Signs She Likes You: What the Evidence Actually Supports

What the evidence says a person does when they like you

Start with the largest thing anyone has measured. Montoya, Kershaw and Prosser pooled 309 effect sizes from 5,422 participants and asked which enacted behaviours track self-reported attraction. The overall association was z = .20. Eleven behaviours reached significance. Indirect ones: eye gaze duration and frequency, mimicry, laughter, smile frequency, head nodding. Direct ones: physical closeness, amount of talking, duration of talking. Two more were specific to speed dating: saying yes to a partner repeatedly, and asking for future contact. Direct behaviours were more strongly related to attraction than indirect ones, which is worth holding on to, because the genre is almost entirely about indirect ones.

An association of z = .20 is the whole relationship between liking and behaviour, across every study they could find. It is real and it is not large. Whatever a list of signs can do, that is its ceiling.

The first behaviour on that list is also the one this genre leans on hardest, and it has a second problem on top of the small effect: held gaze changes how the person holding it feels, so it partly manufactures what it is being used to detect. What eye contact actually predicts goes through that evidence and the measured comfortable duration, which is a little over three seconds.

The other half of the result is the part nobody quotes. Several of the most-repeated cues showed no evidence of a relation to liking at all. In the University of Dayton's summary of the paper, Montoya lists them: hair flips, eyebrow raises, gestures, head tilts, primping, open body posture, leaning in.

Three of those were in the version of this page that stood here until today, written up as if each one were a tell. They are gone. Our piece on what body language actually predicts repeated some of the same claims and was rewritten against this same bibliography the following day, retraction included. Our editorial policy is why that gets said out loud instead of quietly patched.

Those behaviours are liking behaviours, not romance behaviours

Here is the finding that should end the genre and has not.

Montoya's conclusion about the eleven surviving behaviours is that they are the behaviours research has already linked to the development of trust and rapport. They are not a courtship system. They are the general-purpose machinery humans use to signal that another person is being treated as safe and worth attending to.

Montoya put it flatly in the university's write-up: whether we engage in these behaviours has little or nothing to do with romantic desires, and the same set turns up when doctors interact with patients, when parents interact with their children, and when salespeople talk to customers.

So the honest translation of the whole literature on nonverbal signs is this. A woman who holds eye contact, laughs, leans into the conversation, stands close and talks to you at length is displaying, with reasonable reliability, that she likes you. Nothing in the measured behaviour distinguishes the kind of liking. The warm colleague, the friend who is glad to see you and the person who is attracted to you are running the same visible program.

That is not a gap in the research that better research will close. It is the finding. The reason telling flirting from friendliness feels impossible is that at the level of observable behaviour it substantially is.

The first item on that list is also the one most often read as decisive, and it is the one where the measured comfortable duration varies by culture and by neurotype — so its absence says even less than its presence does. What eye contact has actually been shown to predict goes through that evidence.

How often people actually get this right, measured

Hall, Xing and Brooks did the study the genre has never had to survive. Fifty-two single, heterosexual undergraduates of each sex were paired into 52 opposite-sex dyads, seated three feet apart, given a set of conversation-starter cards and left to talk for ten to twelve minutes while being recorded. Then they were separated and each answered two questions without conferring: did I flirt, and did my partner flirt.

Twenty-five of the 104 said they had flirted — 14 women and 11 men.

When a partner was not flirting, people got it right 84% of the time.

When a woman was flirting, her male partner identified it in 5 of 14 cases: 36%. When a man was flirting, his female partner identified it in 2 of 11 cases: 18%. Combined, flirting was detected in 7 of 25 cases, 28%. In the mixed model, the odds of accurate detection fell by 99% when the partner was actually flirting.

The caveat has to be said in the same breath, because it is a real one. Those percentages rest on 25 flirting events and on 5 and 2 correct hits. The uncertainty around 36% and 18% is wide, and the difference between them should not be read as a clean sex difference — the authors could not separate observer sex from target sex, since every female judge was judging a man. The 84% sits on 66 cases and is far more solid. What survives is the asymmetry, not the decimal places: absence of flirting is easy to see, presence of flirting is close to invisible.

Their second study is the one that should change how you read any list of signs. Two hundred and sixty-one observers made 1,557 judgements about one-minute clips drawn from those same interactions. Overall accuracy was 57%. Sitting inside the conversation, with tone of voice and ten minutes of it, bought people essentially nothing over strangers watching a minute of silence-free video.

Women are read less accurately, and the error runs one way

Place, Todd, Penke and Asendorpf used a different design and landed somewhere compatible. Their stimuli came from the Berlin Speed Dating Study: 382 people, three-minute dates in booths, each writing down afterwards whether they wanted to see that person again. Fifty-four American undergraduates who spoke no German — so working from behaviour and prosody, not content — watched clips and guessed who had said yes.

Because daters said yes less than half the time, an observer who knew the base rate could score 52% on male interest and 56% on female interest by guessing. Against those baselines, predictions of male interest were far above chance, d = 2.94. Predictions of female interest were barely above it, d = 0.62.

The shortfall was not people throwing darts. The authors are specific: it was a systematic overperception of female daters' interest, and for the five hardest-to-read women more than 80% of the interest predictions were wrong — wrong in the direction of seeing a yes that was not there.

Three details from the same paper are worth having. Ten seconds of video did as well as thirty; length did not matter. Clips from the beginning of the date were the worst, and clips from the middle and end were the best, which suggests people are still deciding early on and there is genuinely less to read. And accuracy at reading the man and accuracy at reading the woman in the very same clip correlated at r = .00. Whatever skill is involved, it does not transfer.

The authors' explanation is the standard evolutionary one — that the asymmetric costs of a mating mistake push women to stay ambiguous longer while they evaluate. That is an interpretation laid over the result, not something the study measured, and it should be read as such. The result itself does not depend on it.

For the question that brings most people to a page like this, the finding is blunt. This is the harder direction to read, and the characteristic way of failing at it is to see interest that is not there. Male interest is the direction observers read well here, which is a fact about this experiment and not advice about who to watch. What the wider research does and does not support about the signs that someone is falling for you is a separate question, and one that rests on firmer ground than anything on this page.

Overperception: the claim, the challenge, and where it now stands

This is the oldest thread in the literature and it is worth walking through properly, because it is the point where the topic stops being a parlour game.

Abbey ran the original study in 1982 with 144 Northwestern undergraduates. A man and a woman held a five-minute conversation while a hidden man and woman watched. Male actors and male observers rated the female actor as more promiscuous and seductive than female actors and observers rated her.

Abbey's own conclusion is almost always dropped from the retelling. Men also rated the male actor in more sexualised terms. She read the whole pattern as a broader tendency to perceive social situations in sexual terms rather than a specific misreading of female friendliness — a wider and less flattering claim than the one usually attributed to her.

La France, Henningsen, Oates and Shaw later put the accumulated studies through three meta-analyses and found men rating targets higher than women do on flirtatiousness, seductiveness and promiscuousness, with the size of the gap moderated by the target's sex and by how the interaction was observed.

Then the challenge. Perilloux and Kurzban pointed out that nobody can measure a woman's true intent directly — all we ever have is her self-report — so the same data fit a second story: women understating their interest, and men reading it accurately. That is a serious objection, not a rhetorical one.

It was then tested. Murray, Murphy, von Hippel, Trivers and Haselton ran a preregistered study built to pull the two predictions apart, and concluded that men do overestimate. That is where the question sits today: the challenge was made properly and answered properly, and the original effect survived.

Which brings us to the part that is not academic. An overread signal does not stay inside your head. It lands on someone.

Ambiguity is not an invitation. A maybe is not a yes. A signal that requires interpretation to become a yes is not one, and if someone has to decline more than once, the declining is the answer. And where reading turns into monitoring — checking who she is with, tracking how fast she replies, needing to know where she is — that is not attention. That is control, and it has a name.

Mirroring was tested directly, and it did not hold

The single most repeated cue in this genre is that she is unconsciously mirroring your posture. Grammer, Kruck and Magnusson set out to test exactly that in opposite-sex encounters, and their abstract says what happened: they were unable to demonstrate a relation between synchronisation defined as movement echo or position mirroring and the subjective experience of pleasure and interest.

What they did find was something else entirely. Using a search algorithm built for the purpose, they identified what they call hierarchically patterned synchronisation: highly complex patterns of behaviour with a constant time structure, emerging when a woman was interested, specific to that pair and independent of the content of the behaviour.

Sit with that. The synchrony signal that survived their analysis was a temporal structure invisible without software. It is not something a person can watch for across a table, which is a problem for most of what gets written under Body Language, our own older piece included.

Against this, Montoya's meta-analysis does list mimicry among the behaviours significantly related to attraction, and the Dayton summary notes it held in Western samples. Both results can stand. A correlation that emerges across thousands of participants is a different object from a cue you can read off one person in one evening, and the gap between those two things is most of what this genre gets wrong.

There is also a direction-of-causation problem. Hove and Risen manipulated synchrony rather than observing it — having people tap in time with someone — and found that it increased affiliation. If moving in time with a person makes them like you, then noticing synchrony and concluding she already liked you is reading an effect as though it were a cause.

Four cues that were never cues

Pupil dilation first. Bradley, Miccoli, Escrig and Lang measured the pupil while people viewed emotional pictures and found it tracked emotional arousal and sympathetic activation — dilating for unpleasant images as well as pleasant ones. The pupil is an arousal and effort gauge, indifferent to whether the arousal is good news. It also responds far more strongly to ambient light than to anything psychological, which makes the classic dim-bar observation close to uninterpretable.

The eyebrow flash, second. It is a genuine cross-cultural signal, and what it signals is recognition: it is given on greeting, to people you know. Montoya's analysis puts eyebrow raises in the group with no evidence of a relation to liking. It tells you she recognised you.

Third, the claim that 93% of communication is nonverbal. It comes from two 1967 studies about how listeners resolve conflicting cues in a narrow laboratory task, stretched into a law of nature. We take that one apart in full in the piece on texting habits, because it does more damage there than anywhere else.

Fourth, and this one is ours. The version of this page that stood here until today treated a question she did not need to ask as a tell. The largest analysis of real dating conversation points the other way. McFarland, Jurafsky and Rawlings transcribed and analysed close to a thousand four-minute speed dates and found question-asking associated with feeling disconnected: women reported feeling less connected when they had to ask men questions, and when men asked them questions. Questions were what people reached for when a conversation was lagging. What did track a reported connection was appreciative language, expressions of sympathy, shared stories, varying loudness, interruptions that finished or extended a thought rather than redirecting it — and conversations centred on the woman rather than the man. That is self-reported clicking among Stanford graduate students and it is correlational, but it is a thousand real dates, and it does not support the item it replaced.

What the courtship-observation tradition actually established

It would be wrong to leave this sounding as though nothing is known. Moore's 1985 fieldwork is the foundation of the whole area and it found something solid — just not the thing it gets cited for.

She built a catalogue of 52 nonverbal solicitation behaviours from observing more than 200 women, then systematically observed 40 women across four settings: a singles bar, a university snack bar, a library, and Women's Center meetings. Display rates were far higher in the contexts where meeting someone was plausible. And the women who signalled most often were the women most often approached.

State precisely what that establishes, because it is usually inverted. It establishes that these displays do something — they function to solicit an approach, and that the initiative in a first approach is very often the woman's, with the man who believes he chose having in fact responded. It does not establish that any single item in a 52-behaviour repertoire can be decoded on its own, and Moore's own 2010 review reads as the history of a field that has spent forty years describing repertoires rather than validating a decoder.

Hall, Carter, Cody and Albright supply the other half of the problem. Their Flirting Styles Inventory distinguishes five measurably different ways people communicate romantic interest. Someone whose style is polite or sincere may produce almost none of the physical repertoire the listicles enumerate while being entirely interested. The absence of the signs is not evidence of the absence of interest — which is one of the quieter reasons people end up misfiled as friends for months.

What an ambiguous signal is for, and the only thing that resolves it

The last move is to stop treating the ambiguity as a defect.

It is doing work. Signals that can be denied protect both people: the one who sent it keeps the option of not having meant it, and the one who received it keeps the option of not having noticed. Hall's later work on how social-sexual communication gets interpreted makes the same point from the theory side — the identical behaviour carries different meanings depending on the frame the two people are already in. The vagueness is not noise around the message. Some of the time it is itself the message, and reading harder cannot remove it. Sustained on purpose over months it becomes something else, and the measured version of that turns out to be a gap in commitment rather than a missing label.

The only thing that resolves ambiguity is language. Asking once, plainly, in a form that is easy to decline. Not a line, not a moment engineered to be hard to refuse — the opposite of that. A question whose easiest possible answer is no is the only kind that produces information you can trust.

And the cost of asking is smaller than people predict. Joel, Teper and MacDonald found across their studies that people substantially overestimate how willing they would be to reject a potential partner: faced with a real person rather than a hypothetical one, they rejected far less than they had forecast, and the reason was concern for the other person's feelings. Two things follow, and both matter. The rejection you are bracing for is usually gentler than the one you imagine. And a gentle answer can still be a real no — the same finding that makes asking less frightening is the one that makes a soft decline easy to misfile as encouragement.

The other honest limit is how little any of this predicts. Joel, Eastwick and Finkel ran more than a hundred self-report measures on speed daters and used machine learning to predict who would want whom. They could predict how much a person tended to desire others, and how much others tended to desire them. What they could not predict at all, from any combination of variables gathered beforehand, was the part specific to a particular pair — whether these two would click.

Which leaves the one instrument you actually have: your own read of whether the conversation went somewhere. If that read is never available to you — if the loop is compulsive rather than curious, if you have been collecting evidence for months rather than asking a question in five seconds — the thing worth examining is the reader rather than the signals. Our attachment style test takes about two minutes and says plainly what a result like that cannot tell you, the difference between love and anxious attachment covers where the two come apart, and why capable people keep choosing badly is about the same machinery running upstream. More of this kind of thing under Signs.