A match suggestion is a prediction. The platform is guessing which profile you will respond to, based on the record of what you and everyone resembling you have responded to before. The filters you set at signup are a small input into that guess and rarely the decisive one. Most of it comes from behaviour you never declared.
Stated Preferences and Revealed Behaviour
Every platform collects 2 kinds of preference. The stated kind is the age band, the distance radius, the height range and the boxes ticked during setup. The revealed kind is who you actually open, who you message, how long you look before deciding, and who you answer within the hour.
The 2 sets disagree more often than users expect. Somebody can set an age ceiling of 35 and keep opening profiles above it. The system notices the second pattern and reweights towards it, because a click is a stronger signal than a form field. This is why the results can drift away from the settings screen without anybody touching the settings. It also explains a common frustration. Widening a filter often changes very little, because the filter was never the binding constraint on what the queue contained.
Collaborative Filtering in Plain Terms
The core technique is borrowed from music and film recommendation. Suppose you and another user have reacted the same way to 40 of the same profiles. The system treats that other user’s 41st reaction as a reasonable guess at yours, and shows you the profile they liked.
Scaled to millions of accounts, this produces suggestions that no filter could have described.
It also has the flaw the method is known for. If the historical pattern encodes a bias about age, body type or ethnicity, the recommendation repeats the bias and calls it taste. Nothing in the method distinguishes a preference from a prejudice.
Distance Filters and Pool Size
The distance setting is the control that reshapes a queue fastest, and most users leave it where they set it on the first day. A 5 km radius in a dense city produces a large pool. The same radius in a market town produces a pool small enough that the system begins repeating profiles within a fortnight, and it then starts showing accounts that have been inactive for months. Widening the radius restarts the supply and changes the composition at the same time, because somebody 30 km away is on a different commute and a different set of assumptions about what counts as a local night out.
Niche and Special-Interest Platforms
Preference filtering also explains why the market has split into so many categories. A general platform has to serve everybody, so its filters stay broad and its predictions are spread thin across a very mixed pool. A category platform starts with the main filter already applied. A vegan singles service, a platform for walkers, a faith-based service and a sugar daddy website all begin from a narrower pool, which changes the job the matching has to do.
The cost is pool size. A smaller pool means fewer historical interactions to learn from, so category platforms lean harder on stated criteria and lighter on behavioural inference. Users notice this as a system that follows their filters more literally and surprises them less often.
Ranking, Photos and the First Screen
Almost every platform sorts a queue before showing it. The first image is weighted heavily because it is what most decisions are made on, and platforms measure how long an image holds attention before a swipe. Profiles whose first photo produces fast negative decisions get shown less. Their exposure drops regardless of anything written below. The effect compounds. A profile seen by fewer people gathers fewer positive signals, and fewer positive signals push it further down the queue.
Nielsen Norman Group’s usability work on recommended content found that people engage far more with suggestions presented in labelled, explained groups than with an undifferentiated feed. Very few matching platforms do this. The queue arrives with no reason attached, which leaves users guessing about why a particular profile appeared.
Mutual Interest and Matching Theory
Recommendation for shopping only needs to satisfy one party. Matching people needs to satisfy 2, which is a harder problem and an old one. Mathematicians have studied the stable marriage problem since David Gale and Lloyd Shapley published their deferred acceptance procedure in 1962, and the result won Shapley a share of the 2012 Nobel Memorial Prize in Economics.
A platform therefore has to estimate 2 probabilities at once, your interest in the person and that person’s interest in you, because an unreciprocated suggestion wastes 2 people’s attention. Getting one of the 2 right is worth almost nothing. Systems built this way suppress profiles that are far outside your predicted reach, which is the mechanism behind the common complaint that the pool feels narrower than the population.
Feedback Loops and Narrowing Results
Each round of predictions changes the data the next round learns from. Open a run of profiles with a similar look and the queue tightens around that look within days. The system has no way to tell a durable preference from a bored Tuesday evening. A fortnight of idle browsing can reshape a queue that then takes weeks of deliberate opposite behaviour to reset, because the older signals are still in the record and still being counted.
Pew Research Center’s survey of 6,034 people found that 3 in 10 US adults have used a dating site or app at some point, rising to 53% of those aged 18 to 29. At that scale the narrowing reaches roughly a third of the adult population and more than half of everybody under 30.
The Ceiling on Prediction
The ceiling on all of this is lower than the marketing suggests. A team led by Samantha Joel used machine learning on more than 100 self-reported traits and preferences from over 350 speed daters and still could not forecast love. The models predicted how generally desirable a person was, and how open to others they were, but not who would want whom after a 4-minute conversation.
The result marks a boundary. Preference data is good at removing obvious mismatches and bad at identifying the one person in 100 who will hold your attention across a dinner. A platform can tell you who is plausible, and that is a different question from who is interesting.
A Plain Definition of a Match Score
A match score is a probability estimate assembled from 3 inputs. It uses what you said you wanted, what your behaviour says you want, and what people with similar behaviour did next. No platform holds the data required to measure compatibility. The score works well enough to build a shortlist from, and badly as anything more final than that.
For a user, the settings screen matters less than the swiping. If the queue has narrowed to something you no longer want, the fix is to change what you open rather than what you have declared. The system is watching the first of those far more closely than the second.
