Guides
Part of How to tell whether sharing or recommendation carried a viral spike
How to read a viral trends comparison
A skeptical guide to viral trends comparison: spot survivorship bias, test five dimensions, and decide what an episode comparison can actually show.
Every case study of a social-media post or trend that spread covers a survivor. Items made the same way that went nowhere are not written up. So the collection you are reading has been filtered by the outcome you want to explain.
This guide explains how to assess a viral trends comparison: what must match before two episodes are comparable, and which comparisons are worth the effort.
What to take away
- The comparison set is the whole problem. Compare hits to hits and you learn nothing, because everything in the set shares the property you are trying to explain.
- Two episodes are comparable only if the surface, the starting audience, the unit, the period, and the counting rule are close enough to hold constant.
- The honest output of most comparisons is a narrower question, not a conclusion.
Why the standard comparison fails
Take ten things that spread widely, look for what they have in common, and publish the list. The procedure feels rigorous and produces nothing, for a reason easy to state and easy to forget.
Whatever the ten share, thousands of items that went nowhere share it too. Without the failures in the set, no common feature can be distinguished from a feature of the medium.
This is survivorship bias: a collection selected only for spread cannot show which shared features distinguish hits from failures. What happens next to such a collection is how a claim hardens from anecdote into folklore.
The second failure is subtler. Cases are chosen after the fact, so features get named after the fact too, in terms picked to fit. A post is described as having an emotional hook because it succeeded, while an identical post that failed would have been described as trying too hard.
Naming the variable after seeing the outcome is cherry picking with an extra step, and that is why these lists never contradict each other.
What has to match
| Dimension | Why it has to be held constant | What happens if it is not |
|---|---|---|
| Surface | Structural routes to strangers differ enormously | You compare the surfaces, not the items |
| Starting audience | Size and composition set the initial conditions | You measure the head start |
| Unit | A clip, a thread and an image spread differently | The unit dominates every other difference |
| Period | Seasonality and competing events move the baseline | You measure the week |
| Counting rule | Views, reaches and impressions are different quantities | The numbers are not on the same scale |
Meeting all five is rare across operations and routine within one. In practice, match the platform surface, starting audience, unit being counted, observation period, and counting rule. Compare two of your own episodes on the same surface, close in time.
Five dimensions that must match
- Surfacestructural routes to strangers differ
- Starting audiencesize sets initial conditions
- Unitclip, thread, image spread differently
- Periodseasonality moves the baseline
- Counting ruleviews, reaches, impressions differ
People wave through the surface row, but TikTok's For You feed and YouTube Shorts feed are distinct surfaces with different recommendation systems and distribution pathways. The structural properties of a surface make two platforms non-interchangeable. Comparisons across operations are almost always uncontrolled on at least three of the five rows.
A comparison worth making
Compare your own hit against your own ordinary week, with the audience cohorts kept apart. The value is not in explaining the hit. It is in establishing how far the hit is from your own distribution, which tells you whether it is an outlier worth reasoning about or a good week being over-interpreted.
Consider the ALS Ice Bucket Challenge. A case study of its most widely shared posts can describe the campaign, but it cannot establish why it spread. A useful comparison would identify one platform surface and compare challenge posts with less-shared attempts on that same surface, while matching the starting audience, unit, period, and counting rule. Without the non-survivors and those controls, the example remains descriptive, not causal.
Is the hit worth reasoning about?
Compare your hit against your own ordinary week
outlier worth reasoning about
good week being over-interpreted
Two habits keep it honest. Write down what you expect to find before you look, because the outcome is known and the temptation to fit an explanation to it is strong. And include your failures in the set, even though nobody else will, since your own archive is the only place where the non-survivors are available to you.
The design rules for a comparison you cannot flatter are to define the question and counting rule before looking, include failures, and hold the surface, starting audience, unit, and period constant. They apply unchanged here.
Reading somebody else's case study
You will still read them, so read them for the right thing. A case study is a source of mechanisms and a source of vocabulary. It is not evidence about what works.
Three questions extract most of the value. What did the surface allow that made this possible at all, which the narrative rarely answers. What was the starting audience, which is usually the omitted number. And what was the counting rule, which is usually not stated.
If all three are missing, you are reading a story. Stories are useful for generating candidates and useless for settling anything, and treating one as a plan is the standard mistake.
Common questions
Is there any way to include the failures?
In your own archive, yes, and that is the argument for keeping one. Across an industry, no. Nobody publishes their misses, and no sampling method reaches items that were seen by almost nobody.
What about comparing against a competitor's hit?
You will be comparing surfaces, audiences and timing at once, with no way to separate them. It can be worth doing as a prompt to ask what else was happening that week, which is the same limited use described in reading a competitor's numbers.
How many of my own episodes do I need before a comparison means anything?
More than you will have. Treat a comparison of two or three as a way to generate a question rather than to answer one, and be explicit about that when you write it up.
Does a large sample of hits fix the survivorship problem?
No, and this is worth being firm about. The problem is the selection rule, not the sample size. A thousand survivors is still a set defined by having survived, and increasing the count only makes the resulting list more confident.
