As I noted in a previous article titled “Why dating apps suck,” I talked about why the business model of most dating websites and dating applications is fundamentally at odds with creating successful monogamous long-term relationships. The short version is that “pay to play” dating apps make money from frustrated users - not satisfied users. Creating successful monogamous long-term relationships takes away potentially paying clients.
Today, I’m going to talk about OKCupid, which was started roughly twenty years ago. For a period of close to ten years, OKCupid was the titan in the free online dating space, and arguably the best dating website compared to its paid competitors as well. At the time, it was mostly free.1 OKCupid had a strong, transparent, and effective match-making system, one which took multi-dimensional matching to the max.
These features make OKCupid a good case study for dating app degradation of the type I’m describing. OKCupid’s transition from highly successful free dating website to a barely-functional “pay to play” dating app is easy to explain.
The problems with single-dimensional matching
Within any given culture, there is usually a broad consensus about what appears attractive. Many dating websites and apps generate and use attractiveness ratings for their users, including at least some later versions of OKCupid. Others use sophisticated machine learning algorithms that are primarily dominated by attractiveness as a latent variable - since most of their users are swiping left or right in several seconds after a glance at a photo.
I have seen apps use attractiveness ratings in three different ways:
Showing more attractive users more frequently.
Putting more attractive users behind a paywall.
Steering users towards users with similar attractiveness ratings.
In general, most dating apps find that promoting the visibility of attractive profile photos helps promote user engagement. However, this tends to impair matching in two ways. One is a simple framing effect: Next to the most attractive profile photos, the main population of users look unattractive by comparison.
The other is that if most attention is focused on a small number of profiles that have attractive photos, most online matches will involve one of this small number of attractive profiles. However, dozens or hundreds of online matches on a single profile at best will translate to one monogamous long-term offline match. Generating more successful long-term matches requires improving the visibility of less attractive profiles.
Single-dimensional matching at its best
Assortative matching (pairing users of similar overall attractiveness) has some support in the literature on human mating practices. OKCupid experimented with this in the early stages of their transition to a “freemium” model. However, it also has limitations. A key fact: Unattractive people are unattractive for different reasons.
This is true for physical features (“too thin” vs “too fat”) but extends past profile photos. For example, women are more religious than men. Assortative mating by overall profile attractiveness will disproportionately pair women who are unattractively obsessed with religion with men who are unattractively hostile to religion. Similar patterns can hold for video games, fashion, fishing, astrology, and other gendered interests, values, and norms.
It’s easy to find a match for the most attractive 1% of the population; the real test of a match-making system is finding a match for members of the population who are not conventionally attractive. To generate a large number of pairings means finding the odd matches, the people who are appealing to each other but not the larger population. This means considering rare niche interests, sexual fetishes, and other unusual aspects of attraction.
The saga of OKCupid: An example of multidimensional matching
At the core of OKCupid was a matching algorithm that was incredibly multidimensional: A question-and-answer matching system using a large user-generated question bank. This system was highly effective. It has been degraded as part of the transition from free dating website to freemium mobile dating app.
An example of how the system worked is to the left in the above system. To the right is the “gutted” version presented to new users signing up on the OKCupid app as of 2021. The user answered the question (“pizza”), indicated what answers they wanted from a match (“either”), and how important the question was if they cared about the answer— a little, somewhat, very, or mandatory.2
Users could also add long-form text explanations, choose whether or not to answer publicly, submit new questions of their own, and select specific questions to answer that they saw on another user’s profile. The importance levels corresponded to a numerical weight, which was then be used to calculate a fractional compatibility rating. This is a weighted sum, where an acceptable answer is scored as the importance of the question (e.g., 10 points) and an unacceptable answer is scored as 0 points.
In the golden age of OKCupid, these compatibility ratings in each direction were then combined and reduced by a pessimism factor, in a formula something like this:3
The pessimism factor played a key role in OKCupid’s functionality. First, it prevented users from seeing spurious high matches based on having a single question in common. Second, it encouraged users to answer more questions, because answering more questions would lead to higher match percentages and therefore match visibility.
OKCupid’s system was not perfect; a number of questions were badly designed. For example, there were many questions worded in ways like “Would you date someone who was messy?” instead of simply asking “Are you messy?” OKCupid’s two-part answer system already added an implied “would you date” clause to every question!
OKCupid’s two-part answer system meant that “Would you date someone who was messy?” therefore asked someone if they are willing to date someone who was themselves willing to date someone messy. This version of the question is at best difficult to understand and generally does not improve match-making.
The gutting of the algorithm
In the gutted version, users were prompted to answer fifteen questions without asking them if those questions were important or if they cared how another user answered. New users started off with bad data, in other words. Additionally, the pessimism factor was removed, generating spuriously high-percentage matches based on a small number of questions in common.
Third, the algorithms used began to de-emphasize compatibility ratings. While the match percentages of users with few questions answered are generally low (set to 50% at first), users with low compatibility would be thrown into match queues indiscriminately. This is true even if a user selects the “Match %” tab when searching for other users.
There was no “feed” to begin with
In 2004, OKCupid, like many other early social websites, did not have an algorithm dictating which users were shown to each other when. Instead, it had search functionality and a messaging system. Users were free to ignore OKCupid’s compatibility ratings and seek out attractive-looking users with low match percentages by sorting on other attributes such as last log-in time; very few users chose to do so for long.
It’s not clear that a dating website or app with a large modern user base can get away with a “free-for-all” environment without very careful consideration to throttling user activity. This is even true for hookup apps that don’t pretend they exist to facilitate long-term monogamous matchmaking.
One of the few restrictions on the free-for-all nature of early OKCupid was that message inboxes had a limited amount of space. Users were unable to communicate with more than a certain number of other users at a time.4 Users who were very desperate or very attractive soon discovered this limit.
Ironically, while a point of friction with users, a lower limit may have been ideal for match-making activity for the same reason that throttling the visibility of more attractive users is ideal: Focusing too much attention on a small share of users won’t generate long-term matches.
Recap
As a free dating website, OKCupid had a highly effective multi-dimensional matching system that was based on a very large body of compatibility questions and a match % system that encouraged people to answer more questions. This was quite effective at matching couples into long-term relationships based on fundamental compatibility in a way that is impossible for a system based on one-dimensional measures of attractiveness.
However, like most dating apps and websites, OKCupid’s service was degraded in order to successfully monetize its user base, “cashing out” on past success in order to squeeze profits out of frustrated users.
To be precise, monetization started with “A-list” convenience features in 2008, though AFAIK the “service degradation” part of OKCupid’s life cycle didn’t really get into gear until after Match Group bought OKCupid in 2011.
The number of levels had been reduced in the 2021 version of the app as opposed to the older version. If OKCupid has brought back the extra level, let me know and I’ll edit the article appropriately.
A YouTube video from 2013 claims that OKCupid takes the nth root instead of the square root and also - inconsistently - describes this as the geometric mean. Since the geometric mean requires the square root, and the nth root would not produce the numbers typical of OKCupid match ratings in the circa 2004-2011 period I’m talking about, and the video doesn’t include the pessimism factor I recall very well as an OKCupid user during the early OKCupid era, I’m guessing.
I think the earliest inbox limit was 50 messages, but I’m not sure.