How do you run matchmaking at a conference?

What the problem actually is

Attendees cannot tell who is worth their time. The badge says a name and a company; it does not say that the person two metres away is buying exactly what you sell. Matchmaking is an information problem before it is a software problem.

There are two failure modes, and they need opposite fixes. In the first, almost nobody meets anybody new — the tool is there, the profiles are empty, and the directory is a list of names. In the second, matching works well enough that the same twenty people receive every request, and everyone else concludes the event was not for them.

Three ways to match, and what each one costs you

Most conferences want the second, get the first by default, and promise the third to sponsors. Decide which one you are actually running before you write it into the exhibitor prospectus.

“AI matchmaking” — ask what the input is

Matching two people is a comparison between what one offers and what another is looking for. When both are chosen from a list you defined, that comparison is a rule, whatever the feature is called. The label on the button does not change the mechanism underneath it.

This matters for a practical reason rather than a philosophical one: a rule can be explained, checked and corrected, and an unexplained ranking cannot. Ask any vendor one question — what are the inputs, and can an attendee see why they were suggested? If the answer is a topic list you control and a reason on every card, you are looking at rules. If nobody can say why a suggestion appeared, you cannot defend it to the attendee who got a bad one, and you cannot fix it.

Language models are genuinely useful around the edges of this — tidying a profile someone typed in a hurry, or mapping a free-text job title onto your role list. That is preparing the input. It is not the matching.

The input that decides everything: one topic list

If you change one thing about how you run matchmaking, change this. Matching works when both people answer from the same vocabulary, and collapses when they type their own.

One list, asked twice

The same list answers both “what I offer” and “what I am looking for”. Two lists drift apart and stop lining up.

Short enough to read to the end

Long enough to separate two companies in the same category, short enough that nobody gives up scrolling. Write it for your audience, not for an industry taxonomy.

Written in their words

Use what attendees call things, not what your CRM calls them. A list nobody recognises gets answered at random.

Fixed once registration opens

Adding topics later splits the answers: early registrants never saw the new ones, so they can never match on them.

Where to ask is a separate decision. The registration form is the only moment you have everyone's attention, and it is also the form you are trying to keep short — the guide on form design below has the rule for deciding which questions earn a place there and which belong in a follow-up.

Say which roles should meet

Shared topics alone will happily suggest two suppliers to each other. Both sell the same thing, both tagged the same topic, neither wants the meeting.

• Name the roles at your event in the words your audience uses — buyer and supplier, investor and founder, clinician and researcher, whatever the room actually contains.
• Write the pairs that should meet, explicitly. A pair you did not write down is a pair the tool has no reason to prefer.
• Exclude colleagues. Two people from the same company do not need an introduction, and a suggestion list full of them is the fastest way to lose an attendee's trust.
• Allow a role to meet itself where it is true. Two clinicians often do want to meet; two resellers with the same territory do not. This is a judgement about your event, not a default.

Consent is part of the mechanism

A directory is a list of named people at an identifiable event, which is personal data before anyone exchanges a message. Three decisions carry most of the weight:

• Listed by choice, not by default. Opt-in produces a smaller directory of people who want to be found, which is also a better directory. Listing everyone automatically produces complaints and a worse signal.
• Contact details move only on acceptance. A card in a directory can show a name, a role and a company without showing an email address. The exchange is the thing the attendee agreed to.
• Hidden means hidden both ways. Someone who stays out of the directory should not be able to browse it either — otherwise opting out is a penalty and people opt back in for the wrong reason.

Say at the point of collection who will see each answer, and keep networking answers out of exports that go to sponsors. Whether your lawful basis is consent or something else is a question for your own counsel — this guide will not answer it.

A connection is not a meeting

Accepting a request produces two people who have agreed in principle, which is where most networking tools stop and most introductions die. A meeting has a time, a length and a place.

A time both are free

Availability should come from the sessions each person starred, not from a second calendar nobody fills in. Show that a slot is free without showing what is in the other one.

A length decided in advance

Short and fixed. Meetings without an agreed end overrun into the next session and both people blame the event.

A named place

“Somewhere in Hall 2” is how people fail to find each other. Name the tables, and count how many you have before you promise the feature.

A graceful no

Declining has to be as easy as accepting, with an option to suggest another time. Otherwise people stop answering, which looks identical to a broken feature.

The load problem nobody plans for

Good matching concentrates. The best-known speaker, the biggest buyer and the most obvious supplier will receive far more requests than they can accept, and everyone whose requests went unanswered will report that matchmaking did not work.

• Cap outgoing requests per person. A limit forces people to choose, and a chosen request is answered more often than a broadcast one.
• Rank by fit, not by popularity. If the suggestion list is effectively a leaderboard, it makes the concentration worse every hour.
• Refresh the suggestions. Someone who opens the list on the second morning should not see the same six people they already asked.
• Tell the people who will be swamped. Speakers and headline exhibitors can set aside a block, or turn requests off, if you warn them before the doors open rather than after.

Measure coverage, not volume

A headline number of meetings is easy to report and says almost nothing. Three ratios tell you whether it worked, and the third is the honest one.

High volume with low coverage means a small group met everyone — the second failure mode, dressed up as a good result. Low acceptance with high participation usually means the topic list is too coarse: people are being suggested to each other for a reason that is true but not interesting.

Ask the same questions of every vendor, ours included: what the inputs are, whether every suggestion carries a reason, when contact details change hands, and what an attendee sees if they choose not to be listed.