The Swipe Report
What wins a swipe, and what barely matters.
Abstract
We scored 17 million real swipe decisions and asked which parts of a profile changed whether it was swiped right. Three results are reported in full below: in our sample the median profile was decided in under two seconds, profiles with one photo were right-swiped about 20% less often than profiles with four or five, and smiling in the first photo ran about 9% below not smiling for both genders. The rest of the report covers the bio, the right-swipe base rates, the full photo-count curve, one real profile scored, and the limits of the design. We fit the model on 80% of the sample and tested it on the remaining 20%; anything that did not hold on the hold-out was discarded, which is the only reason we are willing to state the three results above as plainly as we do.
The dataset
One dataset: 17 million real swipe decisions, anonymized swipe outcomes from a live US dating platform under a data-use agreement. We don't name the platform, and we publish only aggregates.
The unit of analysis is the profile by viewer outcome: one profile, shown to one person, swiped one way. A profile that was shown many times contributes many outcomes, which is what makes a per-profile right-swipe rate possible at all. An app export cannot produce one, because it knows how often its owner swiped right and not how often its owner's profile was swiped right on.
Photo and text features are extracted on the platform side; we receive only de-identified signals (photo counts, photo attributes, bio length) and swipe outcomes.
So the signals in this report are counts and attributes rather than content. We can tell you what happened to profiles with one photo. We cannot tell you what was in the photo, and we would not publish it if we could.
How we measured
The discipline comes before the results because it is what constrains them.
Model fit on 80% of the sample, tested on the remaining 20% hold-out. Findings that did not hold on the hold-out set were discarded.
Rate figures are computed over profiles with at least 40 impressions, so a profile that was barely shown cannot move a bucket. That criterion is an inclusion rule, not fine print, and it appears in the figure captions where it applies.
Decision time measured as the interval between consecutive swipes within a session, excluding gaps over two minutes; slightly overestimates viewing time since it includes load and animation.
Decision time is not attention. It is the gap between one swipe and the next, and it includes the time the app spent loading and animating, so it reads slightly long and should be read that way. We report it as the outer edge of how long a profile was looked at, not as a measurement of looking.
The full method, including what we do not publish and why, is on the methodology page.
Finding: the decision is fast
In the 17 million swipes we scored: The median profile gets under 2 seconds before the swipe. About two-thirds of swipe decisions happen within 3 seconds.
This is the number that governs the rest of the report. Whatever else a profile is doing, it is doing it inside that window, and a fix that only pays off on the second read does not get a second read.
Schematic. We have not published the distribution of decision times, so this figure marks the two values we did publish and draws nothing between them. Decision time is measured as the interval between consecutive swipes and includes load and animation, so it reads slightly long.
Finding: photo count moves the rate
Across the profiles in our sample: Profiles with a single photo get ~20% fewer right-swipes than profiles with four or five. The gap holds at every age we checked.
The headline is the gap between one photo and several. The shape of the curve, including where it stops paying, is Figure 2 in section 08.
Finding: smiling did not help
In our sample: Smiling in your first photo runs about 9% below not smiling - both genders. The most repeated profile advice does not survive the data.
The axis is a difference, not a right-swipe rate: we have not published the absolute rates behind it. Zero is the not-smiling baseline in our own sample, and the bar is the approximate gap we measured, for both genders.
The counter-argument, first
This is the finding most likely to be argued with, so here is the argument before someone else makes it.
We are not adjudicating that dispute and our result does not settle it. PhotoFeeler scored photos on attractiveness as judged by a rating panel. We scored whether a real person swiped right. Those are two different measurements and both can be correct at once: people can rate a smiling photo as more attractive and still swipe past it.
What we are claiming is bounded to our own data, and the approved wording above says it exactly: the most repeated profile advice does not survive the data. It does not say smiling is bad, and neither do we. It says that in the sample we scored, the advice did not pay.
The full report
Get the full report.
Drop your email and the rest unlocks right here: the right-swipe base rates, the photo-count curve, the bio finding, and a real scored profile.
We'll email occasional SwipeScore updates. Unsubscribe anytime.