How do dating sites decide who to show a member?
They filter first, then rank what is left, and some then learn from what each member does. "The algorithm" is usually these three layers working together, not one formula.
| Layer | How it works | What it needs | Where it goes wrong |
|---|---|---|---|
| Filters and search | Shows only people who fit stated preferences such as age, distance and what they are looking for | Complete profiles | Too many filters leave a small site with nobody to show |
| Ranking | Orders the results, often by recent activity, distance and profile quality | Activity data | The same few profiles get all the attention |
| Compatibility scoring | Compares answers to profile questions and gives a score | Members who answer the questions | People do not always choose what they say they want |
| Learning from behaviour | Uses likes, passes and replies to predict who two people will both like | Large amounts of behaviour data | Does little on a new site, and can narrow what members see |
Search and matching are two views of the same thing. Search lets the member drive. Matching (daily picks, a card deck, suggestions) lets the platform drive. Both draw on the same filters and ranking.
One rule holds across every design: a good match is two-sided. Showing a member someone they will like is half the job. The other person has to be likely to like them back, and to be active enough to reply.
How do Tinder and Hinge say their matching works?
Both publish the main signals they use, and both put activity and stated preferences ahead of any hidden score.
Tinder says its system starts with age, distance and gender preferences. Its help centre says it prioritises members who are "active, and active at the same time", calls regular use of the app "the most important part of our algorithm", and says "proximity is a key factor". It also uses interests and lifestyle details from profiles, and "anonymized cues from photos" to suggest profiles similar to ones a member has liked. Tinder says its old Elo score is "an outdated measure" that its technology no longer relies on, and that the algorithm "doesn't track social status, religion or ethnicity".
Hinge describes a recommendation system made of several algorithms. It uses stated preferences, including age, distance and relationship type, and members can mark some as dealbreakers. Hinge says it "will only show you each other if you both meet the dealbreakers that either of you set". Suggestions reflect the likelihood of mutual interest, and the system learns from likes and matches. Hinge says skipping a profile quickly may be read as a sign that similar people are not a fit. It says it introduced a deep-learning recommendation system in 2025.
Two lessons carry over to a small brand. Activity and distance do most of the work. And both companies explain their systems in plain words, which you can do too.
What is the Gale-Shapley algorithm?
It is a method for pairing two groups so that the result is stable, meaning no two people would both prefer each other to the partners they were given. Dating articles mention it because Hinge used it.
TechCrunch reported that Hinge's Most Compatible feature, which puts one suggested person at the top of a member's feed each day, uses the Gale-Shapley algorithm. The article says the method was created in 1962 by David Gale and Lloyd Shapley for what is called the stable marriage problem. Hinge said that in early tests members were 8 times more likely to go on dates with Most Compatible suggestions than with its other recommendations, measured by an exchange of phone numbers. That is a company figure from its own testing.
The idea is respected well beyond dating. The 2012 Nobel economics prize went to Lloyd Shapley and Alvin Roth for this work. The prize committee said the methods always ensure a stable matching, and that versions of the algorithm are used to match new doctors with hospitals, students with schools and organ donors with patients.
It needs a ranking of who each person prefers, so it works on top of a learning system. It is not where a new site starts.
What is the cold start problem in dating matching?
It is the gap at launch, when you have too few members to show and too little behaviour to learn from. No algorithm fixes an empty room.
There are two versions of the problem. A new site has few people in each town, so strict filters return nothing. A new member has no history, so a learning system knows nothing about them.
Practical answers:
- Launch in one place or one niche, so the members you have are near each other.
- Keep sign-up questions short, but ask the few that matter for filtering.
- Rank by recent activity, so new members see people who will reply.
- Widen distance or age gently when results run out, and tell the member you have done so.
- Give new members early visibility, so they get a first reply quickly.
A shared member network is the other answer. Dating Partners, our sister company run by the same group as this site, runs a shared network across its brands, so a partner's site has members from day one. See shared dating networks explained and our guide to the dating app cold start.
What rules apply to dating matching algorithms?
Two sets of rules matter most: transparency duties for recommender systems in the EU, and data protection law wherever you use personal data to match people.
EU Digital Services Act. Article 27 says online platforms that use recommender systems must set out the main parameters in their terms, in plain and intelligible language, along with any options members have to change them. Article 19 excludes micro and small enterprises from that section of the Act, with some exceptions, so check where you fall. Our DSA guide for dating platforms has more.
UK GDPR. Religious beliefs, ethnic origin, sexual orientation and data about a person's sex life are special category data. The ICO says inferred data counts too if you intend to infer one of these details or to treat someone differently because of it, however confident you are in the inference. A system that works out a member's religion or orientation from behaviour is covered, even if the member never stated it. See GDPR for dating sites.
This is general information, not legal advice.
How do I measure whether matching is working?
Measure replies and conversations, not likes sent. A system tuned only for swipes or time spent can look busy while members meet nobody.
- Mutual likes. The share of likes that are returned.
- Conversations with a reply. A first message that gets an answer is the clearest sign of a good suggestion.
- Time to first reply for new members. Slow first replies are an early warning that people will leave.
- Spread of attention. If a small group of profiles receives most likes, most members are being ignored.
- Empty results. How often a search or daily pick returns nobody.
Test changes on part of your membership before you roll them out, and compare both sides of each match. We publish no benchmark figures here because apps do not report these numbers in a comparable way.
What matching do I get on a white label platform?
The platform's own search and matching, which you inherit and cannot rewrite. You shape results through your niche and your profile questions, not through the code.
Providers describe their systems differently. HubPeople says its matching uses age range, relationship type, lifestyle habits, niche interests, location and "member recency and activity levels". DatingAPI, also run by our group, runs matching and discovery for apps, and its developer page lists search, a card deck and member lookup. Our member experience page shows how members find each other on our own brands.
Before you sign with any provider, ask:
- Which signals decide the order members are shown in?
- Do my own members see each other first, or the whole network?
- Can members hide or change the preferences used?
- Is any paid boost labelled?
- Is special category data used in ranking, and on what legal basis?
Questions
Does Tinder still use an Elo score?
Tinder says no. Its help centre calls Elo an outdated measure and says its technology no longer relies on it.
Do I need machine learning to launch a dating site?
No. Learning systems need a lot of likes, passes and replies before they help. Clear filters and sensible ranking by activity and distance are enough to start.
Can members pay to be shown to more people?
On many platforms, yes. HubPeople, for example, describes its Profile Boost as temporary priority visibility to active nearby members. If you sell a boost, say clearly what it does.
Can matching use religion or ethnicity?
Members can state preferences, and Hinge lists religion and ethnicity among its preference settings. Under UK GDPR this is special category data, so you need a lawful condition for using it. Take advice.
Should I tell members how matching works?
Yes. A short, plain page builds trust, and in the EU the Digital Services Act requires many platforms to set out the main parameters of their recommender systems.
Sources
- Tinder Help: Powering Tinder, the method behind our matching, checked 20 November 2026
- Hinge: How we connect daters, checked 20 November 2026
- TechCrunch: Hinge employs new algorithm to find your 'most compatible' match, checked 20 November 2026
- Nobel Prize: press release, the Prize in Economic Sciences 2012, checked 20 November 2026
- Digital Services Act, Article 27: recommender system transparency (text as reproduced by Springlex), checked 20 November 2026
- Digital Services Act, Article 19: exclusion for micro and small enterprises (text as reproduced by Springlex), checked 20 November 2026
- ICO: What is special category data?, checked 20 November 2026
- HubPeople features, checked 20 November 2026
- DatingAPI developers, checked 20 November 2026
- Dating Partners partner terms, checked 20 November 2026
What changed
- : First published. Replaces our older guide on dating search and matching algorithms.