Were the weakest LinkedIn ties best for finding a new job?

Close contacts often recycle familiar information

People closest to you often know many of the same people, employers and ideas that you do. A looser connection may bridge into a different circle and carry information you would not otherwise receive. That is the intuition behind the strength of weak ties—one of social science’s most influential claims.

For decades, much of the evidence was observational. People with different networks also differ in occupation, ambition, geography and countless other ways. A weak tie appearing before a new job did not prove that the tie caused the job transmission.

LinkedIn’s recommendation tests created a vast experiment

LinkedIn routinely tested changes to its People You May Know recommendation system. Some randomly assigned versions were more likely to suggest people with many mutual connections; others surfaced people with fewer. Those tests changed what kinds of new ties members formed without directly assigning anyone to accept a connection or apply for a job.

Researchers analyzed a 2015 experiment involving more than four million members—98.8% of that edge-level sample was in the United States—and a set of worldwide 2019 experiments involving more than 16 million. Across the second wave, the data contained around two billion new connections, more than 70 million job applications and roughly 600,000 cases the researchers classified as job transmission.

That final measure was specific: member A reported working at a company before member B later reported working at the same company, at least one year after A’s start; their LinkedIn tie had formed at least one year before B’s reported move. It is a platform proxy for sequential employment, not an audited referral or account of who secured the hire.

The sweet spot was weaker—not weakest

Randomized recommendation variants changed the strength of the marginal new ties members formed. The researchers used that assignment as an instrument to estimate how tie strength affected later platform outcomes. This provides a causal interpretation for affected members under the study’s independence, exclusion, monotonicity and interference assumptions.

But weaker is always better was not the result. When tie strength was defined by mutual connections, the relationship was inverted U-shaped: moderately weak ties produced the largest estimated marginal increase in the job-transmission proxy, while the very weakest and strongest ties produced less. When strength was defined by later message intensity, the lowest-intensity ties produced the largest estimated marginal increase.

The answer therefore depends on what weak means. A structurally distant contact and a person you rarely message are related ideas, not interchangeable measures.

Industry changed the pattern

Weak ties produced more job applications to more digitally intensive industries; stronger ties did relatively better for applications in some less digital industries. That variation fits the possibility that different work depends on different information, trust and referral channels.

It also blocks a universal networking formula. A bridge to a new technical community may expose someone to novel opportunities; a role built on tacit local trust may travel differently. The platform experiment identifies an average pattern across a huge population, not the optimal relationship for one specific career decision.

Scale cannot erase the platform’s boundaries

The research was conducted inside LinkedIn. Members, recommendations, applications and inferred hires all reflect one professional platform and the years studied. The experiment changed recommendation exposure; it did not force people to form ties or randomly assign lasting relationships.

The instrumental-variable analysis also depends on exclusion assumptions: the recommendation variants should affect jobs through the ties they induced, not through another unmeasured route. The authors test those assumptions in several ways, but no design turns a platform proxy into every path by which a person finds work.

Two authors were LinkedIn employees when the research was conducted, and the paper reports current or former employment and a significant financial interest in Microsoft, LinkedIn’s parent company. Aggregated replication code and data sufficient for the paper’s correlational regressions are public; individual-level member data are restricted.

Count bridges, not contacts

In these LinkedIn experiments, marginal new ties outside a close circle produced higher estimated values on the platform’s job-transmission proxy. That does not support indiscriminately collecting names or assuming the same result for an existing offline relationship.

The strongest conclusion is relational and modest: in the studied LinkedIn experiments, moderately weak bridges produced the largest marginal transmission estimate under one measure of strength. They do not replace strong ties, and they do not convert a network into a score of personal value.