Type two names into a love calculator and you get a number. Take a twenty question compatibility quiz with a partner and you get a percentage, a label, and usually a short paragraph that feels a little too accurate.
The question worth asking is not whether these are fun. They obviously are. The question is whether a test taken before or early in a relationship can tell you anything real about how that relationship will go. Relationship science has spent a long time on exactly this, and the answer is clearer than you might expect.
What the Research Actually Found
In 2012, a team led by Eli Finkel published a long review in Psychological Science in the Public Interest looking at online dating from the perspective of psychological research. One section dealt specifically with matching algorithms, the proprietary systems that claim to pair people by compatibility.
Their conclusion was blunt. There was no published, peer reviewed evidence that any of these algorithms could do what they advertised, which is predict how well two specific people would get along once they actually met. The algorithms were reasonably good at describing individuals. Turning two descriptions into a prediction about a pair was where the chain broke.
A 2017 study in Psychological Science by Samantha Joel, Paul Eastwick and Finkel put the idea through a harder test. They ran machine learning on speed dating data, using more than a hundred self reported traits and preferences collected before participants met. The models could predict two things reasonably well:
- How much a given person tended to like others in general
- How much a given person tended to be liked by others in general
What they could not predict at all, above chance, was the part everyone wants: the unique spark between one specific pair. Whatever happens when two particular people sit down together was not recoverable from what they had said about themselves beforehand.
Then in 2020, Joel and a large group of collaborators published a study in PNAS that pooled 43 longitudinal datasets covering more than 11,000 couples. Machine learning again, this time predicting relationship quality over time. The finding that matters here: variables describing the relationship itself, such as how appreciated someone felt, how committed they perceived their partner to be, how satisfied they were, predicted outcomes substantially better than either partner's individual traits.
Put those together and the picture is consistent. What predicts a relationship is mostly generated by the relationship. It is not sitting inside two people waiting to be measured before they interact.
Why the Results Still Feel Right
If the predictive value is near zero, why does the paragraph after the quiz so often feel written for you?
Bertram Forer ran the classic demonstration in 1948. He gave students a personality assessment, then handed every single one of them the same generic description assembled from horoscope columns. Students rated it, on average, better than 4 out of 5 for accuracy as a description of themselves specifically.
The statements did the work. Lines like "you have a great deal of unused capacity" or "at times you are outgoing and sociable, at other times reserved" are true of almost everyone, which is precisely why they land as personal. The effect is stronger when the reader believes the result was generated for them, and stronger again when the content is mildly flattering.
Compatibility results are built from the same material. "You two connect quickly but need to work on speaking up when something bothers you" describes nearly every couple that has ever existed.
There is a second effect stacked on top. Once a couple is handed a number, they start explaining it. A high score gets confirmed by remembering the times you agreed. A low score gets laughed off or quietly rationalised. Either way the quiz did not measure the relationship, it gave the couple a prompt and the couple supplied the evidence.
What Does Predict Relationship Quality
The research does point somewhere useful. It just points at behaviour rather than traits.
How conflict gets repaired. John Gottman's observational work on couples found that the differentiator was not the absence of conflict, it was whether repair attempts during conflict actually landed. Contempt, meaning mockery and talking down to a partner, was the standout warning pattern in that research programme. None of this is measurable from a questionnaire taken before the first argument.
Perceived commitment and appreciation. These were among the strongest predictors in the 2020 PNAS analysis, and both are relationship specific by definition. Someone can feel deeply appreciated in one relationship and invisible in the next.
Responsiveness. Whether a partner reliably notices, understands and acts on what you need. It shows up across the literature under different names and it is a behaviour pattern, not a personality score.
Similarity, mostly overrated. Shared hobbies do poorly as predictors. Shared values and life goals do somewhat better. Neither approaches the predictive weight of how a couple treats each other in practice.
The uncomfortable summary is that the most predictive information about a relationship only becomes available after the relationship has been running for a while.
The One Thing Quizzes Genuinely Do Well
They start conversations that would not otherwise happen.
This is not a consolation prize. In 1997 Arthur Aron and colleagues published a procedure now known widely as the 36 questions, a structured set of increasingly personal questions two people answer in turn. The study was not about compatibility. It was about generating closeness, and the escalating self disclosure produced measurably more closeness than small talk did.
That is a real mechanism, and it is available to any quiz. A question like "what would you do if I lost my job" or "how do you want to be treated when you are upset" is one most couples never quite get around to on a Tuesday night. A quiz gives it a container and a reason.
So the practical use is this. The score is decoration. The conversation the score triggers is the product.
How to Use a Compatibility Test Well
1. **Answer separately first.** If you answer together, you converge without noticing. Independent answers preserve the differences, and the differences are the interesting part.
2. **Read the mismatches, skip the total.** A 62 percent means nothing. Two opposite answers on how you each handle an argument means something worth twenty minutes of talking.
3. **Ask why, not who is right.** The follow up question that does the work is always some version of "what makes you answer it that way".
4. **Do not let a high score end an inquiry.** A flattering result is the least informative outcome, because it invites both people to stop thinking.
5. **Do not let a low score become a verdict.** No quiz has the information required to make one. If a low score genuinely unsettles you, the useful move is asking what you were afraid it would confirm.
6. **Retake it in a year.** Comparing answers across time tells you something a single snapshot never can, because now you are measuring change inside an actual relationship rather than guessing from traits.
Where Love Testers Fit
Matcha AI is a love tester and compatibility quiz app, and the honest framing is the one this whole article argues for. It is entertainment that is good at starting conversations. It is not an assessment, and no app on any platform can be, for the reasons the research above lays out.
What it does well is lower the barrier. Handing a phone across a table is easier than announcing that you would like to have a serious conversation about how you each deal with conflict, and the questions get you to roughly the same place. Play it as a game, then talk about the two answers that surprised you.
What it cannot do is tell you whether the relationship will work. That question is answered over months by behaviour, not in ninety seconds by a quiz, and any app that claims otherwise is selling a number.
For related reading, relationship red flags that matter covers the patterns that do have research behind them, the 5 love languages explained looks at what that popular framework does and does not support, how to tell if someone likes you covers early stage signals, and best iPhone apps for couples covers the wider category.
The Bottom Line
Compatibility tests cannot predict compatibility. That is not cynicism, it is the repeated finding of the largest studies that have looked, from the 2012 algorithm review through the 2017 speed dating models to the 2020 analysis of 11,000 couples. The information that predicts a relationship is produced by the relationship, which means it does not exist yet at the moment a quiz is answered.
They are still worth taking. Just move your attention from the number to the two questions where your answers did not match, because that is the only part of the result that contains anything you did not already know.