Transparency
How the compatibility score works
Last updated: October 31, 2026
The compatibility score is an estimate of how likely the two of you would be happy together in the long run. It is plain arithmetic over what you both wrote on Compass. The rules below reflect what's in the open-source code. You can open the breakdown on any profile to see how you both compare.
Keep in mind that the score is only a rough estimate. A score below 50% usually means the two of you clash on a dealbreaker, such as children, distance or age. However, a high score does not mean you would be happy together for sure. It simply makes this person one promising candidate among many, as most of what makes two people compatible only shows up once you talk. So treat it as a starting point, stay open-minded, and connect with different people to figure out what works best for you.
What the evidence says
The evidence is very thin since the gold standard of studies — randomly matching people and measuring how they get along — is hard to implement ethically (you can't set people up and force them to date long-term). And, unfortunately, the ones who deployed matching algorithms at scale (mainstream dating apps running on ads and subscriptions) have financial incentives not to make their algorithms work, let alone to publish their data on whether they work. Let's look at what's been done and what we take from it.
- Attraction after a speed date (Joel, Eastwick & Finkel, 2017). Participants filled a hundred of self-report measures before going on randomly-selected speed dates. Then they rated their attraction to their date. Part of the attraction could be explained by how attractive the date is to people in general and by how attracted the participant is to people in general. However, the questionnaire results didn't explain whether the pair would be more or less attracted to each other. This is not surprising when you dig into the details: the participants were all 19-year-old students at a single US university. The questionnaire also glossed over the questions that seem to us most predictive of a happy relationship: children, location, religion, diet, major, intended career, life plan (marriage, how to raise children, etc.) and other values we have in our own compatibility prompts. This is exacerbated by the fact that all the participants come from the same background, making a questionnaire hard to discriminate among them. Rather, most of their questions are on personality traits such as Big Five, attachment, sociosexuality and self-ratings of themselves and their ideal partners (rational, creative, optimistic, etc.). But the most important reason why we do not think this study should carry much weight in the research of relationship compatibility is the absurdly short duration of the interaction: FOUR minutes. Four minutes measures first attraction, nothing near whether a relationship lasts (i.e., compatibility).
- What predicts a happy couple (Joel et al., 2020). The same lead author then pooled 11,196 couples from 43 studies and asked what best predicts how satisfied each partner is. The facts about each person (personality, mood, age, education, income) explained a fifth of it at most. What each person said about the relationship itself (how committed their partner seems, how appreciated they feel, sex, conflict) explained twice as much, and the partner's own answers added nothing on top. In the authors' words, "who I am" barely matters once you know "who I am when I am with you". This doesn't answer whether two strangers would get along, though. These were couples who had already chosen each other, often years earlier, so anyone with a dealbreaker between them had either never become a couple or had split before the study started. Due to this survival bias, the effect of any mismatch is understated: the couples who crossed a gap (religion, politics, education, etc.) and are still together are the ones for whom it mattered least. The authors themselves call for research on the early stages of relationships, where these filters act.
- Interfaith marriages end more often (Lehrer & Chiswick, 1993; Call & Heaton, 1997; Vaaler, Ellison & Powers, 2009). Satisfaction surveys only hear from couples still together, so they cannot see what a mismatch cost the ones who are not. Divorce is an event, and demographers have tracked it for decades in large US household panels. Couples of different faiths dissolve more often than couples of the same faith; the gap is largest between faiths that differ most and shrinks when the two are close in belief and practice. Shared religion weighed about as much as age at marriage, and a difference in how often the two attend services mattered more than the label each one carries. The usual limits apply: US samples from the 1990s, and people who married across faiths were probably already the ones least attached to theirs, which makes the effect look smaller than it is.
- What couples actually match on (Horwitz et al., 2023). A meta-analysis of 199 studies on 22 traits, plus 79,074 couples in the UK Biobank, asking a simpler question than the ones above: how alike are the two people in a couple? Very alike on political values (correlation 0.58), religiosity (0.56), education (0.55, from almost two million pairs) and smoking (0.49), somewhat on drinking (0.42), and barely on personality: openness 0.21, conscientiousness 0.16, neuroticism and agreeableness 0.11, extraversion 0.08. The study only sees couples that already exist, so a correlation blends three things it cannot separate: people choosing someone like them, people meeting mostly people like them (campuses, workplaces, neighbourhoods), and partners growing alike over the years. It also says nothing about whether the alike couples are happier.
- Do matching algorithms work? (Finkel et al., 2012). A sixty-page review of online dating. On matching, they found no compelling evidence that any algorithm works simply because eHarmony, Chemistry and the others never released their data or their methods. And the few studies they pointed to compared their own customers' success rate with a random slice of daters in the population. That says more about how motivated their customers are (people who paid for a marriage-oriented site, were older, and sat through a long questionnaire) than about whether those same customers would have fared any worse without the algorithm, browsing the profiles themselves. The authors go further and argue that no algorithm fed only with what is known before two people meet can explain more than a small share of long-term outcomes, because it cannot know how the pair will interact, grow and cope with what life throws at them. They do grant that similarity in the values that organise a shared life, such as religion and views on family roles, is "a beginning", and that algorithms can at least weed out people who make poor partners in general. Keep in mind that it reviews an industry that never let anyone look, so "no evidence" means untested, not tested and failed.
- Dealbreakers over dealmakers (Jonason et al., 2015). Six studies with more than 6,500 people. Students listed what would rule someone out, a survey of 2,744 single American adults ticked their dealbreakers from a list, and other groups rated made-up partners described in a few lines, with and without a flaw. People weighed a flaw more heavily than a comparable quality, more so for a long-term relationship than for a fling, and women more than men. The most common dealbreakers were an unhealthy lifestyle, being untrustworthy or uncaring, already dating someone else, and wanting a different kind of relationship. Except for the national survey, the participants were again mostly students, nobody met anyone, and asking people what they would reject is different from actually rejecting it.
- Perceived versus actual similarity (Montoya, Horton & Kirchner, 2008). A meta-analysis of 313 studies. Being similar to someone predicts liking them strongly when the two never meet and only read a description of each other, less when they talk briefly, and not at all once they are in an existing relationship. Feeling similar, on the other hand, predicts liking in all three settings. Most of those studies are old laboratory experiments in which students rated a stranger from a filled-in attitude questionnaire, about as far from a relationship as one can get, and when people like someone they also tend to see them as similar, so part of the effect runs the other way.
- Self-disclosure builds closeness (Aron et al., 1997). Pairs of strangers in a psychology class spent 45 minutes answering 36 questions that gradually got more personal, from whether they would like to be famous to when they last cried in front of someone, while a control group made small talk. Afterwards the question pairs rated their new acquaintance as closer than the closest relationship in the lives of about 30% of comparable students, and weeks later 57% had talked again and 35% had done something together. Matching the pairs on attitudes or telling them they would like each other changed nothing. These were again undergraduates at one university, the pairs were mixed-sex or two women, and the authors say plainly that the goal was closeness at the end of the hour, not a relationship; the "36 questions to fall in love" headlines came much later. We take the mechanism rather than the result: reciprocal, escalating disclosure creates closeness quickly, and agreeing on opinions does not. A prompt answered with a personal explanation is a small first step of that kind.
What we take from it. The evidence is poor, and nobody should claim an algorithm can pick your partner with more than 10% of success rate. Conversely, we have no evidence that an algorithm is useless either. What emerges for us is that the score is helpful for the following reasons:
- Screening out structural mismatches. A profile accurately predicts a lack of compatibility when there is a dealbreaker. They sit at the top of the breakdown, and we filter them out by default in your search results.
- Surfacing the pool of promising candidates. Preferences (values, habits, shared ground) and absence of dealbreakers suggest promising candidates, but without enough precision to single out the one compatible person among them. Similarity in values, plans and background relates to stability where similarity in personality traits largely does not. That is why our prompts ask about values and behavior and why the Big Five row carries little weight.
- Giving you something to talk about. Most of what makes two people compatible is only revealed when they interact. A prompt answered with an explanation is a small act of self-disclosure: it builds closeness and catalyzes the interaction that tells you whether a relationship would work.
Dealbreakers first
When one of them applies, the pair gets no score at all:
- Gender — you are not each other's preferred gender.
- Kind of connection — one of you is only looking for friendship and the other only for a relationship.
- Relationship style — one of you is strictly monogamous while the other is strictly polyamorous or open.
- Children — you both want a relationship and gave opposite answers about wanting children (more than two steps apart on the five-point scale). Neighbouring answers are a difference of degree and count below instead.
What the score is made of
We compare your two profiles across different fields, each producing a number between 0 and 1, plus a weight that measures how important it is.
- Prompt answers — weight up to 8, growing with shared answers. For every question you both answered: your answer, the answers you would accept, and how much it matters to you, in both directions. The weight starts at 0 and reaches its maximum at 100 shared answers.
- Age range — weight 3. Full marks inside each other's stated range. Outside it: half marks for 1–2 years, a fifth for 3–5, almost nothing beyond. The worse direction counts.
- Wanting children — weight 3. Same answer → 1, one step apart → 0.85, two steps → 0.6. Only when you both want a relationship.
- Distance — weight 3. Full marks within 25 miles, then a gentle slide that never goes below 0.6 — unless one of you set a maximum distance: past it the row drops progressively to almost nothing at two and a half times the cap.
- Relationship status — weight 2. One of you single and the other married or in a long-term relationship, with neither open or polyamorous → 0.1. Only when you both want a relationship.
- Shared language — weight 2. At least one language in common → 1, none → 0.1.
- Shared interests, causes and work — weight 2. Compared across all three lists by name, so "effective altruism" counts whether you filed it as an interest or a cause. More overlap, higher score; nothing shared → 0.5.
- Diet — weight 2 if either of you is vegan or vegetarian, else 0.5. Vegan with vegan, or vegetarian with vegetarian → 1; vegan with vegetarian → 0.9; plant-based with pescetarian → 0.6; plant-based with omnivore → 0.3. Two omnivores earn little from it (0.8).
- Religion — weight 1.5. Something in common → 1, nothing → 0.4.
- Politics — weight 1.5. Something in common → 1, nothing → 0.4.
- Relationship style — weight 1. Identical answers → 1, overlapping → 0.8. Only when you both want a relationship.
- Cannabis, psychedelics — weight 1 each. Your stated preference about a partner's use, against their actual use: "prefer they don't" with a regular user → 0.1, "prefer they use" with a never → 0.3, and so on. The worse direction counts.
- Big Five personality — weight 1. 1 minus the average difference across the five traits.
- Smoking — weight 0.5. Same → 1, different → 0.5.
- Drinking — weight 0.5. Closer monthly counts score higher; 20 or more drinks apart → 0.4.
- Exercise — weight 0.5. Same frequency → 1, one bucket apart → 0.8, two → 0.5.
- Education — weight 0.5. Within one level → 1, two apart → 0.8, further → 0.6.
How the rows combine
The score is a weighted geometric mean of the rows. In plain words: multiply the rows together, each raised to its share of the total weight. Unlike an average, one very low row on something that matters pulls the whole number down rather than being cancelled by a dozen good ones — which is what a dealbreaker-lite should do. No row can go below 0.05, so a single bad row never zeroes the score on its own.
The likely range
Every row is an estimate, so every row carries a range, and the score shows one too: "72%, likely between 58% and 84%". Rows both of you answered have a narrow range. A row that one of you left blank is not treated as a mismatch: it is counted at a neutral value with a wide range and half its weight. That is why a half-empty profile reads as "could be anywhere from 40% to 90%" rather than a confident number, and why the range narrows as either of you fills in more. The prompt row's range narrows as you both answer more questions.
Sorting uses the low end of the range. When you sort people by compatibility, the order follows the bottom of each range, not the number in the middle. Two profiles both showing 86% are not equal when one is "80 to 90" and the other "56 to 98": the second has left most fields blank and happens to match on the few it filled in. Ranking by the low end puts the pair you know more about first, and it means filling in your profile or answering prompts moves you up. The number you see on a card is still the estimate in the middle.
What is deliberately left out
Ethnicity and height are not scored since members can't currently set a preference about them on Compass. MBTI is skipped because its validity is weak and the Big Five covers the ground. Neurotype, sexual orientation (your gender preference already carries the match), where you grew up, and all free text — bio, keywords, occupation — are not scored either.
What can we improve?
This is the first release of the score (beyond the one based on the prompts only, which doesn't track deal-breakers), and we are testing it. We designed hand-picked rules and weights, not a model trained on outcomes, so we expect to change them as we get feedback. If a row seems wrong or if the number does not match your experience with someone, tell us!
What would move us most is evidence: a study, a dataset or your own experience showing that profiles and algorithms predict compatibility better than we think, or worse. We are as ready to make the score more ambitious as to make it more modest. Send it through the contact form or open an issue on GitHub.