, the way repair attempts get built, tends to matter more than understanding the theory.
Where to Start
Patterns like these are easier to spot in a partner than in yourself, which is exactly why an outside measure helps. The attachment style test maps the underlying relational patterns that often drive these behaviors in the first place — many instances of contempt or stonewalling trace back to an attachment strategy rather than appearing at random. The Big Five personality assessment covers the broader trait, emotional stability, that Gottman's own later work has linked to how conflict tends to go, independent of any specific technique.
If stonewalling specifically is the pattern you recognize, the dedicated guide to it goes further into what's happening and what actually helps.
Whatever the exact number attached to it, the practical takeaway from decades of this research holds up without needing the headline figure to be exactly right: how a couple fights predicts far more than whether they fight, and that's a pattern within anyone's ability to change.