In the first six months of 2026, US startups raised $412.7 billion, more than the whole of 2025 combined, and 87.5 per cent of that capital landed in rounds of $100 million or larger (PitchBook-NVCA Venture Monitor, Q2 2026). For a conference organiser, those numbers describe a queue: the investors every founder wants to meet are fewer, busier and more heavily requested than at any point in the last decade.
You see the result in your own inbox. A partner confirms for two conference days, lands with a diary of ten pre-booked founder meetings, and asks your team the question that decides next year’s attendance: why these ten? If the honest answer is “the algorithm chose them”, you have a problem no keynote can fix.
This guide explains what actually happens between a founder profile and a confirmed investor meeting. Not at the level of vendor brochures, but at the level of mechanics: the four families of matching logic, the data each one needs, the points where they fail, and the questions that expose a black box before you sign. The short version up front: meeting quality is decided by the rules and data underneath the algorithm, not by how sophisticated the scoring is. A pairing you cannot explain is a pairing you cannot defend.
What is an investor matching algorithm?
An investor matching algorithm is the logic an event platform uses to decide which founders and investors should meet. It scores every possible pairing on criteria such as sector, funding stage, cheque size and declared goals, applies the organiser’s meeting rules, and turns the highest-scoring pairs into a schedule of confirmed one-to-one meetings.
That is the whole definition, and it deliberately says logic rather than AI (artificial intelligence). Some of the best-performing setups contain no machine learning at all. If the concept itself is new to you, start with our organiser’s guide to what investor matchmaking is and come back here for the mechanics.
Why the mechanics matter more in 2026
The capital concentration from the opening paragraph has a direct operational consequence: attention concentrates with it. When three firms collect 48.1 per cent of all new fund capital, the partners of those firms become the most requested profiles at every conference they attend. Without deliberate matching logic, your event reproduces the imbalance instead of correcting it: a handful of funds drown in requests, everyone else waits.
Why matching logic is under pressure
Capital, and attention, keep concentrating
Structured matching is the organiser’s counterweight. Which brings us to the machinery itself.
The four families of matching logic
Vendors describe their algorithms in very different vocabulary, from “16 strategies” to “hundreds of data points”. Under the marketing, nearly every platform combines some of four approaches. There are four families of matching logic used at startup conferences: rule-based meeting matrices, weighted criteria scoring, behavioural matching (collaborative filtering) and semantic matching with machine learning.
1. Rule-based meeting matrix
The meeting matrix is the constraint layer. It defines which participant groups can see, search and request which other groups, and in which direction: founders can request investors, investors see only startups inside their declared thesis, service providers see neither unless invited. Academic work on event recommender systems confirms the value of this layer: a TU Wien study showed that meeting rules such as “founders meet investors, investors do not meet investors” drastically shrink the space of possible pairings before any scoring runs (Bugl, 2023).
Every match is explainable by pointing at a rule, which makes this family the audit-friendly default for programmes under compliance or board scrutiny. Its limit is ranking: within an allowed group, a pure matrix does not sort. Two hundred eligible startups are still two hundred. The matrix decides who can meet. It never guesses.
2. Weighted criteria scoring
Scoring adds the ranking. Each eligible pairing collects points for declared fit: sector overlap, stage match, cheque size inside the founder’s raise, geography, sometimes mandates such as ESG (environmental, social and governance) criteria. The weights are set by you or the platform. A Series A fintech founder scores high with a fintech-focused Series A fund and near zero with a biotech seed investor.
The mechanics resemble a spreadsheet more than a neural network, and that is the strength: you can print the formula and defend every ranking in it. The weakness sits in the inputs. If half of your investors tick every sector to keep their options open, every match looks plausible and none is. Scoring is only as honest as the profiles it reads.
3. Behavioural matching (collaborative filtering)
The third family watches what participants do rather than what they declare: whose profiles an investor opens, which requests get accepted, which get ignored. It then recommends along the pattern “investors similar to you also asked to meet this startup”. This is streaming-service logic applied to deal flow, and at large recurring events it works, because thousands of interactions teach the system what declared profiles hide.
It carries two structural problems. The first is cold start: at a first edition there is no behaviour to learn from, so recommendations are guesses until enough requests flow. The second is the feedback loop: behavioural systems amplify whoever is already popular, steering even more requests towards the five most visible funds, which is precisely the concentration you were trying to correct. If that is your pain point, our playbook on preventing investor overload at demo days covers the distribution side in detail.
4. Semantic matching (embeddings and language models)
The newest family reads free text. It takes the founder’s product description and the investor’s thesis statement and maps both into a mathematical space where “supply chain logistics” sits close to “inventory management automation” even though the words never overlap. No checkbox taxonomy needed, which makes this approach strongest where profiles are rich, written and multilingual.
The trade-off is opacity. The relevance judgement lives inside a model that nobody can print. When a pairing goes wrong, the explanation on file is a similarity score, not a reason. Where AI layers sit well, and where they do not, is a topic we unpack in AI matchmaking at events: how it works.
The four families side by side
| Logic | Data it needs | Transparency | Cold-start risk | Best for |
|---|---|---|---|---|
| Meeting matrix | Group definitions, meeting rules | Full: every rule visible | None | First editions, compliance-heavy programmes |
| Criteria scoring | Structured profiles | High: weights printable | Low | Thesis-driven investor events |
| Behavioural | Interaction history | Low | High | Large recurring conferences |
| Semantic | Free-text profiles | Low to medium | Medium | Rich multilingual communities |
The data your algorithm actually needs
Every family lives or dies on its inputs, and the difference between a useful profile and a useless one is intent. A checkbox that says “interested in AI” cannot tell a buyer from a founder from a job seeker. The fields that move investor matching are more specific:
- Thesis and stage: which sectors, which funding stages, stated as constraints rather than interests. “B2B (business-to-business) SaaS, seed to Series A” is matchable; “innovation” is not.
- Cheque size and raise: an investor writing $500k tickets should not meet a founder raising $30 million, however similar their keywords.
- Geography and mandate: fund restrictions, regional focus, diversity or impact mandates that hard-filter the pool.
- Current intent: what each side wants from this event, this quarter: actively deploying, fundraising now, exploring a market.
Two practical rules follow. Keep the registration form short and demand precision only on these fields, because profile depth predicts match quality better than algorithm sophistication. And schedule a data-quality pass two weeks before matching opens: chase empty thesis fields the way you chase unpaid invoices. Good inputs are unglamorous. They are also most of the outcome.
From match score to meeting schedule
A ranked list of pairings is still not a meeting. The step that turns scores into a schedule is where event matchmaking differs most from dating apps and fundraising databases, and it has older academic roots than most vendor decks admit. The deferred-acceptance procedure that guarantees stable pairings, meaning no two participants would rather swap partners, was introduced by Gale and Shapley in 1962 and earned the 2012 Nobel Prize in economics for Roth and Shapley. Recent work shows it stays robust even when participants rank only a handful of preferences (McCauley et al., 2026), which matters because nobody at your conference will rank two hundred profiles.
Inside the pipeline
How a profile becomes a confirmed meeting
- Signals Profiles and intent enter Structured fields and declared goals from both sides form the raw material.
- Rules The meeting matrix filters Visibility and direction rules remove ineligible pairings before anything is ranked.
- Scoring Eligible pairs get ranked Weighted criteria, behaviour or semantics sort what the rules allow.
- Schedule Accepted matches become slots Mutual confirmations are placed into clash-free time slots, tables and video rooms.
The scheduling layer also carries the operational constraints that pure scoring ignores: investor quotas, no double-booking against pitch slots, table capacity, buffer times. This is where a matching engine becomes an event tool. How many meetings you should allow per investor is its own capacity question; our playbook on maximising investor meetings at a two-day conference works through the numbers.
Where matching algorithms fail
Four failure modes account for most disappointed post-event surveys, and none of them is fixed by a better model.
Cold start hits first editions that rely on behavioural logic: no history, no signal, random-feeling suggestions in week one. Thin profiles hollow out scoring: the algorithm ranks confidently on data that is not there. Overload concentration is the feedback-loop problem from family three: left uncapped, every ranking sends the crowd to the same five funds. And the perfection myth is a buying error rather than a technical one. Matching theory itself shows that no mechanism can be simultaneously stable and fully immune to strategic behaviour (Ravindranath et al., 2021). A vendor promising perfect matches is describing a mathematical impossibility.
The realistic goal is different: explainable logic, measurable acceptance and kept-meeting rates, and caps that spread attention. Imperfect and accountable beats optimal and opaque.
Can you explain the match? The 2026 compliance layer
Explainability used to be a taste question. In 2026 it is a legal one. Article 22 of the GDPR (General Data Protection Regulation) gives participants the right to demand human review of decisions made about them by automated systems, and the EU AI Act, now in full implementation, raises the documentation bar for systems whose decisions carry material consequences, such as which startups get in front of which investors. The practical consequence for organisers: your platform needs an audit trail and an override path, not just good suggestions.
The partner from the opening returns here. When she asks “why these ten founders?”, an answer like “your declared thesis, stage filter, mutual confirmation, rule four” ends the conversation with trust intact. A similarity score does not.
Solution: Converve runs matching on a rule-based meeting matrix that you configure per participant group, with scoring assistance on top. Every pairing traces back to a visible rule plus a mutual confirmation, which keeps the programme auditable under Article 22 and defensible in front of your most demanding attendees. You can see how the B2B matchmaking module works or how investor conferences combine it with pitch sessions on our startup and investor events page.
Seven questions to ask your platform vendor
Use these in the demo, in this order. The answers separate transparent systems from black boxes faster than any feature list.
- Why did participant A meet participant B at your last event, and can you name the rules and profile fields involved?
- What does your algorithm do at a first edition, with no behavioural history to learn from?
- How do you protect the most requested attendees: request caps, quotas, tiered visibility?
- Which profile fields drive the match score, and can we re-weight them for our audience?
- Do both sides confirm before a meeting lands in the schedule, or does the system force-book?
- How do you support Article 22 GDPR and EU AI Act obligations: audit trail, human review, override?
- Which outcome metrics does the platform report out of the box: acceptance rate, kept-meeting rate, second-meeting conversion?
A vendor who answers all seven without reaching for the word “proprietary” is a vendor you can build a programme with.
Frequently asked questions
Do investor matching algorithms need AI?
No. A rule-based meeting matrix plus weighted criteria scoring produces auditable, high-acceptance matches without any machine learning. AI layers add value at scale and on rich free-text profiles, but they are an optional refinement on top of the rules, not a requirement.
What is a good match acceptance rate at an investor event?
Aim for a 40 to 60 per cent match acceptance rate, a kept-meeting rate above 80 per cent and a second-meeting conversion of 20 to 30 per cent (Converve event matchmaking FAQ, 2026). A platform that cannot report these numbers cannot prove its algorithm works; the full benchmark set sits in our event matchmaking FAQ.
Which matching logic is best for a first-time conference?
Rules plus criteria scoring. Both work from day one because they rely on declared profiles rather than interaction history. Purely behavioural systems need data your first edition does not have, so treat “the algorithm learns your community” as a year-two feature, not a launch feature.
How do algorithms prevent top investors from being flooded?
The algorithm alone does not; the constraint layer does. Request caps, meeting quotas per investor and tiered visibility rules keep the most wanted profiles answerable, and behavioural re-ranking should always run inside those limits rather than instead of them.
Conclusion: buy the logic you can defend
Investor matching is not magic, and the vendors who imply otherwise are usually hiding the mechanics this guide just walked through. Four families of logic, one data layer that feeds them, one scheduling layer that turns scores into meetings: that is the whole machine. The organisers who win with it are rarely the ones with the most sophisticated model. They are the ones who can answer “why these ten?” in one sentence, at every edition, to every partner who asks.
Choose rules you can print, data you actually collect, and an algorithm someone in your team can explain. Want to see what an explainable meeting matrix would look like for your conference? Get in touch with Converve and we will walk you through a live setup.
