The software that decides which profiles reach your screen belongs to the same family that recommends films and shopping items. It is a ranking engine. Its objective is a predicted response rate, and every design decision inside it serves that number.
That objective explains most of what users find puzzling about their queues. Profiles arrive in an order somebody else determined. Attractive accounts appear and then stop appearing. A well-written profile gets less traffic than a thin one belonging to someone who logs in twice a day. None of this is arbitrary, and none of it requires a theory of romance to explain.
Desirability Scores and Ranked Order
Early swipe platforms borrowed the Elo rating system from competitive chess. Each account received a hidden score that rose when sought-after users swiped right on it and fell when they swiped left. The weighting worked the way chess ratings do. A right swipe from a highly rated account moved the number further than a swipe from a low-rated one, the same way beating a grandmaster counts for more than beating a club player.
Several companies have since distanced themselves from the Elo label. Ranked ordering remains in place under other descriptions, because a recommendation system has to sort its inventory somehow. The practical consequence for any user is that queue position gets decided before the app opens, and it gets decided by the aggregated swiping behavior of strangers.
Collaborative Filtering
The second layer is collaborative filtering, the same technique that powers retail and streaming recommendations. It groups you with users whose past choices resemble yours and then shows you people that group responded to. No description of your preferences is required. The model infers them.
Every swipe is training data, which produces an effect most people notice without naming it. The feed converges. Early choices teach the model a type, the model returns variations on that type, and those variations generate more of the same training signal. Researchers examining this feedback loop on a large swipe platform found it pushed men and women toward opposite and increasingly extreme approaches, with men liking a large share of the profiles they saw and women becoming steadily more selective. A user who spent their first week swiping right on one kind of face will find the pool has reorganized itself around that face by week four.
Platform Type and Pool Composition
Matching logic changes with the size and makeup of the pool. A general-population app sorts millions of accounts with mixed intentions, so the model spends most of its effort guessing what each user actually wants.
Services organized around a stated purpose behave differently for that reason. A hobby-based service, a faith-based service, or a sugar daddy dating website filters on intent at signup, which leaves the ranking system a narrower job to do and a smaller pool to do it in.
The Gale-Shapley Method
Some compatibility features are built on the stable matching solution David Gale and Lloyd Shapley published in 1962. The original problem was an administrative one. Given two sets of people with ranked preferences, produce pairings in which no two participants would both rather have each other than the partners they were assigned.
Applied to dating, the method gets adapted so everyone is sorted from one shared pool, with machine learning layered on top to estimate what a user prefers from their profile answers and past likes, then pair them with someone whose estimated preferences point back. The property that matters is mutual likelihood. That is why a compatibility suggestion is frequently someone a user would have scrolled past on an open grid.
Aspirational Pursuit and the Response Gap
The largest study of its kind analyzed messaging data from four American cities and found a consistent hierarchy of desirability across all of them. Both men and women contacted people who were on average about 25% more desirable than themselves. The probability of getting a reply dropped steeply as the gap between two users widened.
This creates a permanent tension inside the product. Users reach upward, the ranking system tries to pair them sideways, and the number the system optimizes for is the one users are working against. Suggestions that feel underwhelming are often the model correcting for a pattern the user cannot see in their own behavior.
Activity and Recency Weighting
Login frequency is one of the heaviest inputs in the whole system. A dormant account damages predicted response rate, so it gets pushed down the queue regardless of how good the profile is. Recent activity does the reverse.
The weighting differs across product types. Swipe-based and algorithm-based services handle the same signal in different proportions, since one is sorting a rapid stream of yes-or-no decisions and the other is scoring answers to a questionnaire. Both of them still favor the account that opens the app daily. Two accounts with identical photos and identical answers will receive different volumes of traffic if one logs in every morning and the other logs in every third week.
The Limits of Prediction
A 2017 study in Psychological Science put the whole premise to a direct test. Researchers collected more than 100 self-report measures from participants about traits and preferences relevant to choosing a partner, then ran a machine learning model on that data before anyone met, followed by four-minute speed dates.
The model succeeded at two things. It predicted how much desire a given person would feel in general, and how much desire they would attract in general. It failed at the third question, which was which two specific people would click. One author described romantic desire as closer to an earthquake than a chemical reaction.
That result sets the ceiling on what any matching system can deliver. The prediction is possible at the population level and unavailable at the level of the individual pairing.
The Ranking in Plain Terms
A suggestion from a dating app is an estimate of the probability that a message between two accounts will receive a reply, computed from swipe history, profile responses, and login frequency. Compatibility and chemistry are outside what the system measures or claims to measure. A user who reads the queue as a filtered list of people who might answer will get more out of it, and will spend a good deal less time arguing with the order it arrives in.