What actually predicts a good relationship, and what doesn’t

DANIEL JORDA · OCTOBER 3, 2026 · 7 MIN READ

What relationship research says predicts a good relationship, what it says doesn’t, and how that shaped the questions Veram asks.

Before deciding what Veram would ask people, I wanted to know what relationship researchers have actually found, as opposed to what magazine quizzes assume. Here's the short version. Some of it was humbling, and it changed what we built.

Some context: three in ten US adults have used a dating site or app, and one in ten partnered adults met their current partner on one, according to a Pew Research Center survey from 2022. That's a lot of people being introduced by software. It's fair to ask what it should pay attention to.

Chemistry is hard to predict before two people meet

In two speed-dating studies, Samantha Joel, Paul Eastwick and Eli Finkel had single people fill out more than 100 measures of their traits and of what they wanted in a partner. Then each person had four-minute dates with every opposite-sex participant. Machine-learning models could predict, to a degree, how much a person tended to desire others in general and how much others tended to desire them. They could not predict who would be drawn to whom in particular, from any combination of the answers given before the dates.

That fits a long 2012 review in which Finkel and four colleagues looked at online dating through the lens of psychological science. They found no compelling evidence that matching algorithms lead to better relationships than other ways of pairing people. Part of the reason, they argued, is that algorithms lean on similarity, which matters less to how relationships turn out than people have long assumed.

Similarity: what you feel matters more than what's on paper

"Birds of a feather" is one of psychology's best-known ideas, so this surprised me. A 2008 meta-analysis by Matthew Montoya, Robert Horton and Jeffrey Kirchner compared actual similarity (how alike two people are on questionnaires) with perceived similarity (how alike they feel). Actual similarity predicted attraction strongly when people had never interacted, much less after a short interaction, and not significantly among people already in relationships. Perceived similarity held up at every stage.

A later speed-dating study by Natasha Tidwell, Eastwick and Finkel found the same thing face to face. After four-minute dates, perceived similarity predicted romantic interest. Actual similarity didn't.

What you say you want vs. what moves you

There's a similar gap between stated preferences and real reactions. In a 2014 meta-analysis, Eastwick and colleagues noted that men and women reliably differ in how much looks and earning prospects sway them when rating hypothetical partners. But in studies where people evaluated real partners and dates rather than hypothetical ones, physical attractiveness mattered about equally for both sexes, earning prospects had a small effect for both, and the differences between men and women were small and not significant.

What does predict how a relationship goes

The largest study I found pooled 43 longitudinal datasets from 29 labs, covering 11,196 couples, and used machine learning to rank what researchers had measured (Joel and colleagues, 2020). The strongest predictors of how satisfied and committed people were came from their own view of the relationship: how committed they believed their partner was, how much they appreciated their partner, sexual satisfaction, how satisfied they thought their partner was, and how much they fought. Together, those explained up to 45% of the differences in relationship quality at the start of each study, and up to 18% at the end.

Things about the individual mattered less, but still mattered: satisfaction with life in general, negative mood, depression, and anxiety or avoidance around closeness. Demographics like gender, race and education mattered little. Once a person's own view of the relationship was known, adding their traits, or their partner's answers, added nothing. And none of it could predict whose relationship would get better or worse over time.

The honest reading: most of what predicts a good relationship lives inside the relationship. Before two people meet, there isn't one yet.

Conflict, and coming back from it

John Gottman's lab learned a great deal by watching couples argue. In a study that followed 130 newlywed couples for six years, criticism, defensiveness, contempt and stonewalling during conflict predicted divorce. Gentleness and calming each other down helped, and so did husbands being open to their wives' influence. Active listening, a staple of couples advice, rarely happened in real arguments and didn't predict which marriages lasted.

In another study, Gottman and Robert Levenson asked 79 couples to discuss an ongoing disagreement for 15 minutes, then talk about something pleasant. The emotional tone of that second conversation, the one right after the disagreement, predicted divorce over the next four years at least as well as the argument itself did. How a couple comes back from a fight seems to matter as much as the fight.

One caution: these studies fit their models to the same couples whose outcomes they reported. Finkel's review cites an analysis where a model built on half of a sample was tested on the other half, and the share of divorce it accounted for fell from 65% to 21%. Take the patterns seriously, and the accuracy figures with care.

How this shaped what Veram asks

We don't claim to predict chemistry. A match percentage on Veram is a reason to meet, not a verdict. Every study above points the same way: the real test is meeting.

We show where you rhyme, with a line from each of you. If feeling alike matters more than matching on a checklist, it's better to show you what each of you is actually like and let you judge than to tell you that you both ticked "hiking."

We ask how you handle a disagreement, and every introduction names one honest point of friction. We can't watch two strangers argue, and wouldn't want to. But conflict is too central in this research to leave out.

We ask about likely dealbreakers. Across six studies with more than 6,500 people, Peter Jonason and colleagues found that people weigh dealbreakers more heavily than they weigh dealmakers, and that unhealthy lifestyles and wanting different kinds of relationships were among the common ones. In Pew's 2019 survey, 63% of online daters said it was very important for a profile to say what kind of relationship someone wanted, and 45% said the same about whether they had children. So the seven questions cover kids, relationship shape, smoking, where you'll be living and what the next two years are about.

We ask what a good week with a partner looks like. The 2020 study found the best predictors were how people experience the relationship itself, not traits on paper. An ordinary week is the closest we can get to that before two people meet.

None of the seven questions are about height or income. Earning prospects had only a small effect on people's real romantic evaluations in the 2014 meta-analysis. We didn't think income earned a question.

The research is humbling for anyone building a dating app. It says the biggest part of a good relationship is made by the two people in it, after they meet. Our job is smaller and, I think, more honest: avoid the obvious mismatches, show you something true about each other, and get out of the way.