The reason a viral post is rarely reproducible

A post reaches four million people. The same author, using the same structure, the same hour and the same subject, reaches nine hundred the following week. Both outcomes are drawn from the same distribution, and only one of them got a post-mortem.

What everyone sees

A large cascade is treated as evidence of a property in the thing that spread — a hook, a structure, a formula that can be identified and reused. That inference requires the outcome to be mostly explained by the item. Measurement of actual diffusion says it is mostly not.

What is actually happening

Goel, Anderson, Hofman and Watts measured diffusion across billions of events on multiple platforms and found the overwhelming majority of cascades terminating at or near their origin, with large structurally viral events extremely rare and their size only weakly predictable from content characteristics. Watts’s broader argument about outcome explanation identifies the reasoning error: outcomes drawn from high-variance distributions attract narrative explanation after the fact, and the explanation feels compelling precisely because the observed case is the one being explained. The practical consequence is that a formula extracted from a hit is fitted to a single draw, and its apparent failure on reuse is not a failure of execution.

Why it stays hidden

The variance hides because only successes are analysed. The thousands of near-identical posts that went nowhere are not collected, not compared and not published, so every study of what makes things spread is conducted on a sample selected for having spread. The comparison group exists but has no reason to be assembled.