Can anyone predict who you will click with?
Every matching system sells the same promise. Answer our questions and we will find the person, the therapist, the team, the match. I went looking for whether anyone has ever demonstrated that, and what I found is one unusually careful study. One. That is a real limit, I will keep saying so, and it is still worth your time, because nobody has produced a study showing the opposite, and this one is stranger than a simple no.
Where this ends up: the spark between two specific people is real, and it is the largest single thing in the data, and nobody has predicted it in advance from anything either person can say about themselves. It becomes visible only once they have actually met, which changes what a matching service can honestly claim to be.
Evidence Start with what attraction is made of. When a group of people all meet each other and rate every interaction, the ratings split three ways. How much you tend to like people in general. How much people tend to like you in general. And a third piece that belongs to neither person alone, the spark that shows up only in one specific pairing. In speed-dating data that third piece is the biggest of the three, roughly a third of all the variation (Joel, Eastwick & Finkel, 2017).
So the thing matching services claim to find is real. It is not small, and it is not mystical. Now the finding.
Evidence The same researchers collected more than 100 questionnaire measures from 350 people, one to two weeks before a series of speed dates. Traits, preferences, what they said they wanted in a partner. Then they trained machine-learning models to predict the dates.
The models could partly predict the first two pieces. How much a person would like their dates in general, and how much they would be liked. The third piece, the spark between two specific people, they could not predict at all. Trained on one sample and tested on the other, the correlations were -.06 and .02, which is less than a tenth of one percent of the variation. Some estimates came out worse than just guessing the average for every pair.
Read that plainly. The part everyone actually wants, the reason you would use a matching service at all, was invisible to a hundred questionnaires and a modern algorithm. The parts that could be predicted are the parts nobody is asking for.
Now the red flag, in the open where it belongs. This is one research programme. The two samples share a team, a campus and a method, and the second mostly functions as a test set for models trained on the first, so it is one well-executed study and not a replicated literature. The major 2012 review that reached a similar conclusion about matching sites shares two of the same authors, so it is the same voice agreeing with itself, not independent confirmation. My verification pass went looking for a study showing pair-specific prediction working and did not find one, but part of that search ran out of budget, so absence here is weaker than it sounds. If someone demonstrates pre-meeting pair prediction in an independent sample, this essay changes.
Evidence Here is the strange half. After the dates, the same component stopped being invisible. Measures taken once two people had actually interacted, felt chemistry above all, explained roughly 16 to 29 percent of that same pair-specific spark. The authors’ own conclusion was that desire for a specific partner could not be predicted from traits and preferences measured before the pair had met.
Before, not ever. The information is real. It just does not exist yet when the questionnaire is filled in. It comes into existence when two people interact, and no amount of asking either of them beforehand can reach it, because neither of them has it.
That reframes what a matching algorithm can honestly be. Not a predictor of chemistry, because chemistry is not encoded in what people can report about themselves. At best, a way of arranging meetings and getting out of the way.
Suggestion And it raises an uncomfortable question about what such systems are doing when they appear to work. The predictable component is how much a person is generally desired, which in this study ran on self-rated mate value and physical attractiveness. A system that quietly ranks on general desirability, then presents the results as your compatibility, will look like it works. It is predicting something. It is not predicting the thing on the label.
Two cautions belong in the open, not a footnote. The study is one research programme’s two samples of young students meeting for four minutes, and its authors say plainly that it speaks only obliquely to long-term compatibility. Nobody should read it as proof that matching services have no value, and the authors themselves do not. And there is a darker path to making the numbers look better. The paper notes that a more demographically diverse sample might show predictable matching on age, class, culture and religion. In other words, the easiest way for an algorithm to seem to predict chemistry is to start sorting people by demographics. If a matching system’s accuracy ever suddenly improves, that is the first question to ask it.
Question The question I am left with is not about dating. Every field that matches people, to therapists, to teams, to mentors, to jobs, runs on information collected before anyone has met. If the pair-specific part only comes into existence in the interaction, then the honest design in every one of those fields is cheap first meetings and fast feedback, not longer questionnaires. I do not know of anyone building it that way, and I would like to.
Where this stops. Joel, Eastwick and Finkel 2017, Psychological Science 28(10). Three independent checks in my verification pass read the primary tables. The two samples share one research team, one campus and one method, so this is one well-executed study, not a replicated literature. The percentages for the predictable components vary by how they are computed, and I have used the conservative cross-sample figures. An independent re-execution audit in 2025 found minor errors and let the core conclusions stand.
What I am not claiming. That commercial matching algorithms were tested. They were not; the predictors were academic questionnaires, not swipes or messages. That long-term compatibility is unpredictable. Untested here. And a widely cited study claiming marriages that began online were happier is not cited on this page, because it was commissioned and funded by eHarmony and its lead author was a paid adviser to the company.