Questions That Reveal Compatibility: What Can Be Predicted, and What Cannot

The famous list was built to manufacture closeness, not to detect it

Almost every compatibility questionnaire in circulation descends, at some remove, from one paper. It is worth reading what that paper actually did.

Arthur Aron, Edward Melinat, Elaine Aron, Robert Darrin Vallone and Renee Bator published it in 1997 in Personality and Social Psychology Bulletin, under a title that gives the game away: The Experimental Generation of Interpersonal Closeness. The problem they were solving was methodological, not romantic. Researchers could compare people who were already close against people who were not, but they could not assign closeness at random, which meant they could never separate cause from selection. So they built a procedure that produces it on demand: two strangers work through three sets of twelve slips over forty-five minutes, fifteen minutes per set, with the disclosure escalating within each set and across the three.

Study 1 tested whether the content of the tasks was doing the work. The experiment ran inside a large psychology class five weeks into term. Students were paired with someone they had not listed as an acquaintance — thirty-three cross-sex pairs and seventeen all-women pairs — and half of them got the closeness slips while half got small-talk slips matched for length and structure. On a composite of two standard closeness measures, the closeness condition came out at 4.06 against 3.25 for small talk, a difference of about 0.88 standard deviations, F(1, 37) = 5.68, p < .05.

That is a real result and a large one. Notice what it is a result about. The outcome measured is how close two strangers feel at the end of forty-five minutes. Nothing in the design ranks the pairs against each other, scores their fit, or follows them to see who lasted. To calibrate it, the authors compared their scores against a separate sample of 296 students at the same university who rated their closest, deepest, most intimate relationship: that group averaged 4.65, and the forty-five-minute strangers landed around the thirtieth percentile of that distribution. Impressive for three quarters of an hour. Not a measurement of anything about the two people.

The word love does not appear in the findings. The only marriage in the paper is a parenthetical aside about an earlier pilot. The instrument creates a feeling, and the dating advice shelf that borrowed it — this site's own older version of this page included — quietly repurposed it into a test.

In the same paper, matching people on their opinions did nothing

The same article contains the result that should have ended the compatibility-quiz genre, and almost nobody quotes it.

Study 2 ran the identical forty-five-minute procedure on seventy-one new pairs — fifty-two cross-sex, nineteen all-women — and manipulated the two things the folklore treats as decisive. Half the pairs were matched so that they did not disagree on anything either of them had rated as important, using seventeen attitude items chosen precisely because a clash on them would make someone undesirable as a partner. The other half were deliberately paired into real disagreement: one strong clash on a highly important issue, or two on moderately important issues, or three moderate clashes on issues rated very important. Crossed with that, half were told the researchers had matched them carefully and expected them to like each other, and half were told there was no reason to assume anything of the kind.

Neither manipulation moved anything. Not a small effect, not a marginal one: on the closeness composite and on a two-item liking measure, every F was below 1, for both variables and for their interaction. The study had roughly 90% power to detect a large effect and about 90% power to detect even a marginal medium-sized one, which makes this an informative null rather than an empty one. Two strangers deliberately paired on a live disagreement about something they both cared about finished as close as two strangers who agreed on everything that mattered to them.

The authors say they were surprised, and offer candidates: perhaps the disclosure process is powerful enough to swamp agreement; perhaps the slips never steered anyone onto the topics they disagreed about; perhaps everyone assumed similarity because their partner came from the same lecture hall. Those are live possibilities, and one design is not a verdict. But the single most famous exercise in this field is also the only experiment that tried matching people on their stated positions, and it measured nothing.

Similarity predicts attraction right up until people know each other

The obvious objection is that the similarity literature is enormous and mostly points the other way. It is enormous. It also splits in a way that matters here.

R. Matthew Montoya, Robert Horton and Jeffrey Kirchner pooled 460 effect sizes from 313 laboratory and field investigations to separate two things that usually get merged: actual similarity, meaning the measured overlap between two people's traits and attitudes, and perceived similarity, meaning how similar one person believes the other to be. Across the whole pile both were substantial — actual similarity correlated with attraction at r = .47, perceived similarity at r = .39.

The split is where it gets useful. Actual similarity mattered in studies where participants never interacted and in studies with brief interactions, and its effect dropped significantly once you moved beyond no-interaction designs. In existing relationships, the effect of actual similarity was not significant. Perceived similarity predicted attraction at every stage: no interaction, short interaction and established relationship alike.

Read that as a sentence about questions. Scoring how much two people's answers overlap predicts how much they will like a description of each other. It stops predicting once there is an actual relationship to measure. What keeps predicting is the sense of being understood, which is a different quantity and is not obtained by tallying agreement. This is one of the reasons capable people keep making the same choice badly, a pattern covered in why smart people keep dating the wrong people: the checklist is being run on the wrong variable.

The same wedge — between the person you have information about and the person you have filled in — opens wider the scarcer the contact is, which is why separated couples rate their partners more positively than couples who see each other every day. That is covered in what distance does to attraction.

Your list of ideal qualities stops predicting the moment you meet

The other half of the folklore says compatibility means finding someone who matches what you are looking for. That has been tested directly, and there is a clean point at which it stops working.

Paul Eastwick, Laura Luchies, Eli Finkel and Lucy Hunt reviewed the field in Psychological Bulletin in 2014 and ran a new meta-analysis across 97 samples. Physical attractiveness predicted romantic evaluations at about r = .40 and earning prospects at about r = .10 — for both sexes. The sex difference in those correlations averaged around .03 and was uniformly non-significant, even though men and women reliably differ when rating hypothetical partners on paper.

Their review reaches two conclusions about ideal partner preferences. They predict evaluations of hypothetical targets, and they do not predict evaluations of face-to-face partners in attraction settings. And they do predict evaluations of established partners, though only through one particular source of variance.

The cleanest demonstration is a laboratory study they describe. A single participant reads a profile of an opposite-sex person listing three traits that person supposedly picked out for themselves, and rates their romantic interest. Matching the participant's stated ideals moved that rating. The two then take turns describing pictures in a short, scripted live interaction. Afterwards, the ideal versus non-ideal manipulation had no effect at all, and the interaction between condition and stage was significant. The authors call the first face-to-face meeting a point of disconnect, after which stated ideals lose their predictive power.

So the questionnaire you fill in before meeting someone is genuinely predictive — of how you will evaluate a profile. What survives the first coffee is something else. It fits with what large surveys find about couples who met through an app: once they are together, they look much like couples who met any other way, which is one of the findings in the survey numbers on modern dating.

Nothing measured in advance predicted whether two people would click

In 2017 Samantha Joel, Paul Eastwick and Eli Finkel put the question to a machine, and the result is the most useful thing in this whole literature.

They used two speed-dating samples: 163 undergraduates across seven events in 2005, and 187 across eight events in 2007. One to two weeks before the dates, each participant completed more than a hundred self-report measures covering the traits and preferences that relationship science has spent decades identifying. Then each person had roughly twelve four-minute dates and rated romantic desire immediately after each one.

The first step was to split desire into three independent pieces: actor variance, how much a person tends to want others; partner variance, how much others tend to want them; and relationship variance, how much this specific person wanted that specific person over and above both. All three were substantial, and relationship variance — the piece everyone means by compatibility — was the largest of the three, so it was the easiest target, not the hardest.

Random forests predicted 4% to 18% of actor variance and 7% to 27% of partner variance. On relationship variance, using any combination of the hundred-plus advance measures, they predicted nothing: the cross-validated estimates came out slightly below zero. The same models, given measures collected after the four minutes, predicted 16% to 29% of it. Whatever determines whether two people click exists, it is large, and it was not present in anything either of them reported beforehand.

The authors are careful, and so should anyone quoting them be. These are undergraduates in a speed-dating room. Four minutes is not a relationship. They say explicitly that their findings address long-term compatibility only obliquely, and that relationship desire might consist of thousands of tiny actor-by-partner interactions that would need genomics-scale samples to detect. What the study rules out is narrower and still large: no self-report battery collected in advance predicted unique desire, in a design built to give it the best possible chance. It is the same limit that makes reading someone from the outside so unreliable, the theme running through why interest gets misread in both directions.

What does predict, once there is a relationship to ask about

Three years later the same lead author ran the largest test of the other end of the question, and this one has an answer.

Joel and 85 co-authors pooled 43 longitudinal datasets from 29 laboratories: 11,196 couples, 22,163 people at baseline, 2,413 measures, followed across an average of four waves over fourteen months. Random forests again, preregistered, with everything given the same chance to compete.

The strongest predictors of relationship quality were all relationship-specific and all reported by the person themselves: how committed they perceived their partner to be, how much appreciation they felt toward their partner, sexual satisfaction, how satisfied they perceived their partner to be, and conflict. Those variables explained up to 45% of satisfaction at baseline and up to 18% at follow-up. The strongest individual differences — life satisfaction, negative affect, depression, attachment avoidance, attachment anxiety — managed 21% and 12%.

Then the finding that reframes the genre. Adding individual differences and the partner's own ratings contributed nothing beyond what the person's own relationship-specific reports already carried. Random forests detect interactions by construction, and the authors note that a model containing both partners' data would have caught matching effects if any robust ones existed. None showed up. Whatever the two people are like separately appears to reach the relationship only through what each of them experiences inside it.

That is what a question can and cannot reach. Asking what someone is like produces information about the weakest predictor in the set. Asking what a relationship with them tends to look like — how appreciated their last partner felt, what commitment looked like in practice, how conflict actually went — aims at the strongest one. Those three, appreciation, commitment in practice and how conflict actually goes, are also the ones with the most observational work behind them, collected in what 137 studies of breakups found predicts staying together. And since attachment anxiety and avoidance were among the individual differences that did carry weight, knowing your own is worth two minutes; our attachment style test gives a rough read and says plainly what it cannot tell you, and the difference between love and anxious attachment covers where the two come apart. That same list is what the relationship satisfaction test is organised around: it reports satisfaction and commitment as separate scores because in this analysis the strongest single predictor was not how satisfied a person was but how committed they believed their partner to be. More of this under Mindset.

What the Gottman work supports, and where its famous numbers stop

No article on this topic can avoid John Gottman, and most of them cite him badly, so here is the careful version.

The underlying research is real observational work. Gottman and Robert Levenson followed married couples with behavioural coding and physiological measurement and reported processes that preceded dissolution. Gottman, James Coan, Sybil Carrere and Catherine Swanson studied 130 newlywed couples and reported predicting divorce and stability with 83% accuracy and satisfaction with 80%. Those numbers are the ones that made it into the culture.

Richard Heyman and Amy Smith Slep took them apart in the same journal three years later. Their point is statistical rather than personal: a prediction equation fitted to one sample and evaluated on that same sample is not a prediction, and once you cross-validate and account for how common divorce actually is, accuracy and predictive value drop sharply. Their conclusion is that results without cross-validation analyses should be interpreted with extreme caution no matter how impressive they look.

Then somebody tried it on a fresh sample. Hyoun Kim, Deborah Capaldi and Lynn Crosby applied the models to 85 married or cohabiting couples, 54 of whom were still together about two and a half years later. Two of twenty-two affective processes predicted separation. The wife's negative start-up, one of the signature findings, did not. Their summary sentence is that the major findings of the Newlywed Study were not replicated.

One more figure deserves the same treatment, because it is quoted in nearly every article on this subject: that 69% of a couple's conflict is about perpetual problems that never get solved. We could not locate a peer-reviewed paper reporting that percentage with its method. It appears in Gottman's trade books and on the Gottman Institute's own site, and it circulates online in mutated forms with other numbers attached. Treat it as a clinical observation from a large practice rather than as a measured quantity.

What survives is the distinction itself, and it is genuinely useful: some disagreements are problems to be solved and some are permanent features of two particular people, and the second kind is managed rather than resolved. That frame is what makes the difference between a fight that repairs and one that does not, which is the subject of how to reconnect after a fight.

Questions about how conflict actually goes

Everything above narrows the useful question set considerably. Conflict was one of the five strongest relationship-specific predictors in the 11,196-couple analysis, and it is one of the few things a person can describe concretely from their own history. Three questions, and what each one can and cannot give you.

What happens after an argument — who reaches out first, and how long does it take? The value here is that repair is a behaviour with a timestamp, so the answer is checkable in a way that a stated value is not. Somebody who says they need a day and then takes a day is telling you something. Somebody who says they always talk things through and then describes three-day silences is telling you something else, and the gap between the two answers is the information. It is a fair question to answer about yourself first; the conflict style test takes a few minutes and separates pressing for resolution from pulling away.

What has a partner asked you to change, and what did you do about it? This is the closest a question can get to what the Gottman work calls accepting influence, and it is deliberately awkward. The useful part is not whether they changed. It is whether the request is remembered at all, and whether it is described as a reasonable thing someone wanted or as an unreasonable thing someone demanded.

How would your last partner describe the end of it? Asking for the other person's version rather than their own is the whole trick. Everyone has a fluent account of why their last relationship ended. Far fewer have thought about how it looked from the other chair, and the ones who have tend to describe something with two people in it.

None of these detects a good partner. What they do is put a person's behaviour under stress on the table early, instead of leaving you to discover it in month eight.

Questions about the ordinary week, and one caution about them

The second cluster is about the texture of normal time, and it is the part of the folklore that holds up best — with a caveat that the folklore usually gets backwards.

What does a normal Tuesday look like? Not a Sunday, and definitely not a holiday. The unremarkable weekday is the unit that repeats, it is hard to perform, and the answer tells you what you would actually be sharing. This is a description, not a test: there is no correct Tuesday, and two people with very different ones can build something that holds both.

What do you do when you are bored? Here is the caveat. The intuitive reading is that a compatible partner is one who is content with the same quiet routine you are, but the self-expansion research points somewhere else. Arthur Aron, Christina Norman, Elaine Aron, Colin McKenna and Richard Heyman found across a questionnaire study, a door-to-door survey and three laboratory experiments that couples who did a novel and arousing activity together reported greater relationship quality afterwards than couples who did a mundane one, with a no-activity control confirming the novel task was doing the work. Their survey data linked shared exciting activities to satisfaction with relationship boredom as the mediator. Note the size of what was tested: the laboratory effect comes from a seven-minute task measured immediately afterwards, not a transformation tracked over years. What it suggests is that shared novelty is a thing couples do, not a trait they match on.

What is something you have been appreciated for, and what do you wish somebody noticed? Felt appreciation was among the top five predictors in the 11,196-couple study, and unlike most of the others it is something you can offer rather than screen for. The answer is only worth asking for if you intend to use it.

One thing worth watching across all three answers: whether a real week has any room in it for another person. That is a structural question, not a values question, and it is often the one that quietly decides things — the failure mode described in the research on unequal commitment.

The things that do not bend

There is a category of question the research above does not cover, and it is the category where the folklore is right for once.

Children, or none. Where you are willing to live. Money, meaning both what someone earns and what they do with it. Care of ageing parents. Faith, where it is practised rather than merely held. These are not preference-matching exercises and they are not personality items. They are facts about the shape of a life, and two incompatible answers stay incompatible no matter how good the conversation is.

Ask them plainly and early, and ask for the fact rather than the value. Do you want children is a values question, and people answer it aspirationally. Where do you expect to be living in five years, and what would have to happen for that to change, produces a description. The difference is the same one running through this whole article: descriptions of behaviour and circumstance carry information, stated positions carry much less.

The part people get wrong is what to do with a mismatch. A difference here is not automatically a verdict, and it is not automatically survivable either. This is where the perpetual-problem frame earns its keep: some differences are permanent, and the real question is not whether you can talk each other out of them but whether you can build something that holds both answers without one person quietly disappearing into the other's plan. That is a decision, not a diagnosis, and it is worth making consciously rather than by drift.

And the timing matters more than the tact. A factual incompatibility discovered in month two costs a conversation. The same one discovered in year four costs considerably more.

When a difference is not incompatibility

One warning, because the compatibility frame is actively dangerous when applied to the wrong thing.

Being told who you may see or message. Having your phone checked, or being required to share your location when it is not mutual and not freely chosen. Being made to account for your time. Money controlled or withheld. Escalation — anger, punishment, silence used as a sanction — when you say no to something. Being isolated from friends or family a little at a time.

Those are not differences in style, and they are not compatibility problems. Monitoring, restricting contact and demanding constant availability are recognised markers of coercive control. The reason the framing matters is practical: treating control as an incompatibility invites you to negotiate with it, to find the middle ground, to ask better questions. There is no version of this list that a better conversation fixes.

If any of that is recognisable, the useful next step is a person rather than an article. Find A Helpline, in the sources below, lists free and confidential services by country. In the United States the National Domestic Violence Hotline is 1-800-799-7233, or text START to 88788, 24 hours a day.

Where the feeling underneath is jealousy rather than control, the distinction between the two is worked through in much more detail in the psychology of jealousy, and the phone-specific version of the same line is in what phone behaviour on dates actually shows.

How to ask without conducting an interview

The delivery has been studied, which is more than can be said for most of the advice about it.

Susan Sprecher and Stanislav Treger ran a live-interaction study comparing two ways of doing exactly the same amount of disclosure. In one condition pairs took turns, alternating disclosures back and forth. In the other, one person disclosed at length while the other listened, and they swapped roles in a second interaction. Same content, same total quantity, different shape. The turn-taking pairs reported greater attraction, particularly after the first interaction, and three process variables — responsiveness, enjoyment and the perception of being liked — were also higher in that condition and mediated the effect.

That is the mechanism underneath Aron's original design, which is reciprocal and escalating by construction: each slip is answered by both people, and the intensity rises gradually across forty-five minutes rather than starting deep. Two practical consequences follow. Answer your own question, ideally first, and at the depth you are hoping to receive. And move one step at a time rather than opening with the heaviest item on your list — a question that outruns where the other person currently is reads as an interrogation regardless of how good the question was. The same gradual-escalation point, and what it means when somebody does not reciprocate, is worked through in what the research on getting someone to open up actually shows.

The honest limit on both studies: they are laboratory work with strangers, measuring attraction and closeness over minutes. Nobody has shown that turn-taking disclosure predicts a relationship lasting. It shows that the form of the exchange changes the outcome independently of its content, which is a smaller claim and still a useful one.

One thing to leave out: the timing games. There is no evidence that withholding a question, or an answer, or a reply buys anything — the only experiment on post-date timing found a short wait beat no wait and a long one lost on every measure, which is the subject of the 72-hour rule. The first-meeting version of the same point is in first date mistakes.

What none of this can tell you

The limits are the part this genre skips, so they go at the end rather than in a footnote.

Change is not predictable. In the 11,196-couple project, models predicting change in satisfaction over the course of each study never accounted for more than about 5% of the variance, from any combination of self-report variables. That is the best-resourced attempt anyone has made at this question, and it says that whether a relationship gets better or worse from here is largely not visible in what either person can report today. It is also why no calendar-based rule can deliver what it promises: if change is that hard to see from the inside with two thousand measures available, a checkpoint at month six is not going to surface it.

The samples are narrow. The speed-dating work is undergraduates in one country. The 43-dataset project is overwhelmingly North America, Western Europe and Israel, and 99.4% of its relationship-specific variables were self-report rating scales rather than observations of behaviour. The Aron studies are one university's psychology classes in the 1990s. Everything here is evidence about particular people in particular places, and the honest reading of a body of work this uneven is that it constrains our confidence more than it licenses it.

And the questions are not a test. Nothing above scores a person, and no combination of answers earns a verdict. What a good question does is put real information on the table earlier than it would otherwise arrive: what a normal week looks like, how somebody behaves when something goes wrong, what they are unwilling to move on. What happens after that is not knowable in advance, and the largest study ever run on the subject is the reason for saying so. The few markers that do carry weight once there is a relationship to observe are collected in the signs a new relationship has real potential.

Every empirical claim on this page carries a linked source, which is the promise made in our editorial policy. The rest of this section is under Dating.