How event matching algorithms work
Event matching scores every possible pair of guests by how closely their answers agree, removes pairs that break the night’s rules, and then chooses the set of pairs with the highest combined score, so the whole room does as well as it can instead of the first pair taking the best match.
Published Updated By the HowWeMet team
The short version
This page explains how HowWeMet matches a room, step by step, including the parts that are compromises. The same approach applies to most event matching: turn answers into a score for every pair, decide which pairs are allowed, then choose the pairs together.
- Pool. Take every guest who verified their phone and finished the questions.
- Rules. Remove pairs that are not allowed: people not looking for each other, colleagues on work nights, pairs the host kept apart, and partners from an earlier round.
- Scores. Score every remaining pair from 0 to 1 by how closely their answers agree.
- Odd rooms. If the room is odd, pick the best group of three, when one is allowed.
- Pairs. Choose the set of pairs with the highest combined score. If not everyone can be paired, place as many people as possible first.
- Explain. For each match, list the answers the pair shared strongly or saw very differently.
Scoring a pair from answers
Guests answer every question on a five-point scale, from 1 (disagree) to 5 (agree). For each question, the agreement between two guests is:
Per-question agreement
agreement = 1 minus (the gap between the two answers, divided by 4)
The same answer scores 1. One step apart scores 0.75, two steps 0.5, three steps 0.25, and opposite ends of the scale (1 and 5) score 0. Dividing by 4 keeps every result between 0 and 1, because 4 is the largest possible gap.
The pair score is the weighted average of those agreements across the questions. Most questions weigh 1. A few positions in each mode carry a small nudge up or down. The weight belongs to the position in the list, not to the wording, so it applies to whatever question the host puts in that slot:
| Mode | Heavier questions | Lighter questions |
|---|---|---|
| Romantic | 1st and 4th weigh 1.25, 2nd weighs 1.1 | None |
| Platonic | 1st and 6th weigh 1.15 | None |
| Professional | 1st, 2nd and 4th weigh 1.2 | 3rd weighs 0.9 |
A score of 0.83 means the pair agreed closely across the questions, weighted as above. Guests see it rounded to a percentage. It is a measure of how similarly two people answered, not a prediction that they will get along.
Who is allowed to be matched
Before any pair is chosen, pairs that break a rule are removed completely. A removed pair cannot be picked, however high its score would have been.
Romantic nights: both people must be seeking each other
On romantic nights, guests say their gender and who they are seeking: women, men or anyone. A pair is allowed only if each person is seeking the other. Someone seeking women and a woman seeking anyone can be matched. A woman seeking men and a woman seeking women cannot, because the first is not seeking the second.
Work nights: colleagues are kept apart
On professional nights, guests can give their company, and two people from the same company are never matched. Names are compared ignoring capital letters and extra spaces, so “Acme” and “ acme ” count as the same company. A blank company never counts as a match. On Elite nights, hosts can switch this ban on for any mode.
Pairs the host keeps apart
On Pro and Elite nights, hosts can mark specific pairs that should never be matched, such as exes or a manager and their report.
Previous partners across rounds
On multi-round Elite nights, anyone who was matched together in an earlier round is banned from being matched again. A group of three bans all three of its pairs. Each new round re-solves the whole room with those extra bans, so nobody meets the same person twice.
Choosing the pairs together
The obvious approach is greedy: take the highest-scoring pair, remove those two people, and repeat. It is fast and it is often wrong, because an early pick can leave the people behind it with poor options.
Here is a small illustration with made-up scores. A and B score 0.90, A and C score 0.85, B and D score 0.80, and C and D score 0.30. Greedy takes A and B first (0.90), which forces C with D (0.30), for a total of 1.20. Pairing A with C and B with D instead totals 1.65. Nobody gets their single best score, but the room as a whole does much better.
So the solver looks for the set of pairs with the highest combined score, known as a maximum-weight matching. How it searches depends on the size of the room.
Rooms of 12 or fewer: exact
With 12 or fewer people in the pool, the solver checks every possible way to pair the room and keeps the best one. The answer is exact.
Larger rooms: a strong heuristic
The number of possible pairings explodes as rooms grow, so above 12 people the solver starts from the greedy answer and then repeatedly tries swapping partners between two pairs, and re-pairing any three pairs exactly, keeping every change that raises the total. This is fast and does well on the dense score tables that questionnaires produce, but it is not guaranteed to find the best possible total. We say so plainly: in a large room it can land slightly below the true best, and in a large, imbalanced room it can occasionally leave someone sitting out whom a perfect solver would have paired.
Romantic nights: a stable matching pass
On romantic nights, the solver also runs a Gale-Shapley pass, the classic stable matching method, where people propose to their highest-scoring allowed partners in turn. If that pass pairs everyone, its result is used. Otherwise the highest-combined-score result stands.
Odd rooms, trios and sit-outs
Groups of three
When the pool has an odd number of people, one group of three can absorb the extra person. Every trio in which all three pairs are allowed is ranked by the sum of its three pair scores. For pools of up to 13 people, the solver tries the 25 best trios and keeps the one that lets the most people be placed overall, counting the best pairs it can make from everyone else. It also checks the option of using no trio at all, and only prefers that option if it places strictly more people. A tie goes to the trio. For larger pools it takes the single best trio.
A trio’s score is the average of its three pair scores. Because every pair inside it must be allowed, trios are routine on platonic and professional nights and rare on romantic nights, where all three people would have to be seeking both of the others.
Sit-outs
Sometimes not everyone can be paired: an imbalanced romantic room, many banned pairs, or an odd room with no allowed trio. The solver then chooses the pairing that places the most people, then the highest combined score among those, then a fixed tie-break. Anyone left over sits out that round. A guest sits out only when no allowed partner is available for them in the best arrangement.
The host sees who is sitting out, by first name, before the reveal. Those guests see “No match this round” on their phone at the reveal, not a forced pairing that breaks what they asked for.
Romantic overflow
On Pro and Elite romantic nights, hosts can turn on romantic overflow. If at least two people are left over, they are matched again among themselves as friends, using the same questions, bans and rules but without the seeking requirement. Those guests see “Matched as friends”.
A score floor for Elite nights
Elite hosts can set a maximum score gap: pairs that score more than that gap below the best pair in the room are dropped. The floor is only applied if everyone can still be paired without those pairs. Otherwise the solver ignores it rather than leave people out.
A worked example: 4 guests, 3 questions
The numbers below were produced by running the real matching code on these answers. You can check every one with a calculator.
The answers
A platonic night with three questions. On platonic nights the first question weighs 1.15 and the other two weigh 1, so the weights add up to 3.15.
| Guest | Q1 “I believe in platonic soulmates” (1.15) | Q2 “The group chat is sacred” (1) | Q3 “Karaoke is a personality test” (1) |
|---|---|---|---|
| Asha | 5 | 4 | 2 |
| Ben | 4 | 4 | 1 |
| Chen | 2 | 1 | 5 |
| Dev | 1 | 1 | 5 |
Scoring one pair: Asha and Ben
- Q1: answers 5 and 4, one step apart. Agreement 0.75, times weight 1.15, gives 0.8625.
- Q2: answers 4 and 4, the same. Agreement 1, times weight 1, gives 1.
- Q3: answers 2 and 1, one step apart. Agreement 0.75, times weight 1, gives 0.75.
- Total 2.6125, divided by the total weight 3.15, gives 0.8294, shown as 83%.
All six pairs
| Pair | Score | Shown as |
|---|---|---|
| Chen and Dev | 0.9087 | 91% |
| Asha and Ben | 0.8294 | 83% |
| Ben and Chen | 0.2619 | 26% |
| Asha and Chen | 0.2500 | 25% |
| Ben and Dev | 0.1706 | 17% |
| Asha and Dev | 0.1587 | 16% |
Choosing the pairs
Four people can be split into pairs in exactly three ways. The solver compares the totals:
| Pairing | Combined score |
|---|---|
| Asha with Ben, Chen with Dev | 1.7381 |
| Asha with Chen, Ben with Dev | 0.4206 |
| Asha with Dev, Ben with Chen | 0.4206 |
The first pairing wins by a wide margin, so Asha meets Ben (83%) and Chen meets Dev (91%). With four people this is an exact search.
What they see
The explanation comes from the answers too. An answer counts as shared when both people gave the same 4 or 5, and as a difference when their answers are 3 or more steps apart. Asha and Ben both answered 4 to “The group chat is sacred”. Chen and Dev both answered 5 to “Karaoke is a personality test”. Each pair sees that line under “You both agreed”. Had Asha been paired with Chen, all three questions would have shown up as differences.
A fifth guest joins
Add Esi, who answers 5, 3 and 2. Her best score is with Asha (0.9206), then Ben (0.7500). The room is now odd, and the best allowed trio is Asha, Ben and Esi, with pair scores 0.8294, 0.9206 and 0.7500. Their trio score is the average, 0.8333, shown as 83%. Chen and Dev stay together at 91%, and all five people are placed.
Round two
Back to the four guests, now in a second round. Asha and Ben, and Chen and Dev, are banned from meeting again. Only two pairings remain, and both total 0.4206. On a tie the solver falls back to its fixed internal order, and in this run Asha meets Chen (25%) and Ben meets Dev (17%). That is the honest trade-off of multiple rounds in a small room. Later rounds reach further across the room, so scores are usually lower.
Same answers, same pairs
The whole process is deterministic. Given the same guests, answers, questions and rules, it always returns the same matches. Guests are sorted into a fixed internal order before anything runs, and every tie, between two pairs, two pairings or two trios, is broken by that order rather than by chance or by who joined first. Shuffling the input changes nothing.
This matters for hosts. The pairs you preview are the pairs that will reveal, provided nothing changes in between, and if a guest asks why they were matched, there is a clear answer.
Why no AI picks the matches
No language model or trained model is involved in choosing matches. That is a deliberate choice.
- Every score can be explained. A score is arithmetic on answers the guest gave, which is why the explanation a guest sees is just a list of their own answers.
- It is repeatable. A model can give different answers to the same input. A host needs the preview to match the reveal.
- The rules are hard rules. Seeking preferences, colleague bans and kept-apart pairs are enforced before scoring, not requested politely from a model that might ignore them.
- It uses only the night’s answers. Matching reads the questionnaire answers and the rules above for that event, nothing else. Read more on how we handle guest data.
The limitation is the flip side of the strength: the matching only knows what the questions ask. Good questions matter more than anything in this page, and the guide to hosting a matchmaking party covers how to write them. To watch the process on sample guests, try the demo.
Questions
Does the matching use AI?
No. Scores are calculated from answers with a fixed formula, and pairs are chosen by a deterministic algorithm. The same answers always produce the same matches.
Why did I get a lower score than my friends?
Scores depend on how closely two people answered and on who else was in the room. The solver maximises the combined score of all pairs, so a slightly lower individual score can let several other people get a much better match.
What happens with an odd number of guests?
One group of three is formed when all three of its pairs are allowed, which is routine on platonic and professional nights. On romantic nights a trio is rare, so an odd room usually means one person sits out that round.
Is the matching exact?
For 12 or fewer people it is exact. For larger rooms it uses a strong heuristic that can land slightly below the best possible total.