Who engages matters more than how many

You got twenty likes in the first hour. All from the same group. Your running crew. Or your coworkers. Or that one Discord you are in. The post stops there. The next one gets eight likes. From eight different places. It runs for two days.

You look at the two. Twenty beats eight. The math is clear. But the second one moved and the first one did not.

Here is a study that measured exactly that. It watched 54 million people decide whether to join something. What made them join was not how many people invited them. It was how many separate parts of their life those people came from. One invite from five unconnected people beat five invites from one friend group. The study calls that structural diversity. The number of separate clusters.

The result is the strongest scientific case for varied engagement over a pile of it from one source. It is also a study of Facebook sign-ups in 2012. Not Instagram. Not a ranking algorithm. Sign-ups to a website. Where you could see who invited you. That difference matters and this post will not skip it. The study also explains one variable, not the whole outcome. Which post takes off stays unpredictable even when the inputs are known.

What follows is a reading of one paper and what it actually proves. It covers what structural diversity is. It says what the study measured and what it did not. The source is linked, dated and quoted. One section names what the study cannot tell you.

The study: 54 million invitations, two outcomes measured

Four researchers studied how Facebook grew. Two worked at Facebook. Two worked at Cornell. They are Johan Ugander, Lars Backstrom, Cameron Marlow and Jon Kleinberg. Their paper is called Structural Diversity in Social Contagion. It ran in PNAS on 17 April 2012. PNAS is a journal. Its full name is Proceedings of the National Academy of Sciences.

The paper studied 54 million email invitations. Facebook users sent them to people not yet on Facebook. Then the researchers watched. Did the invited person join? If they joined, did they stay? The study tracked 10 million new members for three months after they signed up.

It measured two things. The first was adoption. Did you join or not. The second was engagement. That meant logging in on six out of seven days in a row. Three months after joining. Both are binary. Yes or no. You either did or you did not.

Here is what they were testing. You get invited to Facebook by five friends. Does it matter whether those five friends know each other? The old model said no. It counted the invites. Five is five. The new model said yes. Five people from five separate parts of your life is a stronger signal than five people from the same group.

The paper backs the new model. From a news article about the study, quoting Kleinberg:

The decision depended on how many different social contexts the inviters represented. — Jon Kleinberg, quoted in Cornell Chronicle, 3 April 2012

The article is called To convince people, come at them from different angles. It was published by Cornell University. The study was announced the week of 2 April 2012 in the online edition of PNAS. The print issue came out 17 April 2012.

Structural diversity is the number of separate clusters

The term sounds technical. The idea is not. You have ten friends on a platform. All ten have done something. Joined a group. Bought a thing. Started posting reels. You see that they did it. Now you are deciding whether to do it too.

Structural diversity asks: are those ten people connected to each other, or are they separate? If all ten know each other, that is one cluster. One signal, repeated nine times. If they are ten separate people from ten separate parts of your life, that is ten clusters. Ten independent signals.

The paper says the number of clusters predicts adoption better than the total number of friends. From the Facebook research page for this study:

The probability of contagion is tightly controlled by the number of connected components in an individual's contact neighborhood, rather than by the actual size of the neighborhood. — Meta Research, Structural Diversity in Social Contagion

A connected component is the paper's term for a cluster. A contact neighborhood is everyone you know who has already adopted. The finding is that counting the clusters beats counting the people.

The same page gives the second half of the finding. Once you account for structural diversity, neighborhood size stops helping. It starts hurting:

When we controlled for structural diversity, we found that the size of the neighborhood becomes a negative predictor of contagion. — Meta Research, Structural Diversity in Social Contagion

Controlled for means they did the math to separate one effect from another. A negative predictor means more became worse. If two people have the same structural diversity, the one with fewer total contacts is more likely to adopt.

That is the result. Varied beats big. And beyond a point, big without varied makes things worse.

What it studied and what it did not

The study measured Facebook adoption. People got email invites from friends already on Facebook. They could see who sent them. The decision to join was visible and social. You knew that five people invited you. If you joined, you joined because of those invites.

Instagram does not work that way. You do not see how many people liked a post before you saw it. The platform's models see it. They use it as one input among many. The cause runs through software, not through your eyes.

The second difference is the behaviour. The study measured joining a website. Then staying active on it. Both are big decisions. Liking a post is not. Watching a reel for three seconds is not. The paper does not measure those.

The third difference is the time. Facebook in 2012 was still growing. It had not reached everyone yet. So an invite carried information. It said this thing is real and people you know are on it. Instagram in 2024 carries no such signal. Everyone is already on it.

So the mechanism is different. The behaviour is different. The time is different. That does not make the finding useless. It makes it indirect. The paper proves that varied social proof beats concentrated social proof when the proof is visible. It does not prove that varied engagement beats concentrated engagement inside Instagram's ranking models.

No published experiment we could find shows that structural diversity affects reach inside Instagram's own software. That is the gap between this paper and the platform most people care about.

Two things sit on either side of that gap. On one side, the mechanism is plausible. Instagram's models read engagement as a signal. The cold-start problem is real. A post with some engagement beats a post with none. And if the models are learning from patterns like the ones this study found, varied engagement would read as a stronger signal than clustered engagement. But that is reasoning, not proof.

On the other side, we do not know what Instagram's models actually do with clustering. Meta does not publish that. The system cards list signals. They do not publish weights. They do not publish whether clustering is even detected. So if someone tells you varied engagement gives you a 3x lift, or any other number, they are making it up. No such number has been published.

The honest position is this. The study proves varied social proof matters when people can see it. Whether it matters when only the models can see it is unproven.

What the author said about limits

Kleinberg gave a limit himself. The same Cornell news article quotes him:

These rules might not apply in all situations. — Jon Kleinberg, quoted in Cornell Chronicle, 3 April 2012

His example is a party. Three friends who know each other invite you. Then three friends who do not know each other invite you. The second case sounds less exciting, he says. Because if they are connected it suggests the party is real and your whole group is going.

The article says more research was planned to explore such cases. We could not find that follow-up study. If it exists and we missed it, the correction goes in this post's changelog.

The other limit is the platform. The study is about Facebook. A social network where you connect to people you know. The paper does not study Instagram. It does not study TikTok. It does not study anything where the primary signal is algorithmic distribution, not friend invites.

What this means if you make things

Start with what it does not mean. It does not mean you should avoid getting engagement from one group. A running creator getting engagement from other runners is not a problem. That is the audience. The study does not say clustered engagement is bad. It says varied engagement is better.

It also does not mean you should chase engagement from random unconnected people for the sake of variety. The study measured real invites from real connections. Not fake ones. Not bought ones from people with no connection to you. The signal worked because it was real.

What it does mean is this. If your post is only seen by one corner of your world, it stays in that corner. If it reaches separate corners, it has more chances to spread. That is true whether the signal is visible or not. It was true for Facebook sign-ups in 2012. It is probably true for Instagram posts in 2024. But probably is not proven.

The second thing it means is that size is not the goal. Twenty likes from twenty separate people beats a hundred likes from one group. If the study holds. And if Instagram's models detect clustering the way Facebook's adoption data did. Both of those are ifs.

Where this stops being useful

Here is what would prove this post wrong. Say Meta publishes a study showing that clustering affects Instagram ranking. Or say they publish the weights in their models and structural diversity is in there. Then the gap closes. The finding becomes direct instead of indirect.

The second limit is harder. Even if the mechanism is real, it does not tell you which post wins. The study found that varied social proof beats clustered social proof. It did not find that varied social proof guarantees success. Most posts with perfect structural diversity still go nowhere. This explains one variable. It does not predict the outcome.

The question worth someone's time is narrower. It can be answered. Does Instagram's software detect clustering in engagement? And if it does, how much weight does that carry against everything else the models see?

Until then you have this. A post that reaches five separate groups has more paths forward than a post that reaches one group five times. That was true for Facebook adoption. Whether it is true for Instagram reach is reasoning, not proof. But the reasoning is backed by the strongest study we could find on the question.

What we could not determine

Three things, named so no one assumes we checked and stayed quiet.

Whether Instagram's models detect or use clustering in engagement signals. Meta's system cards list signals. They do not describe structural diversity anywhere in the five documents we read for the post on how Instagram treats a new post as a cold item. So we do not know if the models see it. We do not know if they weight it. And we do not know how much it matters if they do.

Whether the finding holds for low-stakes behaviours. The study measured joining a website and staying active on it. Both are decisions that take effort. Liking a post takes none. Watching a reel for three seconds takes none. We do not know if the clustering effect is as strong for those.

Whether a follow-up study was published. Kleinberg said more research was planned to explore counter-examples. The Cornell article is from 2012. We could not find a later study by the same authors testing the limits he named. If it exists, we missed it.

How we did this

Scope: this post reads one published study and describes what it does and does not prove.

How we read it: the PNAS page for the full paper returned a 403 error and could not be opened on 18 August 2026. The PDF hosted at Stanford returned corrupted binary data that could not be read. So we read two sources. The first is the Facebook Research page for the study. It gives the citation, the finding, and a description of structural diversity. The second is the Cornell Chronicle news article from 3 April 2012. It gives the sample size, the two outcomes measured, and Kleinberg's own stated limit. Both were retrieved 18 August 2026.

We also attempted to open the PNAS page at a second URL. That also returned a 403 error. Web search confirmed the citation details: volume 109, issue 16, pages 5962-5966, DOI 10.1073/pnas.1116502109.

Limits: we could not read the paper itself. Every claim above is sourced to the Facebook Research page or the Cornell news article. Both are secondary summaries of the study, not the study's own text. The full paper may carry details, caveats, or numbers we could not report.

Sources

1. Structural Diversity in Social Contagion, Meta Research publication page. Authors: Johan Ugander, Lars Backstrom, Cameron Marlow, and Jon Kleinberg. Published in PNAS, 17 April 2012, volume 109, issue 16, pages 5962-5966, DOI 10.1073/pnas.1116502109. Retrieved 18 August 2026.
2. To convince people, come at them from different angles, Cornell Chronicle, 3 April 2012. Source of the sample size (54 million invitations, 10 million new members), the two measured outcomes (adoption and engagement defined as 6 of 7 consecutive days), Kleinberg's quote about different social contexts, and his stated limitation about counter-examples. Retrieved 18 August 2026.

Changelog

• 2026-08-18: First version drafted. Could not open the full PNAS paper directly; relied on Meta Research summary and Cornell news coverage.