Social media holidays statistics: what can be counted

What to take away

Why this belongs on an algorithms, trends and outages site

A themed day is a trend signal. It tells you what an audience expects to see in feeds on a known date. The algorithm decides which posts carrying that day's tag reach which people, and the visibility rule behind a tag count is an algorithmic choice you cannot see.

An outage during the same window removes posts and data from the record, so the day's numbers describe a disrupted period rather than the day itself. For this site, the directory below is not a calendar of holidays. It is a list of dates when ranking systems, trend cycles and platform failures are easiest to observe.

Read as a watchlist, each date is a scheduled stress test for ranking systems. A known spike in themed posts gives a clean window to observe how a platform orders or suppresses content. The same window makes an outage legible, because missing posts and data stand out against an expected surge. Use the directory to plan observation, not promotion.

The figures matter less than the provenance, because provenance is what tells you whether a number came from an algorithm, an organizer or a retail tracker.

The directory: the days and the source that documents each

Every entry in the first table falls on a fixed calendar date, and each has an institution that publishes it.

Giving Tuesday falls the Tuesday after US Thanksgiving, and Black Friday the day after it, so both move every year and neither has a single institutional publisher.

Small Business Saturday, the Saturday between them, was started by American Express in 2010.

The floating ones are a separate problem. For those, the date is arithmetic, not a fact you can cite.

The rest of the calendar is folklore. National Donut Day has a documented origin in the Salvation Army's 1938 Chicago fundraiser, but most food and animal days trace to a marketing agency or a PR calendar with no publisher at all. That absence is the finding, not a gap in the research.

Of the eleven days listed, ten carry a named publisher; Black Friday carries none, and the wider calendar has no publisher at all.

The figures that exist, and who publishes each one

Every row above carries a publisher, a year and a stated method.

The final column also shows what each figure counts: people with a condition, dollars spent, journalists killed, countries taking part. Not one of them counts the share of people who observed the day.

Earthday.org's billion and GivingTuesday's $3.1 billion are an organizer's estimate for an event it runs. Quote them with that label attached.

The retail rows come from different machinery. Adobe Analytics counts transactions it tracks, and American Express reports spending at small businesses. Neither describes observance.

Why none of these figures counts participation

The four figures people ask for break in four different places.

The third row looks answerable and is not. Outperform compared to what? A themed post differs from your average in the day, the subject, the format, the effort, and often the hour you published.

Any of those could explain a difference, and none was held constant. Naming the comparison group first is the whole of building a test that cannot flatter you.

Where a sampling frame would have to come from

Every defensible statistic starts with a list of who could have been counted. For a national survey that list is constructible. For an observance it is not, and the reason is structural rather than practical.

Participation is not a defined act. Somebody who thought about the day and posted nothing is not counted. Somebody who mentioned it without the tag is not counted. Somebody who used the tag ironically is counted as a participant.

Whatever gets counted is therefore a group that assembled itself. Anyone reporting on an observance describes people who chose to be visible about it, which is selection bias built into the definition rather than introduced by a sloppy method. Nothing downstream repairs it.

The survey route hits the same wall one step later. Ask people whether they observed a day and the ones who answer are the ones for whom the question is salient. The response rate tells you how much room the non-answerers have to be different, and here that room is almost the whole thing.

AAPOR's best practices for survey research list what a survey must disclose before its numbers can be read. Themed-day figures supply almost none of it. The same shortage runs through most published numbers in the field.

Wording adds another layer. Asking whether someone took part implies that taking part was expected. Pew Research Center's account of how survey questions are written shows how far phrasing can move an answer in a case like this.

What you can actually count

Two things, and they cover the decision you face.

Provenance is the first, stated in the takeaways above. Your own activity is the second, because you own the property, define the unit, and know the denominator. That is a real measurement about a real population of one operation.

Removing the season before you read your own numbers

Your own figures still need care, because themed days are not randomly placed in the year. They cluster in seasons when your audience already behaves differently, so a comparison against last month is contaminated by the calendar.

The correction is the ordinary one: compare like periods and account for the recurring annual pattern. Seasonal adjustment does this in official statistics, though you will not do anything so formal. The discipline transfers.

Compare the same week across years, or against a holdout of ordinary posts in the same window. Do not compare against a monthly average that spans several different conditions.

A plausible figure survives because nobody asks what it was compared against. An argument that rests on no figure at all is harder to knock down.

Common questions

Which figures in the table can I use?

Use them with the publisher, the year and the thing counted attached. Earthday.org's billion is Earthday.org's estimate. Adobe's $10.8 billion is US online spending Adobe tracked on Black Friday 2024, not all retail sales that day.

Surely somebody has counted tag usage?

Several parties have counted something and called it that. What each counted is the visible subset available through whatever access it had, on one surface, under an undisclosed visibility rule. Two such counts of the same day are not comparable, and neither is a count of participation.

Can I use a figure if I attribute it clearly?

Attribution fixes the courtesy, not the arithmetic. A source label does not supply the missing frame, so quoting the figure repeats an unsupported claim with a footnote. Say what is known instead, which is usually less than a sentence.

Is the honest answer just that nobody knows?

For the population-level questions, yes, and saying so beats an invented figure. For your own operation the honest answer is a measurement you can run, which is a better answer than the one you were asking for. The decision rarely needs a number in the first place, as the six worked situations show.

What do I do when a reader asks for benchmarks?

Give them the provenance check and your own historical numbers, and say plainly why no external benchmark exists. That reasoning is more persuasive than a figure would be, because it can be checked.