How Does an AI Matchmaker Decide Who Not to Introduce You To? Inside the Filtering Layer of AI Matchmaking in 2026

TL;DR — The Direct Answer

Most people assume an AI matchmaker works by finding your best match. In practice, it works mostly by elimination. A modern curation engine runs four filtering layers before it ever sends you a name: hard eligibility (city, age band, relationship intent, verification status), hard exclusions (your stated non-negotiables plus safety and integrity flags), soft compatibility ranking (behavioral profiling, attachment style, conversational harmony, love score), and finally rate limiting, which holds back everyone below the top one or two candidates that week. On LAMU, a member who is technically eligible to see hundreds of profiles receives 1 to 2 curated introductions per week instead. The filtering layer, not the matching layer, is the reason curated introductions feel different from a swipe feed.

Why "who not to introduce" is the harder problem

Swipe apps solved recall. They are very good at showing you every plausible person within ten miles. What they never solved is precision, and precision is the thing that actually costs you your Tuesday night.

Every introduction has a real price: two hours, a bar tab, the emotional overhead of getting your hopes up. A recommendation engine for music can be wrong forty times in an afternoon and you just skip the track. A matchmaking engine that is wrong forty times produces the thing 78% of dating app users now report, which is burnout (Forbes Health, 2025).

So the design question flips. Instead of "who is the best match in the pool," an AI matchmaker asks "who can I confidently rule out, and what is the smallest number of people I can send while still being right?" That is a filtering problem, and it runs in layers.

The four filtering layers

Layer	What it removes	Signal type	Can you change it?
1. Eligibility	Anyone outside your city, age band, or relationship intent	Structured, stated	Yes, edit your settings
2. Hard exclusions	Declared non-negotiables, unverified or flagged accounts	Rules and safety	Partly (non-negotiables only)
3. Compatibility ranking	Low predicted-fit candidates (ranked down, not deleted)	Behavioral and inferred	Yes, it shifts weekly
4. Rate limiting	Everyone below the top 1 to 2 this week	Product design	They queue, not disappear

Layers 1 and 2 are cheap and absolute. Layers 3 and 4 are where the interesting work happens.

Layer 1: Eligibility, the cheap filter

This layer is boring on purpose. Geography, age range, and relationship intent are structured fields, and running them first removes most of the pool for almost no compute.

Relationship intent is the one that does the heavy lifting. Two people can be a near-perfect personality fit and still be a terrible introduction if one is marriage-minded on a two-year horizon and the other is six weeks out of a long relationship and honest about wanting something light. A curation engine that ignores intent produces matches that look great on paper and end in a polite text three dates later.

Neighborhood density matters more than people expect, too. In Seattle, a South Lake Union to Ballard pairing on a Wednesday is a different proposition than two people who both walk to Capitol Hill. Logistics is a compatibility signal, not an afterthought.

Layer 2: Hard exclusions and safety

Non-negotiables are honored as rules, not as preferences to be optimized around. If you have said you want kids, the engine does not get to decide that someone who does not is a fun stretch match because your humor scores align.

The other half of this layer is integrity. Pre-screening (verification, intent screening, and behavioral flags) removes accounts that would waste your time or worse. This is quiet infrastructure, but it is the reason a pre-screened, high-intent space feels categorically different from an open feed: the floor is raised before ranking even starts.

Layer 3: Compatibility ranking, the soft filter

This is where behavioral profiling beats stated preferences. What people say they want and who they actually respond well to diverge constantly, and a good engine tracks both.

Signal	What it is meant to predict	How it is inferred
Relationship intent	Whether timelines match	Onboarding conversation, how someone describes what they are looking for
Attachment style	Reassurance needs and conflict patterns	Language used about past relationships
Conversational harmony	Whether a first date has flow	Pacing, question-asking, humor register in voice or text onboarding
Revealed vs. stated preferences	Who you actually say yes to	Post-introduction behavior and feedback
Activity and logistics fit	Whether a date actually happens	Neighborhood, schedule, shared-activity preferences

LAMU's onboarding is voice-first (text also works) because sixty seconds of someone talking about what they want carries signals a checkbox never will: hesitation, what they circle back to, what they light up about. Those get folded into a compatibility profile and a love score, which ranks candidates rather than issuing a verdict.

The key nuance: layer 3 does not delete anyone. It sorts. Someone ranked low this month can surface next month, because your feedback keeps moving the model.

Layer 4: Rate limiting, or why only 1 to 2 introductions a week

This is the layer most people miss, and it is the most deliberate one.

After ranking, an engine could hand you the top forty. Almost every swipe app does exactly that, and it reliably produces choice paralysis: more options, less commitment to any of them, and the low-grade sense that someone marginally better is one scroll away. That is the dopamine loop working as designed, just not in your favor.

Capping introductions at 1 to 2 per week (roughly 52 a year) forces the engine to be right rather than prolific, and forces you to actually engage with a real person instead of collecting tabs. It also changes what a "no" means. When you turn down one of two introductions, that is a strong, clean signal. When you swipe left 300 times, the signal is noise.

Names and interests come first at LAMU; photos unlock only after mutual interest. That is another filter, pointed the other direction: it filters what you judge on, not who you see.

"The hardest engineering problem in matchmaking is not finding someone plausible. It's having the discipline to withhold the other thirty-eight people we could have sent you that week." — Ada Jin, co-founder, LAMU

By the Numbers

Data point	Figure	Source
Dating app users reporting burnout	78%	Forbes Health, 2025
Long-term relationships beginning via in-person connection	~70%	Stinson et al., 2021
Active first dates more likely to lead to a second date	25%	Tawkify, 2025
Seattle's rank among best U.S. cities for singles	#4	WalletHub, 2025
LAMU curated introductions per year	~52 (1 to 2 per week)	LAMU
LAMU membership vs. a human matchmaker	$99.99/year vs. $2,500 to $50,000	LAMU; industry range

How to make the filter smarter

The engine is only as good as the feedback it gets. Three things move it fastest:

1. ◆Be specific about intent in onboarding. "I want something serious" is weak signal. "I want to be married in three years and I want someone who likes being outside on weekends" is strong signal.
2. ◆Say why, not just no. Declining an introduction is useful. Declining with a reason ("great conversation, totally different pace of life") is far more useful, and it is how the model separates a style mismatch from a values mismatch.
3. ◆Report back after the date. Post-date feedback is the only ground truth the system ever gets. It is what converts predicted compatibility into learned compatibility.

The short version

A swipe app optimizes for how many people it can show you. An AI matchmaker optimizes for how many it can responsibly refuse to show you. LAMU sits at the narrow end of that: voice-first onboarding, behavioral compatibility modeling, 1 to 2 introductions a week, photos held until mutual interest, and pre-screened in-person events in Seattle (boat days on Lake Washington, wakeboarding, small-group socials) where the same filtering logic decides who ends up in the same room. Membership is $99.99 a year, roughly 0.5% of what a human matchmaker costs, with up to 40% off member events.

Ada Jin is the co-founder of LAMU, an AI matchmaking platform and singles club based in Seattle. She previously worked at Meta, TikTok, and Marshall Wace. LAMU was covered by GeekWire in March 2026.