Research boundary: Every number in this paper is a group total computed inside Murudo's database. No name, photo, message or individual record was read or exported, and no group of fewer than 100 people is reported.
Abstract
Background. Dating platforms ask members what kind of relationship they want and print the answer on every profile. Whether that answer guides anyone's choices is rarely tested, and almost all evidence on stated and revealed partner preferences comes from North American and European samples.
Methods. We analysed all 2,080,648 like and pass decisions made between men and women on Murudo, a dating platform for Zimbabweans, from 1 May to 30 September 2026 (6,798 members deciding, 8,927 profiles judged). Stated intent was coded as committed (four labels, such as "Dating to marry") or open-ended (six labels, such as "Figuring it out"). The primary estimate is a Mantel-Haenszel odds ratio stratified by member, profile age and verification, tested against equivalence bounds of 0.90 to 1.11 fixed in advance. A within-member contrast, four robustness specifications and three benchmark attributes accompany it.
Results. Members who state a committed intent did not favour committed profiles. The adjusted odds ratio was 0.96 (95% CI 0.95 to 0.98) for men, statistically equivalent to no preference, and 0.90 (0.87 to 0.92) for women. Within members, committed men liked 16.7% of committed profiles and 18.4% of open-ended ones; committed women liked 6.1% of each. By contrast, a profile on the preferred side of the age line had 3.1 times the odds of a like from men and 5.1 times from women. Stated intent did predict the behaviour of the person stating it: likes sent to women who say committed were returned more often (5.1% against 3.2%; odds ratio 1.53, 1.45 to 1.61). Matches in which both people say committed were less likely, not more, to become a two-way conversation (48.8% against 56.8% when both say open-ended).
Conclusions. On this platform stated relationship intent works as a description of the person who states it, not as a criterion that others apply. Platforms should not infer intent preferences from likes, and should treat age and reciprocity as the signals that operate.
Keywords: online dating; stated preferences; revealed preferences; relationship intent; mate choice; age preferences; reciprocity; Zimbabwe
Introduction
Almost every dating profile answers one question before any other: what are you looking for? The answer is short, it is chosen from a list, and it sits near the top of the profile. The design assumes that people use it. Someone who wants to marry is expected to look for others who want to marry, and to pass over those who say they are only exploring.
That assumption has two parts, and they can be tested separately. The first is that the label is true of the person who chose it. The second is that other people act on it. Economics has long distinguished what people say they prefer from what their choices reveal (Samuelson, 1938), and psychology has shown that the two can differ sharply in partner choice. In speed-dating studies, the qualities people said they wanted beforehand did not predict whom they wanted after meeting (Eastwick and Finkel, 2008; Todd et al., 2007). These studies concern traits such as attractiveness and earning prospects. Relationship intent is a different kind of statement. It does not describe a quality of the partner. It describes the purpose of the search.
The evidence base is also narrow. Most studies of partner preferences draw on Western, educated and wealthy samples, which are unusual by global standards (Henrich, Heine and Norenzayan, 2010). We are not aware of a comparable published analysis of stated intent and revealed choice from a Zimbabwean platform, where the vocabulary of intent is itself different. Murudo's list includes "Dating to marry", "Ready to introduce families" and "Open to lobola conversation" beside the labels found on global apps.
Murudo's first research paper examined the platform's recommendation system from its source code and documentation alone, and ended by calling for local evidence (Murudo, 2026). This paper supplies some. It asks a single question: do people who say they want a committed relationship favour profiles that say the same? It answers with every like and pass decision made between men and women on the platform over five months, analysed as anonymous totals.
The paper makes four contributions. It tests stated intent against revealed choice inside each member, so that every person serves as their own control. It tests for the absence of an effect formally, using equivalence bounds fixed before the estimates were computed. It sets the size of the intent effect beside three benchmarks: age, distance and a verified badge. And it follows the label past the like, to whether likes are returned and whether matches talk.
The short version is that say and do part ways. Members who state a committed intent like committed and open-ended profiles at nearly the same rate. Age, by contrast, divides choices sharply at the point where two people are the same age. The label is not empty, though. It predicts what its owner does next.
Background and hypotheses
1 Stated and revealed preferences
People have limited insight into the causes of their own choices and often report plausible theories in place of actual reasons (Nisbett and Wilson, 1977). Partner choice is a well-studied case. Before a speed-dating event, men and women stated the usual sex-typed ideals. At the event, those ideals failed to predict whom they desired (Eastwick and Finkel, 2008). In a second sample of 46 speed-daters, stated preferences again did not predict actual choices (Todd et al., 2007). Across 10,526 participants in commercial speed-dating, choices followed attributes that can be seen in seconds, and education, income, religion and children had surprisingly little effect (Kurzban and Weeden, 2005). A meta-analysis of 97 samples found that physical attractiveness and earning prospects predicted romantic evaluations about equally for men and women, although the sexes state different ideals (Eastwick et al., 2014).
Online platforms record choices at a scale that surveys cannot reach. Hitsch, Hortaçsu and Ariely (2010a; 2010b) estimated preferences from browsing and first-contact decisions on a United States dating site and found no evidence of strategic behaviour, which supports reading such decisions as preferences. Reviews of the field conclude that platforms widen access to partners but that pre-meeting information predicts attraction poorly (Finkel et al., 2012; Joel, Eastwick and Finkel, 2017).
2 Intent as cheap talk
An intent label costs nothing to choose and nothing to change. In the economics of communication a costless message is cheap talk: it informs the receiver only to the extent that sender and receiver want the same thing (Crawford and Sobel, 1982). A signal becomes reliable when it is costly to send falsely (Spence, 1973). A field experiment on a Korean dating site showed the difference. Participants were given a small number of virtual roses to attach to proposals, and proposals carrying a rose were more likely to be accepted (Lee and Niederle, 2015). A scarce signal changed behaviour. A free label may not.
Two further forces weaken a stated intent. People bias self-reports toward what is socially approved (Fisher, 1993), and in many settings the approved answer to "what are you looking for?" is a serious relationship. People also use dating apps for several motives at once (Sumter, Vandenbosch and Ligtenberg, 2017), and a single label cannot carry them all.
3 Age and distance as benchmarks
An effect is easier to judge beside a known one. Age preferences are among the most replicated findings in the study of partner choice. Men on average prefer younger partners and women older ones, in 37 cultures (Buss, 1989), across the life course (Kenrick and Keefe, 1992) and in a replication across 45 countries (Walter et al., 2020). Platform data show the same asymmetry in choices: on OkCupid, men of every age rated women in their early twenties most highly (Rudder, 2014). Dating markets also have steep hierarchies of attention (Bruch and Newman, 2018), and men and women differ widely in how selective they are, with women liking a far smaller share of profiles (Tyson et al., 2016). We therefore report age, distance and verification as benchmarks, measured in the same way as intent.
4 Hypotheses
If stated intent guides choice, four things should hold.
- H1, concordance in likes. Members like profiles whose stated intent matches their own more often than profiles whose stated intent does not.
- H2, label premium. A committed label attracts more likes than an open-ended one, because most members state a committed intent themselves.
- H3, concordance in reciprocity. A like is more often returned when sender and receiver state the same intent.
- H4, concordance in conversation. A match is more likely to become a two-way conversation when both people state the same intent.
Setting and data
1 The platform and the intent label
Murudo is a dating platform for Zimbabweans at home and abroad, available as an Android app and on the web. A member builds a profile, sees other profiles one at a time in Discover and in a set of Daily introductions, and decides on each: like or pass. Two likes in opposite directions make a match, and a match opens a chat.
Every member chooses one relationship intent from ten labels. We group them into two classes. Committed covers "Long-term", "Dating to marry", "Ready to introduce families" and "Open to lobola conversation". Open-ended covers "Figuring it out", "Exploring", "Just exploring", "Friendship", "Short-term" and "It's complicated". The label is printed near the top of each profile card, beside gender and height, so it is in view at the moment of decision.
Three features of the product shape what members see, and so shape this study. Unless a member sets an age range, Discover shows profiles within three years of the member's own age. Members on a paid plan can filter Discover by intent. And after a member's first fifteen decisions the ranking system begins to learn from that member's likes and passes (Murudo, 2026). Age range and filter settings are stored on the member's device and were not available to us.
2 Sample
The study window runs from 1 May to 30 September 2026. In those 153 days members recorded 2,098,753 decisions. We kept decisions by men on women's profiles and by women on men's profiles in which both people had a stated intent, and removed three staff accounts. That leaves 2,080,648 decisions, 99.1% of the total. Decisions involving members of other genders were too few to report without risking the identification of individuals, and are a limit of this paper, not a judgement about whose choices matter.
The stratified odds ratios draw on all 2,080,648 decisions. The within-member contrast uses the 3,778 members.
Table 1. The analytic sample. Like rates per member are for members with at least 50 decisions (2,581 men, 1,867 women).
| Men deciding | Women deciding | |
|---|---|---|
| Members | 3,912 | 2,886 |
| Profiles judged | 4,200 women | 4,727 men |
| Decisions | 1,413,017 | 667,631 |
| Likes | 260,296 | 35,588 |
| Share of decisions that were likes | 18.4% | 5.3% |
| Like rate per member, median | 11.7% | 3.5% |
| Like rate per member, middle half | 5.4% to 23.1% | 1.5% to 7.6% |
| Mean age in years | 23.1 | 25.1 |
| Verified badge | 22.6% | 20.2% |
| State a committed intent | 66.3% | 81.9% |
These counts are smaller than Murudo's public totals of more than 20,000 sign-ups and more than 27,000 matches, and the difference is expected. The sample counts only members who still held an account when the data were read on 5 October 2026 and who made at least one decision in the window. Sign-ups that were never completed and accounts that were later closed are not in it, and neither are their matches.
The sample is young, with a mean age of 23 for men and 25 for women. Men and women differ in selectivity by a factor of more than three, as on other platforms (Tyson et al., 2016). The median man liked 11.7% of the profiles he judged and the median woman 3.5%.
3 Privacy safeguards
The analysis ran as read-only queries inside Murudo's production database. Each query returned totals for a group, such as the number of likes sent by men who state a committed intent. No row describing one person left the database, and no name, photograph, biography or message text was read. Message activity enters only as counts. We set a floor of 100 people: any group smaller than that is merged with another or left out, which is why several intent labels appear only in combination.
Measures and methods
1 Outcome and exposure
The unit of analysis is one decision: a member is shown a profile and likes or passes. The outcome is a like. The exposure is the intent class printed on the profile, committed or open-ended. The member's own intent class defines four groups, which we analyse separately: men and women who say committed, and men and women who say open-ended.
2 Within-member contrast
Members differ greatly in how often they like anyone, so pooled like rates mostly compare different people. For each member i with at least 20 decisions on each class of profile, we compute the difference between that member's own two like rates:
Δᵢ = Lᵢ(committed) / Nᵢ(committed) − Lᵢ(open-ended) / Nᵢ(open-ended)We report the mean of Δi across members with a 95% confidence interval, a one-sample t test against zero, and the standardised effect size dz, the mean difference divided by its standard deviation (Cohen, 1988). Each member counts once, however many decisions they made.
3 Stratified odds ratio
The primary estimate uses every decision. We compute the Mantel-Haenszel odds ratio (Mantel and Haenszel, 1959), which compares committed and open-ended profiles only inside strata and then pools:
ORMH = Σₛ (aₛ dₛ / nₛ) / Σₛ (bₛ cₛ / nₛ)where, in stratum s, a and b are likes and passes on committed profiles, c and d are likes and passes on open-ended profiles, and n is their sum. A stratum is one member deciding on profiles of one exact age in years and one verification status. Strata that contain only one class of profile carry no information and drop out. The estimate therefore compares profiles of the same age, with or without a badge, judged by the same person. Confidence intervals use the Robins-Breslow-Greenland variance, which remains valid when strata are small (Robins, Breslow and Greenland, 1986). An odds ratio above 1 means committed profiles are liked more.
4 Equivalence bounds
A result that is not significantly different from 1 does not show that an effect is absent. We therefore use an equivalence test (Lakens, 2017). Before computing any stratified estimate we set the smallest effect of interest at a 10% change in the odds of a like, which gives bounds of 0.90 and 1.11. An estimate whose 90% confidence interval lies wholly inside the bounds is declared practically equivalent to no preference.
5 Benchmarks and later outcomes
Three benchmarks use the same estimator. The age benchmark compares profiles on the side of the age line that each sex tends to prefer (younger than a man deciding, older than a woman deciding) with profiles on the other side. The distance benchmark compares profiles within 50 km with those farther away, among the 62% of decisions in which both locations are known. The verification benchmark compares profiles with and without a verified badge.
A like counts as mutual if the receiver also liked the sender at any time up to 5 October 2026. A match counts as a two-way conversation if both people sent at least one message, and as sustained if both sent at least five. Proportions carry Wilson intervals (Wilson, 1927).
6 What was fixed in advance
The study was not preregistered with a third party. After a first pooled look at raw like rates, and before any within-member or stratified estimate was computed, we fixed the two intent classes, the within-member contrast and its threshold of 20 decisions, the stratified estimator, the equivalence bounds, the four hypotheses, the three benchmarks and two robustness checks (each member's first fifteen decisions, and September alone). Adjustment for profile age was planned as a check and was promoted to the primary estimate once the first results showed that committed and open-ended profiles differ in age. Both versions are reported. The label-by-label analysis and the specification with photo count and sign-up month were added after results were seen and should be read as exploratory.
Results
1 What members say
Most members state a committed intent: 81.9% of women and 66.3% of men. "Dating to marry" alone is chosen by 55.8% of women and 33.5% of men, and it is the most common label for both.
The four other labels are "Ready to introduce families", "Open to lobola conversation", "Short-term" and "It's complicated", combined because each has fewer than 100 members of one sex.
2 Say versus do
Finding 1Members who say they want commitment do not favour profiles that say the same.
Take the 1,369 men who state a committed intent and who judged at least 20 profiles of each class. On average each liked 16.73% of the committed profiles he was shown and 18.37% of the open-ended ones. The within-member difference is −1.64 percentage points (95% CI −1.94 to −1.34; t = −10.7; dz = −0.29). Only 36.9% of these men liked committed profiles at the higher rate. For the 1,436 women who state a committed intent, the two rates are 6.11% and 6.14%, a difference of −0.03 points (−0.25 to +0.18; p = 0.76), and the women split 49.9% to 50.1%.
Members with at least 20 decisions on each class of profile. Each member counts once.
Men who say open-ended lean the same way as men who say committed, slightly more strongly (−2.06 points). Their choices are consistent with their label. The choices of committed men are not consistent with theirs.
Part of the raw gap is age. Committed profiles are about one year older on average than open-ended ones, for women's profiles (25.2 against 24.1) and for men's (23.3 against 22.3). Men tend to like younger women and women older men, so age pushes the unadjusted comparison in opposite directions for the two sexes. The primary estimate removes this by comparing profiles of the same age.
Table 2. Within-member like rates by the member's stated intent and the profile's stated intent, and the adjusted odds of liking a committed profile relative to an open-ended one. Unadjusted odds ratios are in Table 3.
| Members deciding | Liked, committed profiles | Liked, open-ended profiles | Difference, points [95% CI] | dz | Adjusted odds ratio [95% CI] |
|---|---|---|---|---|---|
| Men who say committed (n = 1,369) | 16.73% | 18.37% | −1.64 [−1.94, −1.34] | −0.29 | 0.96 [0.95, 0.98] |
| Men who say open-ended (n = 678) | 15.96% | 18.01% | −2.06 [−2.50, −1.61] | −0.35 | 0.87 [0.85, 0.89] |
| Women who say committed (n = 1,436) | 6.11% | 6.14% | −0.03 [−0.25, +0.18] | −0.01 | 0.90 [0.87, 0.92] |
| Women who say open-ended (n = 295) | 4.72% | 5.13% | −0.41 [−0.84, +0.02] | −0.11 | 0.83 [0.78, 0.89] |
Mantel-Haenszel odds ratios stratified by member, profile age in years and verification, with 95% confidence intervals. Log scale.
After adjustment, committed men like committed profiles with odds 0.96 times those for open-ended profiles (0.95 to 0.98). The 90% interval lies inside the equivalence bounds, so this group shows, in the formal sense, no preference of practical size. Committed women have an odds ratio of 0.90 (0.87 to 0.92): at the same age they like committed men slightly less. No group favours committed profiles. H1 is not supported.
The member's own label does leave a trace. Committed members are less averse to committed profiles than open-ended members are. The ratio of the two odds ratios is 1.10 for men (1.07 to 1.13) and 1.08 for women (1.00 to 1.16, p = 0.057). What a person says shifts what they do by about a tenth, in the stated direction. It is not enough to produce a preference for the kind of profile they say they want.
3 Which labels attract likes
Finding 2A committed label earns no premium. Marriage-track labels attract somewhat fewer likes than "Long-term" at the same age.
Pooling members of each sex, a committed label lowers the odds of a like by 7% from men (odds ratio 0.93, 0.92 to 0.94) and by 11% from women (0.89, 0.86 to 0.91). H2 is not supported. An exploratory comparison of single labels against "Long-term" shows where the difference lies.
Mantel-Haenszel odds ratios with 95% confidence intervals, log scale. Labels with fewer than 100 profiles of one sex are combined or omitted. Ten comparisons are shown; those for "Dating to marry" (both sexes), "Exploring" (both sexes) and, among men deciding, the family labels and "Friendship" remain significant after a Bonferroni correction.
"Dating to marry" profiles receive likes at odds 0.90 times those of "Long-term" profiles from men and 0.85 from women. "Exploring" profiles receive more: 1.13 from men and 1.07 from women. These are associations. Members choose their own labels, and people who choose "Dating to marry" may differ from others in their photographs or biographies, which this study did not read.
4 What does move a like
Finding 3Age moves the odds of a like three to five times over, and distance up to twice. Intent moves them by about a tenth.
Mantel-Haenszel odds ratios stratified by member with 95% confidence intervals, log scale. Intervals for age and distance are narrower than the markers. The distance row uses the 62% of decisions with both locations known.
A woman younger than the man deciding has 3.07 times the odds of a like that an older woman has (3.02 to 3.11). A man older than the woman deciding has 5.13 times the odds that a younger man has (4.93 to 5.33). Living within 50 km roughly doubles the odds for men deciding (2.02) and raises them by a third for women deciding (1.33). A verified badge adds 2% and 10%. The intent label sits on the other side of 1.
The age effect is not gradual. It is concentrated at the point where two people are the same age.
All decisions with both ages known. The outermost points pool gaps of 12 years or more. Unless a member sets a range, Discover shows profiles within three years of the member's age (shaded), so points outside the band come from members who widened their range.
Men liked 21.7% of women one to three years younger than themselves, 18.8% of women their own age and 11.2% of women one to three years older. Women liked 2.4% of men one to three years younger, 4.9% of men their own age and 7.1% of men one to three years older. Each curve is nearly flat on either side of the line and steps at it. On the right of the chart the two curves meet: men liked 6% to 9% of women four or more years older than themselves, about the same share that women liked of men four or more years older.
Zero means that liked and shown profiles were the same age on average. Ages with at least 100 members are shown singly; the last point for each sex pools three years of age (29 to 31 for men, 30 to 32 for women). Means are weighted by decisions.
Men at every age liked women about three quarters of a year younger than the women they were shown. Women liked men older than those they were shown, and the shift grew with their own age, from almost nothing at 18 to more than two years by the late twenties. The age structure of the platform works against both tendencies: women in the sample are on average two years older than men.
Distance behaves as a gradient. Men liked 26.7% of profiles less than 10 km away and 16.0% of profiles 300 to 1,000 km away. For women the figures were 7.0% and 5.4%.
5 What the label says about its owner
Finding 4Likes sent to women who say committed are returned about one and a half times as often.
Of the 260,296 likes men sent, 4.7% became mutual. The rate was 5.11% when the woman states a committed intent and 3.21% when she states an open-ended one. Stratified by sender, the odds ratio is 1.53 (1.45 to 1.61). It holds whatever the man's own label: 1.60 for men who say committed and 1.39 for men who say open-ended. So this is not concordance, and H3 is not supported as stated. It is a property of the women who choose the label. Women who say committed are more active on the receiving end: they made a decision on 25.0% of the men who liked them, against 20.4% for women who say open-ended, and when they decided they liked back more often (20.4% against 15.7%). They also liked a larger share of all profiles they judged (mean 6.4% against 4.9%; difference 1.5 points, 0.7 to 2.3).
For likes sent by women the pattern is the same in direction and much smaller. Of 35,588 likes, 34.5% became mutual: 35.0% when the man says committed and 33.5% when he says open-ended (odds ratio 1.09, 1.03 to 1.15).
A. Men's likes that became mutual scale 0 to 6%
B. Women's likes that became mutual scale 0 to 40%
C. Matches in which both people wrote scale 0 to 70%, with 95% interval
The three panels use different scales, stated beside each title. Message activity is counted, never read.
6 After the match
Finding 5Pairs who both say committed are the least likely to turn a match into a conversation.
The window produced 15,710 matches between men and women. In 64.0% at least one person wrote, in 50.2% both wrote, and in 14.2% both wrote at least five messages. Where only one person wrote, it was the man nine times in ten.
Two-way conversation was most common when both people say open-ended (56.8%, 53.2 to 60.3) and least common when both say committed (48.8%, 47.8 to 49.8). The two mixed pairings fall between. The four rates differ more than chance allows (χ² = 27.4, 3 degrees of freedom, p < 0.001), and the gap between the two concordant pairings is 8.0 points (4.4 to 11.7). Sustained conversation shows the same order with less separation: 13.4% when both say committed and 15.3% in each of the other three pairings (χ² = 10.9, p = 0.012). H4 is not supported. Its sign is reversed.
7 Robustness
Table 3. Odds of liking a committed profile relative to an open-ended one under five specifications, with 95% confidence intervals.
| Specification | Men, committed | Men, open-ended | Women, committed | Women, open-ended |
|---|---|---|---|---|
| Stratified by member only | 0.89 [0.88, 0.90] | 0.82 [0.80, 0.83] | 1.00 [0.97, 1.02] | 0.89 [0.83, 0.95] |
| Primary: member, profile age, verification | 0.96 [0.95, 0.98] | 0.87 [0.85, 0.89] | 0.90 [0.87, 0.92] | 0.83 [0.78, 0.89] |
| Primary plus photo count and sign-up month | 0.96 [0.94, 0.98] | 0.87 [0.84, 0.89] | 0.88 [0.85, 0.91] | 0.83 [0.76, 0.90] |
| September 2026 only | 0.94 [0.91, 0.97] | 0.86 [0.82, 0.90] | 0.93 [0.87, 0.99] | 0.83 [0.72, 0.96] |
| First 15 decisions of each member | 0.97 [0.87, 1.07] | 0.77 [0.66, 0.89] | 0.83 [0.71, 0.97] | 0.91 [0.64, 1.29] |
The conclusion does not depend on the specification. Across twenty estimates none exceeds 1.00. Adding the profile's number of photographs and month of sign-up to the strata changes nothing of substance. September alone, where the current label is most likely to be the one shown at the time, gives the same pattern. Each member's first fifteen decisions, made before the ranking system has learned anything about that member, are fewer and the intervals are wide, but the estimates point the same way.
Discussion
1 Summary of evidence
Table 4. Hypotheses and verdicts.
| Hypothesis | Prediction | Result | Verdict |
|---|---|---|---|
| H1, concordance in likes | Committed members favour committed profiles | Odds ratio 0.96 for men and 0.90 for women | Not supported |
| H2, label premium | Committed profiles receive more likes | Odds ratio 0.93 from men and 0.89 from women | Not supported |
| H3, concordance in reciprocity | Likes are returned more when labels match | Likes to committed receivers are returned more whatever the sender says: 1.53 for women receiving, 1.09 for men receiving | Not supported as stated |
| H4, concordance in conversation | Matched pairs with the same label talk more | Both wrote in 48.8% of committed pairs and 56.8% of open-ended pairs | Reversed |
| Benchmark | Age, distance and verification move likes | Age 3.07 and 5.13; distance 2.02 and 1.33; badge 1.02 and 1.10 | Confirmed |
All four predictions of the view that stated intent guides choice fail. The label does carry information, but about the sender. It is associated with how readily a woman responds to interest, and hardly at all with who shows interest in her.
2 Why say and do part ways
Several explanations fit the data, and this study cannot choose between them.
The label does not discriminate. When four women in five and two men in three state a committed intent, the label separates few people from the rest. In the language of signalling, almost everyone sends the same costless message, and a message that everyone sends carries little news (Crawford and Sobel, 1982; Spence, 1973).
The decision is fast and visual. A like is made in seconds, mostly on photographs. Earlier work finds that choices at this stage follow what can be seen at once and that stated criteria are poor guides to them (Kurzban and Weeden, 2005; Todd et al., 2007). The step in Figure 7 is that kind of cue: age is printed beside the name and takes no reading.
The label may be read as a description, not a requirement. A member who wants marriage may not treat "Figuring it out" as a reason to pass. Many people who want a committed relationship would describe themselves as still working out the details. On this reading members do not ignore the label. They do not take the two classes to be opposed.
The approved answer may be overstated. If "Dating to marry" is the respectable thing to say, some members will say it who are in practice open to less (Fisher, 1993). That would explain why men who state a committed intent choose much as men who do not. It would not explain the women's results, where the label does track behaviour.
The lower rate of conversation among pairs who both say committed is the least expected result. One possibility is that these members hold more matches at once and spread their attention more thinly. Another is that people with serious aims are slower to open a conversation they are not sure of. We did not read messages and cannot say.
3 Relation to earlier work
The gap between stated and revealed preferences reported for traits (Eastwick and Finkel, 2008; Eastwick et al., 2014) extends here to the purpose of the search itself, in a sample from outside the settings where it has mostly been studied. The age results replicate a pattern documented across cultures (Buss, 1989; Walter et al., 2020), with one refinement that survey data cannot show: the preference acts as a threshold at equal age more than as a slope. Men's liked ages rose by about 0.4 years for each year of their own age, a weaker fixation on the youngest profiles than Rudder (2014) reported, though the default three-year window limits what members here could show. The difference in selectivity between men and women matches Tyson et al. (2016).
The finding on reciprocity bears on the argument of our first paper, that a dating recommender should estimate whether interest will be returned and not only whether it will be expressed (Murudo, 2026). Here the two diverge measurably. Men lean slightly toward open-ended profiles, while their likes are returned half again as often by committed ones.
4 What this means for members
Three practical points follow, each within the limits of an observational study.
- Saying what you want costs a few likes, not matches. Profiles that say "Dating to marry" receive somewhat fewer likes at a given age. But a like received by a member who states a committed intent more often becomes a match, most clearly among women, because those members answer more of the interest they receive.
- Do not assume the other person chose you for your label. Most likes are sent without regard to it. If intent matters to you, read theirs and ask early.
- Age decides more than people admit. A profile one to three years on the wrong side of equal age is liked about half as often by men, and about a third as often by women, as one on the preferred side.
5 What this means for platform design
- Do not learn intent preferences from likes. Murudo's ranking system keeps like and pass counts by stated intent for each member (Murudo, 2026). These results show that the signal in those counts is small and, for committed members, runs against what they say they want. A system that learned from it would show committed members fewer committed profiles. Intent is better treated as a declared constraint that the member sets than as a taste inferred from behaviour.
- Rank on returned interest. The label that predicts a returned like is not the one that attracts the like. Ranking by the chance of mutual interest would use the information that the label does carry.
- Give intent a cost or a consequence. A statement that is free to make and free to ignore will be both. Scarce signals change behaviour where free ones do not (Lee and Niederle, 2015).
- Test it properly. Whether the lower like rate for marriage-track labels is caused by the label can be settled only by an experiment that varies how prominently intent is shown.
- Record intent over time. Keeping the date of each change of label, and the declared age and distance settings, with members' knowledge, would allow the stated side of this comparison to be measured as precisely as the revealed side.
6 Limitations
- Observational design. Members choose their labels and the platform chooses what to show. Stratifying by member removes differences between people who decide, and stratifying by age and verification removes two differences between profiles. Other differences between profiles remain.
- Current labels. The database holds each member's present intent, not the one displayed when a decision was made. A member who changed label is classified by the later one. The September-only estimates reduce this problem without removing it.
- Selected exposure. Decisions are observed only on profiles that were shown. The default age window, paid intent filters and learned ranking all shape that set, and filter settings were not available. A paying member who filtered to one class contributes nothing to within-member estimates.
- Survivorship. Likes and passes are deleted with an account. The sample covers members who still had an account on 5 October 2026, and members who left after finding a partner are missing.
- Sample. Members are young and chose to join a dating platform. Results describe them and not Zimbabweans in general. Only decisions between men and women are analysed.
- Coding. The two classes are our grouping of ten labels. Figure 5 shows that labels inside a class differ, and other groupings are defensible.
- Location. Distance uses current coordinates and covers 62% of decisions.
- Multiple comparisons. Many estimates are reported. The primary analysis was fixed in advance, but analyses marked exploratory were not.
- One platform, five months. Nothing here establishes that the pattern holds elsewhere or will persist.
7 Next studies
Three follow-ups would test the explanations above. An experiment on the prominence of the intent label would separate the effect of the label from the effect of who chooses it. A comparison of Daily introductions, where members see a few chosen profiles, with open browsing in Discover would show whether having fewer options makes members weigh stated intent more. And a count of who writes first and who replies, by intent, would show whether the conversation gap arises at the first message or later.
Conclusion
People on Murudo say clearly what they want. Four in five women and two in three men state a committed intent, and more than half of women say they are dating to marry. Their choices do not follow the label. Members who say they want commitment like committed and open-ended profiles at almost the same rate, a committed label earns no premium from either sex, and pairs who share it talk no more than pairs who do not. What members act on is what can be seen at once, above all whether the other person is older or younger than they are.
The label still tells the truth about something. Women who state a committed intent respond to more of the interest they receive and return more of it. On this evidence a stated intent describes how its owner will behave, and it is not a filter that others apply. A platform that wants to connect people who want the same thing cannot rely on members to sort themselves by a label. It has to do some of that work in how it ranks, and it should measure the result by interest returned and not by interest expressed.
Statements
Ethics and privacy. The study used records that Murudo already holds in order to run its service. It changed nothing that any member saw and involved no contact with members. All statistics were computed as group totals inside the production database, no individual record was exported, and no group of fewer than 100 people is reported. The study was not reviewed by an institutional ethics board. Zimbabwe's Cyber and Data Protection Act governs personal information held by the platform (Government of Zimbabwe, 2021).
Data and code. The underlying records are personal data and are not shared. Every estimate was computed from group totals by the procedures set out under Measures and methods.
Competing interests. Murudo Labs is part of Murudo, the company that operates the platform studied. The results were not selected for commercial advantage, and several describe weaknesses in how the platform uses stated intent.
About Murudo Labs. Murudo Labs is the research side of Murudo. It studies how people look for a partner, using anonymous totals from the platform, and publishes what it finds.
References
Bruch, E.E. and Newman, M.E.J. (2018) ‘Aspirational pursuit of mates in online dating markets’, Science Advances, 4(8), eaap9815. Available at: https://doi.org/10.1126/sciadv.aap9815.
Buss, D.M. (1989) ‘Sex differences in human mate preferences: Evolutionary hypotheses tested in 37 cultures’, Behavioral and Brain Sciences, 12(1), pp. 1–14. Available at: https://doi.org/10.1017/S0140525X00023992.
Cohen, J. (1988) Statistical Power Analysis for the Behavioral Sciences. 2nd edn. Hillsdale, NJ: Lawrence Erlbaum Associates.
Crawford, V.P. and Sobel, J. (1982) ‘Strategic information transmission’, Econometrica, 50(6), pp. 1431–1451. Available at: https://doi.org/10.2307/1913390.
Eastwick, P.W. and Finkel, E.J. (2008) ‘Sex differences in mate preferences revisited: Do people know what they initially desire in a romantic partner?’, Journal of Personality and Social Psychology, 94(2), pp. 245–264. Available at: https://doi.org/10.1037/0022-3514.94.2.245.
Eastwick, P.W., Luchies, L.B., Finkel, E.J. and Hunt, L.L. (2014) ‘The predictive validity of ideal partner preferences: A review and meta-analysis’, Psychological Bulletin, 140(3), pp. 623–665. Available at: https://doi.org/10.1037/a0032432.
Finkel, E.J., Eastwick, P.W., Karney, B.R., Reis, H.T. and Sprecher, S. (2012) ‘Online dating: A critical analysis from the perspective of psychological science’, Psychological Science in the Public Interest, 13(1), pp. 3–66. Available at: https://doi.org/10.1177/1529100612436522.
Fisher, R.J. (1993) ‘Social desirability bias and the validity of indirect questioning’, Journal of Consumer Research, 20(2), pp. 303–315. Available at: https://doi.org/10.1086/209351.
Government of Zimbabwe (2021) Cyber and Data Protection Act [Chapter 12:07], Act 5 of 2021. Available at: https://zimlii.org/akn/zw/act/2021/5.
Henrich, J., Heine, S.J. and Norenzayan, A. (2010) ‘The weirdest people in the world?’, Behavioral and Brain Sciences, 33(2–3), pp. 61–83. Available at: https://doi.org/10.1017/S0140525X0999152X.
Hitsch, G.J., Hortaçsu, A. and Ariely, D. (2010a) ‘Matching and sorting in online dating’, American Economic Review, 100(1), pp. 130–163. Available at: https://doi.org/10.1257/aer.100.1.130.
Hitsch, G.J., Hortaçsu, A. and Ariely, D. (2010b) ‘What makes you click?—Mate preferences in online dating’, Quantitative Marketing and Economics, 8(4), pp. 393–427. Available at: https://doi.org/10.1007/s11129-010-9088-6.
Joel, S., Eastwick, P.W. and Finkel, E.J. (2017) ‘Is romantic desire predictable? Machine learning applied to initial romantic attraction’, Psychological Science, 28(10), pp. 1478–1489. Available at: https://doi.org/10.1177/0956797617714580.
Kenrick, D.T. and Keefe, R.C. (1992) ‘Age preferences in mates reflect sex differences in human reproductive strategies’, Behavioral and Brain Sciences, 15(1), pp. 75–91. Available at: https://doi.org/10.1017/S0140525X00067595.
Kurzban, R. and Weeden, J. (2005) ‘HurryDate: Mate preferences in action’, Evolution and Human Behavior, 26(3), pp. 227–244. Available at: https://doi.org/10.1016/j.evolhumbehav.2004.08.012.
Lakens, D. (2017) ‘Equivalence tests: A practical primer for t tests, correlations, and meta-analyses’, Social Psychological and Personality Science, 8(4), pp. 355–362. Available at: https://doi.org/10.1177/1948550617697177.
Lee, S. and Niederle, M. (2015) ‘Propose with a rose? Signaling in internet dating markets’, Experimental Economics, 18(4), pp. 731–755. Available at: https://doi.org/10.1007/s10683-014-9425-9.
Mantel, N. and Haenszel, W. (1959) ‘Statistical aspects of the analysis of data from retrospective studies of disease’, Journal of the National Cancer Institute, 22(4), pp. 719–748. Available at: https://doi.org/10.1093/jnci/22.4.719.
Murudo (2026) ‘Data science in reciprocal dating platforms: A socio-technical case study of Murudo’, Murudo Blog, 13 July. Available at: https://murudo.co/blog/data-science-and-dating-apps/.
Nisbett, R.E. and Wilson, T.D. (1977) ‘Telling more than we can know: Verbal reports on mental processes’, Psychological Review, 84(3), pp. 231–259. Available at: https://doi.org/10.1037/0033-295X.84.3.231.
Robins, J., Breslow, N. and Greenland, S. (1986) ‘Estimators of the Mantel-Haenszel variance consistent in both sparse data and large-strata limiting models’, Biometrics, 42(2), pp. 311–323. Available at: https://doi.org/10.2307/2531052.
Rudder, C. (2014) Dataclysm: Who We Are (When We Think No One's Looking). New York: Crown.
Samuelson, P.A. (1938) ‘A note on the pure theory of consumer's behaviour’, Economica, 5(17), pp. 61–71. Available at: https://doi.org/10.2307/2548836.
Spence, M. (1973) ‘Job market signaling’, Quarterly Journal of Economics, 87(3), pp. 355–374. Available at: https://doi.org/10.2307/1882010.
Sumter, S.R., Vandenbosch, L. and Ligtenberg, L. (2017) ‘Love me Tinder: Untangling emerging adults' motivations for using the dating application Tinder’, Telematics and Informatics, 34(1), pp. 67–78. Available at: https://doi.org/10.1016/j.tele.2016.04.009.
Todd, P.M., Penke, L., Fasolo, B. and Lenton, A.P. (2007) ‘Different cognitive processes underlie human mate choices and mate preferences’, Proceedings of the National Academy of Sciences, 104(38), pp. 15011–15016. Available at: https://doi.org/10.1073/pnas.0705290104.
Tyson, G., Perta, V.C., Haddadi, H. and Seto, M.C. (2016) ‘A first look at user activity on Tinder’, Proceedings of the 2016 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 461–466. Available at: https://doi.org/10.1109/ASONAM.2016.7752275.
Walter, K.V., Conroy-Beam, D., Buss, D.M. et al. (2020) ‘Sex differences in mate preferences across 45 countries: A large-scale replication’, Psychological Science, 31(4), pp. 408–423. Available at: https://doi.org/10.1177/0956797620904154.
Wilson, E.B. (1927) ‘Probable inference, the law of succession, and statistical inference’, Journal of the American Statistical Association, 22(158), pp. 209–212. Available at: https://doi.org/10.1080/01621459.1927.10502953.