Match Group's Preference Problem: Are Algorithms Chasing the Wrong Data?

The dating app industry has a stated-preference problem. Match Group and its competitors have spent years refining algorithms that serve up what users say they want, but new research suggests those stated preferences bear little relation to actual attraction. The gulf between what people claim to desire and what they respond to in person may undermine the entire foundation of algorithmic matching.

Match Group (MTCH) has spent the last three years optimising its platforms around stated preferences—algorithmically serving up what users say they want. A new study from UC Davis tracking 4,500 blind dates suggests that exercise might be fundamentally flawed. According to research published in the Proceedings of the National Academy of Sciences, both men and women demonstrated higher attraction to younger partners during in-person encounters, contradicting decades of survey data showing women claim to prefer older men.

The disconnect isn't subtle. Whilst demographic data consistently shows men partnering with younger women—and women routinely tell researchers they seek older partners—the blind date research reveals something different happens when people actually meet. Lead researcher Paul Eastwick told media outlets the finding would be 'shocking to many people', and he's not wrong. It upends the prevailing wisdom that youth preference is primarily a male trait.

Create a free account

Unlock unlimited access and get the weekly briefing delivered to your inbox.

Register free

No spam. No password. We'll send a one-time link to confirm your email.

Why algorithms can't solve for what users won't admit

The implications for product development are uncomfortable. Dating platforms have collectively invested hundreds of millions into recommendation engines and compatibility scoring, almost all of it built on user-supplied preference data. Bumble (BMBL) lets users filter by age range. Hinge prompts members to specify dealbreakers. Match's entire value proposition rests on compatibility matching derived from stated preferences.

You can't A/B test your way out of users not knowing what they actually want, or knowing but being unwilling to admit it.

If those stated preferences don't predict in-person attraction, the feature set isn't wrong—it's solving for the wrong variable. The age gap presents an especially thorny challenge because it intersects with trust and safety concerns.

Most major platforms already restrict age-gap matching at the extreme ends—Tinder, for instance, won't show profiles with more than a decade's difference if one party is under 27. But those guardrails were built partly on the assumption that large age gaps primarily benefit older men pursuing younger women. If women are also drawn to younger partners but self-censoring that preference in their profile settings, platforms may be inadvertently restricting matches that would have produced genuine attraction.

There's also a commercial tension. Subscription dating platforms generate revenue by keeping users engaged long enough to convert to paid tiers, but not so engaged that they're matching efficiently and churning out. A feature that helps users overcome their own preference blind spots—surfacing profiles they wouldn't have selected but might actually be attracted to—could theoretically improve match quality whilst reducing time-to-partnership. That's good for users, questionable for lifetime value metrics.

The profile problem gets worse with AI

The stated-preference gap will become more problematic as platforms lean harder into AI-driven matching. Grindr (GRND) recently disclosed it's testing AI tools to help users craft better profiles. Match Group has repeatedly flagged AI as a strategic priority, with CEO Bernard Kim telling investors last quarter that the company sees 'significant opportunity' in using large language models to improve recommendations.

Training AI on profile data that doesn't accurately reflect what users want is expensive way to build a more sophisticated version of the same problem. If the model learns that women overwhelmingly set age preferences for older men—and then optimises to serve those matches—it's encoding the stated preference bias directly into the algorithm. The blind date data suggests that approach might be systematically under-serving matches that would produce higher attraction.

Some platforms are already moving away from heavy reliance on stated preferences. Hinge famously removed filters from its free tier, forcing users to encounter profiles outside their specified criteria. The company framed it as an engagement play—more profiles viewed means more session time—but it also functions as a hedge against preference blind spots. Users might skip profiles they'd otherwise filter out, but at least they see them.

The more interesting question is whether platforms should surface this research to users directly. Imagine an onboarding flow that tells new members: 'Research shows people are often attracted to partners outside their stated preferences. We'll show you profiles that don't match your filters.' It's honest, evidence-backed, and directly contrary to the prevailing product philosophy of giving users more control over their experience.

No major platform will do this, of course. User control has become a core product value, particularly for women-focused apps like Bumble that position filtering as a safety and empowerment feature. Telling users their preferences might be wrong—even if the data supports it—runs counter to a decade of product positioning.

The UC Davis research won't change how platforms match users next quarter. But it should prompt product teams to question whether preference optimisation is the right north star, or whether the industry has spent years building increasingly sophisticated tools to help users avoid the partners they'd actually be drawn to in person.