Viral trends: planning, execution and measurement

"Viral" is an analogy, and it is worth taking the analogy seriously for a moment, because it says something precise. An infection spreads when each case produces, on average, more than one further case. Below that line it dies out however dramatic the individual cases are. Spread on a platform obeys the same arithmetic: each view has to produce, on average, more than one further view, and it has to do so before the item's window closes.

That single ratio is the whole subject. Planning is about raising it before publishing. Execution is about timing so the window is open. Measurement is about estimating it from your own numbers rather than from how it felt.

What to take away

• Spread has two engines, sharing and recommendation, and they leave different fingerprints. Most of what gets called viral is the second engine, and it is measured differently.
• You can plan the participation cost, the copyable shape, and the timing. You cannot plan the spark, and anyone who says otherwise is selling the spark.
• Measure the ratio, not the total. A total tells you what happened; the ratio tells you whether it was going to keep happening.

Two engines

	Sharing	Recommendation
Who moves the item	People, deliberately, to people they know	The system, to people predicted to respond
What raises the ratio	Something worth being seen to pass on	Something that produces the measured response quickly
The fingerprint	Arrives in clusters, by group; slow start, then bursts	Arrives smoothly; steep early rise; follows a few strong early signals
Where it stops	When the groups it can reach have all seen it	When the response rate falls below what the system can find elsewhere
What you can do	Make it cheap and safe to pass on	Make the first seconds and the completion strong

The two engines interact. Sharing produces the early response that recommendation looks for; recommendation produces the audience that sharing draws from. But they answer to different things, and an item built for one can fail on the other. Something people want to be seen passing on may be slow and thoughtful, which the recommendation engine reads as weak early response. Something the recommendation engine loves may be nothing anyone would put their name to.

Knowing which engine carried a spike is the first measurement job, because it tells you which of your choices mattered. The plain mechanics of how a recommender picks what to push are in what a recommender can and cannot infer.

Planning what can be planned

Participation cost. The trends that spread furthest are the ones cheapest to join. A shape that someone can copy in a minute, with their own material, beats a shape that requires skill. If you want your thing to be copied, strip out everything that makes copying hard.

A copyable core. Separate what must stay the same for the thing to be recognizably itself from what each participant changes. Trends with a small fixed core and a large variable part run longest, because each copy is new enough to be worth seeing and familiar enough to be understood.

Hooks for both engines. A reason to pass it on, and a reason to keep watching. These are different reasons, and the plan should name both.

A ready follow-up. If the ratio does go above one, the window is short. The most common execution failure is having nothing ready when attention arrives. Prepare the second and third item before the first one is published.

What cannot be planned is the spark: the particular reason this one, out of a thousand similar attempts, catches. Sparks look inevitable in hindsight and are invisible in advance. The honest plan raises the ratio for every attempt and accepts that most attempts will still sit below one.

Executing on timing

Every trend has a life: a small group starts it, a larger group joins once it is recognizable, then the volume peaks, then the copies outnumber the audience's patience. Joining early means being seen as part of the thing; joining at the peak means competing with the most copies for the least remaining attention; joining after means being seen as late.

Where a trend sits in that life is visible from your own feed, though your feed is a biased sample. The better instrument is a note of when you first saw it, and how many copies you had seen a week later. Doubling weekly is early. Ten copies a day is the peak. Copies that comment on the trend rather than doing it mean it is over.

Trends at the product layer move more slowly and are easier to time; the distinction between product, behavior and content trends is in the three layers of platform change. Content trends, the kind this page is about, are the fastest and least forgiving.

Measuring the ratio

You cannot see the ratio directly, but you can estimate it.

• Shares per view, per day. If each view yields a share, and each share yields several views, the product of those two is your ratio for the sharing engine. Watch it fall, and it will, because that fall is the end arriving.
• The time signature. A sharing-driven spike is lumpy: quiet, then a step, then quiet, then a bigger step. A recommendation-driven spike is smooth and front-loaded. Plot views per hour, not per day, and the shape shows which engine.
• Response in the first hours against your own baseline. The recommendation engine decides early. If the first hours match your normal items, the system saw nothing unusual, whatever happens later through sharing.
• The decay. Every spike decays. The decay is not a failure and is not a change in the ranking inputs; it is the ratio dropping below one as the reachable audience is used up, and the bounded slots of a feed being refilled with whatever responds next. Note the half-life. It is a property of the audience, and it will be similar next time.

The tool for thinking about this exists in another field, the basic reproduction number, and the reason it transfers is that the arithmetic of chains does not care what is being passed along. The related idea of an information cascade, where people act on what others appear to have chosen rather than on their own judgment, explains why spread can be large and still tell you little about the item's quality.

Common questions

Something of mine went viral once. Why can I not repeat it?

Because the spark was not yours to control and the ratio was probably only just above one. Repeat what you can plan, and expect the same distribution of outcomes: mostly nothing, occasionally a lot.

Should I join every trend I can?

No. Each one costs production and dilutes what you are known for. Join the ones with low participation cost that fit what you already make, and only while they are early.

Is a paid push the same as a viral spread?

It raises views without raising the ratio. When the push stops, the spread stops, which is the test that tells the two apart.

How do I tell if a spike came from recommendation or sharing?

The hourly shape, and where the viewers came from if your analytics report it. Smooth and early is recommendation. Stepped and clustered is sharing. Both together is the rare case that goes furthest.