Do Dating Apps Actually Work? What the Data Says

Four different questions are hiding inside this one

Do dating apps work is asked as if it were a single question. It is at least four, and they have different answers.

The first is whether the apps put you in front of people you would not otherwise have met. The second is whether relationships come out of them. The third is whether those relationships are worse than the ones that begin some other way. The fourth is whether the thing the companies actually sell — matching, the claim that a questionnaire and an algorithm can work out who suits you — does anything at all.

Each has been measured separately, by different researchers, using different methods, and the results do not point the same way. The first two are about as settled as anything in this field gets, and the evidence is good enough that arguing with it is not reasonable. The third is a null finding with a known size attached, which is a far better kind of answer than it sounds. The fourth is where the honest reply is no, and the distance between the fourth answer and the first two is the whole story of this subject.

Underneath all of them sits a fifth question, which is the one most people actually mean: will it work for me. That one has no published answer. No dating company releases a rate of relationships formed per user, and the surveys that do exist count couples who already formed rather than the people swiping this month. What can be done instead is to describe the distribution you would be stepping into, and the measurement studies that did that are the least quoted and the most useful work in the area.

The market-level figures — how many people use these apps, what the companies file with regulators, what share of the population is partnered with someone they met online — are laid out in the piece on modern dating statistics. This page is about whether the thing works, and for whom. Our editorial policy explains why every number below carries a linked source.

At the level of the population, the answer is not in doubt

The best measurement of how couples meet is a sociological survey, not an app's press release. Michael Rosenfeld, Reuben Thomas and Sonia Hausen used two waves of How Couples Meet and Stay Together, 2,473 heterosexual couples from 2009 and 2,997 from 2017, and charted the route into each relationship against the year the couple met.

Among heterosexual couples in the United States, meeting online went from about 2% in 1995 to about 39% in 2017 (Z = 23.43). Meeting through friends went the other way over the same period, from 33% to 20%. The crossover — the year online overtook friends — was around 2013, ending roughly six decades in which friends had been the most common route. The authors report separately that phone apps accounted for at least half the growth in meeting online between 2010 and 2017.

The detail nobody repeats is in the same table. Meeting in a bar or restaurant rose from 19% to 27%, which reads like a revival of meeting people in person, and it is not. The paper says the post-2010 rise in that category is due entirely to couples who met online and then had their first in-person meeting at a bar or restaurant. Take those couples out and the category was significantly declining after 1995. One of the few apparently offline trends in modern dating turns out, on inspection, to be the apps again.

The argument sits in the paper's title, Disintermediating your friends, and it rests on two numbers. In the 2009 survey, 11.2% of couples who met online had been introduced or helped along by some third person. By 2017 that was 3.7%, and 89% of couples who met online had no personal connection of any kind beforehand. The internet did not merely add a venue. It took over a job that friends and family used to do.

The limit matters and the authors are clear about it: this counts couples that formed, not people who are using an app. A route can be the most common way into a relationship and still be a frustrating month for almost everyone on it, which is exactly what the later sections of this page describe. It is also a United States sample, and the categories are softer than they look — plenty of people who say they met through a friend exchanged their first messages on a phone.

The relationships are not worse, and this null has a size

The standard worry is that something begun by swiping is flimsier than something begun through a friend. It has been tested on national longitudinal data, and the test is more careful than its summaries.

Michael Rosenfeld followed the How Couples Meet and Stay Together respondents from 2009 through follow-ups in 2010, 2011, 2013 and 2015: 1,341 couples who met in 1995 or later, 57,081 couple-months of exposure, 407 breakups. The raw comparison looks damning — couples who met online had odds of breakup 2.46 times higher. It is also meaningless, and he says so. At the first wave, couples who met online had been together 4.4 years on average and couples who met offline 19.1 years, and long-established couples break up less. Nearly all the couples who met before 1995 met offline.

Restrict the comparison to couples who met during the internet era and the raw odds ratio falls to a non-significant 1.31 (0.91 to 1.89). Control for relationship duration and the rest, and it is 0.96 (0.66 to 1.39). In plain rates: couples who met offline after 1995 broke up at 7.3% a year, and couples who met online would have broken up at a modelled 7.0% if they had the same demographic profile.

What makes this a real null rather than a shrug is that Rosenfeld reports the power of his own data. If the true odds of breakup were 1.5 times higher for couples who met online, this study had a 0.802 chance of detecting it; at 2.0 times higher, 0.999. So the finding is not we looked and saw nothing. It is that an effect of the size the critics describe would very probably have shown up, and it did not.

The study usually quoted for the opposite conclusion — that meeting online is actively better — needs its conflict-of-interest statement read first. John Cacioppo and colleagues surveyed 19,131 Americans who married between 2005 and 2012 and found marital break-up at 5.96% among those who met online against 7.67% offline, and satisfaction at 5.64 against 5.48 on seven-point scales. The survey was run by Harris Interactive, commissioned by eHarmony.com; the first author was a scientific advisor to eHarmony and the second a former Director of eHarmony Laboratories. The paper reports an agreement reached before analysis that results bearing on eHarmony would not affect publication, which is a sensible precaution and not the same thing as independence. The authors' own verdict on the size of the effects is that they are understandably quite small and do not hold for everyone.

So the defensible position is narrow and worth stating exactly. There is no evidence that relationships which start online are worse. The evidence that they are better amounts to a gap of 0.18 on a seven-point scale in a survey paid for by a dating company. And nobody can randomly assign a person to meet their partner on an app, so selection sits underneath all of it. Once two people are together, how they met stops explaining much, and what the research on couples who last actually reports takes over.

They are not a hookup machine either, and the best evidence is Swiss

The other standing objection is that apps select for people who do not want a relationship, so whatever comes out of them is unserious. That is an empirical claim, and it has been checked against a national survey.

Gina Potarca used a 2018 nationally representative Swiss survey and a sample of 3,245 respondents who had met their partner within the previous ten years. The findings run against the stereotype in three directions. Couples formed through dating apps had stronger cohabiting intentions than couples formed in non-digital settings. Women who had found their partner through an app had stronger fertility desires and intentions than women who met their partner offline. And there were no differences in relationship satisfaction or life satisfaction between app couples and everybody else.

There is a fourth result that is more interesting than any of those and almost never mentioned. Apps were associated with couples who were more educationally diverse and more geographically distant than couples who met other ways. A technology built around location filters is, on this data, mixing people who would not have crossed paths in a workplace or a university. What distance does to a relationship once it exists is a separate and much better-measured question.

Two limits. This is Switzerland, not the United States, and it is again a survey of couples who formed rather than of people searching. What people say they are searching for is a different measurement, and Pew Research Center asked it: among US adults who had used a dating site or app in the previous year, 44% said a major reason was to meet a long-term partner, 40% to date casually, 24% to have casual sex and 22% to make new friends. Men were much more likely than women to name casual sex as a major reason, 31% against 13%.

None of that says apps are not used for casual sex. It says the population on them is mixed, that the mix differs sharply by gender, and that the couples who come out the other end are, on the one national dataset that has looked, at least as oriented towards building something as anyone else. The arrangements that stall in between are a third thing with their own literature.

The thing they actually sell has never been shown to work

Matching is the product. It is what justifies the questionnaire at sign-up, the subscription, and the word compatibility in the marketing. It is also the part with no supporting evidence, and the gap between what is claimed and what has been demonstrated is the largest in this subject.

Pew asked Americans directly whether they thought these programs could determine whether two people will eventually fall in love. 21% said yes. 35% said they could not and 43% were unsure.

Samantha Joel, Paul Eastwick and Eli Finkel ran the test. In two speed-dating studies, unattached participants completed more than 100 self-report measures covering the traits and preferences that the mate-selection literature had identified as relevant, and then went on four-minute dates with every opposite-sex participant at the event. Random forests were then asked to predict three separate things from the pre-date measures. They predicted 4% to 18% of actor variance — how much a given person tended to like others. They predicted 7% to 27% of partner variance — how much others tended to like that person. And they were unable to predict relationship variance at all, with any combination of the measures: the part specific to one pair, which is precisely and only what a matching algorithm claims to deliver.

There is a structural reason, and the biggest study in relationship science makes it visible. Joel and 85 colleagues pooled 43 longitudinal datasets from 29 laboratories, 11,196 couples, 2,413 measures, with couples tracked across an average of four time points over 14 months. The variables that predicted relationship quality were relationship-specific judgments — perceived partner commitment, appreciation, sexual satisfaction, perceived partner satisfaction, conflict — and they accounted for up to 45% of variance at baseline and up to 18% at the end of each study. Individual differences, the kind of thing a sign-up questionnaire can capture, managed 21% and 12%. Partner reports added nothing beyond a person's own relationship-specific judgments.

Read those two results together and the problem is not that the algorithms are badly built. Every variable that does substantial predictive work is a judgment about a relationship that already exists. None of it can be collected before two people have met. An algorithm is being asked to forecast the output of a process using only the inputs that were available before the process started, and when researchers tried exactly that with a hundred measures and modern methods, the part they could not reach was the part that matters. What a set of questions can and cannot reveal is worked through in the piece on compatibility questions.

What did predict a better first date was believing the algorithm worked

Liesel Sharabi ran a two-wave longitudinal study of online daters, surveying them before and after a first date with someone from a dating site. Her summary of the result is the most quotable sentence in this literature: whether the algorithms actually worked mattered less than whether participants had the perception that they worked. Participants reported better first dates to the extent that they believed in the efficacy of the compatibility matching process.

Read that carefully, because it is self-reported and correlational and an obvious alternative explanation is sitting right there. People who are optimistic about a date may both believe in the algorithm and enjoy the evening, with the belief doing none of the causal work. What the study establishes is an association, not a mechanism. What it does suggest — and this is a defensible thing to pay for, if you want to — is that part of what a matching service delivers is confidence walking into the room, which is not nothing and is not what the advertising says it is.

Two other studies have looked at what happens in the gap between the match and the meeting, which is the part of this process a user can actually change.

Sharabi and John Caughlin surveyed 186 people before a first date with someone from a dating site and 94 of them afterwards, and had participants hand over the actual emails they had exchanged through the site. First-date success was predictable from features of their online impressions and from the relational dynamics in that correspondence — which is to say the messages carried real information, and not only about whether somebody writes well.

Artemio Ramirez, Erin Sumner, Christina Fleuriet and Megan Cole surveyed a national sample of online daters and found a curvilinear relationship between how much communication happened before the first in-person meeting and what the meeting was like: intimacy, composure, informality, social orientation, forecasts about the future of the relationship, and how much information-seeking people did when they met.

Curvilinear is the important word and the one that gets flattened in retellings. It means neither end is the good place: too little messaging and too much both predict a worse meeting than something in between. The paper does not license a number of days, and anyone who gives you one — three days, a week, two weeks — is inventing it, in the same way the three-day rule was invented. What the result supports is only this: time spent messaging before meeting is a variable with an interior optimum, not a resource to maximise. The wider evidence on what a thread can and cannot carry is in the piece on texting.

There is no average experience, and a measurement study on Tinder shows why

Everything above is about couples and populations. The question a person actually brings to this page is about their own month, and the most honest answer is that the distribution is so skewed that an average is close to useless.

Gareth Tyson, Vasile Perta, Hamed Haddadi and Michael Seto created fourteen Tinder profiles — stock photographs, a real volunteer with one picture and with three, a profile with no picture, a profile whose picture said the account had been disabled — placed them in London and New York, liked large numbers of other users and recorded what came back. They collected 230,000 male and 250,000 female profiles in the process.

The male profiles matched with 0.6% of the users they liked. The female profiles matched with 10.5%.

The pattern continues after the match. 21% of the female profiles' matches sent a message; 7% of the male profiles' matches did. Men messaged almost immediately — a median of 2 minutes after the match, with 63% of their messages arriving within five minutes — against a median of 38 minutes for women. And the median message length was 12 characters from men against 122 from women, with a quarter of men's opening messages under six characters, which is to say hi. In the authors' survey, 49% of the women said fewer than 30% of their matches turned into a conversation.

One result is worth sitting with on its own: 86% of the matches the male profiles received came from other men, even though the dataset was roughly evenly split by gender. Whatever the male experience of this app was in 2016, a large part of it was not happening with the population the user was looking for.

The limits are real and should travel with every number here. This is 2016, it is one app, and it is a measurement study using constructed profiles rather than a sample of real users' outcomes. The authors describe their match rates as a lower bound, because they could not confirm that every user they liked was shown their profile in return. But none of that touches the structural point. A single figure for the match rate does not exist. Two people using the same app in the same week are running experiments with different odds by more than an order of magnitude, and advice built on one person's experience of it will be worse than useless to the other.

The same asymmetry in the inbox, measured across four cities

Elizabeth Bruch and Mark Newman took every first message sent between users of a popular online dating site in New York, Boston, Chicago and Seattle during January 2014, and used the resulting network to rank desirability — not by attractiveness ratings but by who receives messages from people who themselves receive a lot of messages.

The distribution is long-tailed in both directions. Most people receive a handful of messages at most. The most popular individual across those four cities, a 30-year-old woman in New York, received 1,504 messages in the month, which is one every half hour, day and night, for thirty-one days.

Two findings follow, and both are more useful than the usual advice. First, people reach: both men and women message partners about 25% more desirable than themselves. Second, reaching has a price. The probability of a reply falls as the desirability gap grows. Men are more than twice as likely to get a reply from a woman less desirable than themselves than from a more desirable one, and for messages sent to more desirable women the reply rate never rises above 21%. Women's average reply rate overall is under 20%.

The paper also tested the obvious advice. Both men and women write longer messages to more desirable targets, up to twice as long. Longer messages were positively associated with response rates in exactly one of the four cities, for men in Seattle — who were also the only men writing longer messages to more desirable targets in the first place. Writing more positively was, if anything, slightly counterproductive for men in all four cities.

Pew measured what that distribution feels like from inside it. Among people who had used a dating site or app in the previous year, 64% of men said they had felt insecure because of the lack of messages they received, against 40% of women, while 54% of women said they had felt overwhelmed by the number, against 25% of men. Those are two populations describing the same product.

Which means the question does it work is partly a question about where you sit in a distribution nobody will show you. It also explains why this subject generates so much bad folklore: someone whose experience sits at one end of that curve writes advice that is simply inapplicable at the other, and the pattern of drawing general rules from a position you cannot see is the same one examined in why intelligent people keep choosing badly.

Too much choice: the standard objection, and three results that do not line up

The most common intellectual criticism of dating apps is choice overload: that an endless menu makes people less satisfied with whoever they pick and less willing to commit. Three pieces of evidence bear on it and they disagree, which is worth showing rather than resolving.

As a general law, it is not there. Benjamin Scheibehenne, Rainer Greifeneder and Peter Todd meta-analysed 63 conditions from 50 published and unpublished experiments, N = 5,036, and found a mean effect size of virtually zero, with considerable variance between studies and no sufficient conditions they could identify for when it appears. The famous jam experiment is real; a reliable general phenomenon built on it is not.

In this specific domain, it is there. Jonathan D'Angelo and Catalina Toma ran an experiment in which online daters selected a partner from a set of 24 or a set of 6. A week after choosing, the participants who had chosen from the large set were less satisfied with their choice and more likely to want to change it. Reversibility on its own made no difference, but the combination of a large pool and the ability to reverse produced the least satisfied participants of all.

And at the level of actual relationships, it is not there. If an overabundance of options were destabilising couples, then the couples who met in the largest choice set — online — should break up more. In Rosenfeld's national longitudinal data they do not, with the power to detect a moderate effect if one existed.

The way to hold those three together without picking a side: there is decent experimental evidence that a big menu makes you less happy with what you chose for a while afterwards, and good national evidence that this does not show up in whether couples last. Both can be true if the effect is real, small, and about the choosing rather than about the relationship. The public is split on the same question: Pew found 43% of US adults saying people have the right amount of options on these platforms, 37% saying too many and 13% saying not enough.

What does appear to build something over time, once a choice has been made, has nothing to do with the size of the pool it came from, and is covered in what actually builds attraction over time.

What using them does to the person using them

There is a literature on this and the kindest thing to say about it is that it is young. Gabriel Bonilla-Zorita, Mark Griffiths and Daria Kuss ran a systematic review across PsycINFO and Web of Science and found 43 studies. Heavier use was associated with neuroticism, sociability, sensation-seeking and sexual permissiveness. Using the apps for sex-search and for self-esteem enhancement predicted problematic use. The authors' own conclusion is substantially about methodological weakness, and the review should be read as a map of what has been asked rather than a verdict on what is true.

The users' own ratings are mixed and not catastrophic. In Pew's survey of 6,034 US adults, among those who had ever used a dating site or app, 53% said their personal experiences had been positive and 46% negative. Men were more positive than women, 57% against 48%, with women almost evenly split. 35% had paid at some point, and paid users reported positive experiences more often than people who had never paid, 58% against 50% — which could be the product being better, or could be that people who are enjoying it are the ones willing to pay.

Then there is a category that is not an experience-quality question at all, and it is the part this genre skips.

52% of people who have used a dating site or app say they came across someone they think was trying to scam them, rising to 63% among men under 50. 38% have received unsolicited sexually explicit messages or images. 30% have had someone keep contacting them after being told they were not interested. 24% have been called an offensive name. 6% have been threatened with physical harm, and among younger women who use these platforms it is 11%.

Those numbers change what the word works means for a large share of users. An argument about yield — matches, dates, relationships — assumes the main problem is that not enough happens. For many women on these apps the defining problem is the opposite, and fixing the matching would not touch it.

The line is worth naming plainly, because it is the same one that runs through the rest of this site. Someone continuing to contact you after you have declined is not persistence. Pressure after a no is coercion, not poor etiquette. Being asked for your location, your schedule or your phone before you have met anyone in person is a request you are allowed to refuse without explanation. Where that behaviour comes wrapped in intensity and reads as interest, the piece on jealousy and control sets out the difference between a feeling and a behaviour. The helplines at the top of this page and in the sources below are free and confidential.

So: do they work?

Four questions, four answers, and they are not the same answer.

As a way of meeting people: yes, decisively. Online is the most common route into a heterosexual relationship in the United States and has been since around 2013, and it got there by replacing introductions that friends used to make rather than by adding to them.

As a producer of relationships that hold: yes, as far as anyone can tell. Couples who met online break up at a rate indistinguishable from everybody else's, in a study with the power to have caught a moderate difference. The claim that they do better rests on a smaller effect in a survey commissioned by a dating company.

As a matching service: no. Nothing collected before two people meet has been shown to predict the part of attraction that is specific to the pair, and the variables that do predict relationship quality are all judgments about a relationship that already exists. Believing the algorithm works predicted a better first date; the algorithm working has never been demonstrated.

As an experience: it depends enormously on who you are, and nobody will show you the distribution before you join it. Match rates differed by more than a factor of seventeen between the male and female profiles in the one open measurement study, reply rates fall as you reach upwards, and the most common complaint among men and the most common complaint among women are opposites.

What follows practically is short. Use them for the function the evidence supports — access to people outside your network — and discount the function it does not. Judge a slow month against the distribution rather than against somebody's story. Move to a real meeting earlier than feels natural, because the only pre-meeting variable with measured consequences has an optimum in the middle rather than at the far end. And then stop asking whether the app worked, because the moment two people have met, the question becomes one the app has no bearing on at all: what happens between them. That is the level at which signs that a new relationship has real potential is the useful page, and where our relationship satisfaction test and attachment style test are aimed — both of which say on their own pages what a score cannot tell you. More in this vein under Dating.