On Sunday, May 22, 2022, at 1:54 in the afternoon Pacific time, I posted a photo with a one-line question under it. It earned 19,285 likes, which is still the most any post of mine has ever earned and roughly 2,000 times what a typical post of mine earns. For four years that single afternoon has quietly poisoned every average I have looked at. This piece is about the difference between an average and a median, why that difference matters more for creators than for almost anyone else, and what my complete archive of 20,797 posts actually says about hour of day, day of week and post length once the outlier is put in its place. It is a short statistics lesson with my timeline as the dataset.
The Problem With One Good Afternoon
An average is the sum divided by the count. A median is the middle value when you line everything up. For most things people measure, height, say, the two sit close together. For attention they do not, because attention follows a power law: a handful of posts earn a huge share and almost everything else earns a little. When the distribution is shaped like that, the average is dragged toward the handful and stops describing anything you will experience on a normal day.
My numbers make the point. Across the complete window of my archive, March 1, 2022 to September 23, 2023, I made 20,797 posts and earned 454,160 likes. The average post earned 21.8 likes. The median post earned 8. If I planned my week around the average, I would be disappointed nine days out of ten. If I planned it around the median, I would be right about what a Tuesday feels like and still get the occasional Sunday.
Data and Method
The dataset is my own X archive export, complete through September 23, 2023, filtered to my own posts (retweets removed) and timestamped in Pacific time. I used only likes, because the export does not carry views and reposts are too sparse to analyze. For every grouping below I report three things: the count of posts, the mean, and the median. Where they disagree, trust the median. The 2022 outlier is left in the data on purpose, because it is real, but every claim below is checked against the median so that it cannot carry the argument on its own.
Two definitions. Hour is the clock hour the post went up. Length is the number of words excluding links.
Hour of Day: The 1 PM Mirage
Group my posts by hour and the 1 PM hour (1:00 to 1:59 PM) shows an average of 41.0 likes on 826 posts, far above every other hour, which cluster between 16 and 29. A naive reading says post at one in the afternoon. The median for that same hour is 7, which is the same as 8 AM and lower than any evening hour. The 1 PM average is one Sunday afternoon in 2022 wearing a trench coat.
Look instead at the medians. Morning hours, 6 to 11 AM, run a median of 7 to 8 likes on the largest volume of posts (6,769 posts, 32.5 percent of the total). Afternoon hours run 7 to 9. Evening hours, 6 to 11 PM, run a median of 10 to 12, the highest of the day, with 10 PM and 11 PM both at 12. Late night, midnight to 5 AM, falls back to 7 to 11.
So the honest statement is: my typical post does best in the evening, my biggest posts have happened in the afternoon, and I have historically posted most in the morning. Those are three different facts, and the average mashes them into one wrong instruction.
Day of Week: Sunday Is Real, Mostly
By weekday the averages are Sunday 29.0, Thursday 23.3, Friday 22.7, Wednesday 20.6, Monday 20.1, Tuesday 19.3, Saturday 19.0. Sunday leads by a wide margin. Is that the outlier again? Partly. Remove the single May 22 post and Sunday's average drops to about 21.6, still at the top of the pack but no longer a different species. The medians are flatter: Sunday, Monday and Thursday at 9, the rest at 8. Sunday is a modest real edge plus one giant lucky afternoon. Treat it as a good day, not a magic one.
Volume runs the other way: I posted least on Sundays (2,616 posts) and most on Tuesdays (3,202). Part of Sunday's edge may simply be less competition from me against myself.
Length: Six to Twelve Words
Bucketing posts by word count, the averages are 21.4 likes for 1 to 5 words, 27.4 for 6 to 12, 17.7 for 13 to 25, and 15.4 for 26 or more. The medians are 6, 9, 9 and 9. Here the average and the median agree on the shape even if not on the size: the shortest posts underperform, the one-sentence posts win, and long posts fall off. The median post in my archive is 11 words. The sweet spot is one sentence with a little texture, not a fragment and not a paragraph.
Replies Are Not Posts
A last split that matters for anyone counting. Within the window, 491 of my posts began with another account's handle, which X treats as replies. Those average 3.7 likes with a median of 1. My original posts average 22.3 with a median of 9. Mixing the two in one average understates the originals and overstates the replies. Replies are for relationships. They should be measured by what the conversation produced, not by likes.
Why Creators Get Fooled More Than Most
Two reasons. First, the platforms report averages because averages are easy to compute and flattering to show. A dashboard that said your median post earns 8 likes would be accurate and demoralizing, so nobody builds it. Second, creators remember peaks. The May 2022 afternoon is a story I can tell; the 1,874 posts that earned zero likes in my archive are not. Memory is an average weighted by emotion, and emotion loves outliers.
The research on cultural markets says this is not a personal failing. Salganik, Dodds and Watts showed in 2006 that in a market where people can see what others like, which items win is highly unpredictable and the inequality of outcomes is extreme, even holding quality constant. That is a description of a timeline. Nassim Taleb's distinction between Mediocristan, where averages work, and Extremistan, where a single observation can dominate the total, is the same idea in a different vocabulary; attention lives in Extremistan.
Limitations
This is one account and one platform. Likes are a thin proxy for what I actually care about, which is who shows up and stays. The archive has no view counts, so I cannot separate reach from resonance. The complete window ends in September 2023 and the platform's ranking has changed since, so the hour and weekday patterns are a historical finding, not a forecast. Hour-of-day analysis also cannot separate when I post from when my audience is awake, because I never varied my schedule on purpose. A real test would be to post the same kind of thing at different hours for a month and compare medians. I may do that and report back.
What This Means
Report medians to yourself. If a dashboard only gives you averages, divide your posts into buckets and look at the middle one. Expect your typical post to do what your median says, and treat anything above the 90th percentile as weather rather than skill. Post one-sentence things in the evening if you want the typical post to land, and save the thing you believe in for a Sunday afternoon, not because the data promises anything, but because that is where the weather has been kind before. And when someone shows you an average, ask what single afternoon is hiding inside it.
Sources
- Hellcat Blondie, X archive export (data download), complete through 2023-09-23; own posts only, 20,797 rows in the analysis window 2022-03-01 to 2023-09-23. Analyzed November 2026.
- X Help Center, "How to download your X archive", https://help.x.com/en/managing-your-account/how-to-download-your-x-archive (accessed 2026-11-05).
- Salganik, M. J., Dodds, P. S., and Watts, D. J. (2006). "Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market." Science, 311(5762), 854-856. https://doi.org/10.1126/science.1121066
- Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House. (Chapter 3, "The Speculator and the Prostitute", on Mediocristan and Extremistan.)
- Huff, D. (1954). How to Lie with Statistics. W. W. Norton. (Chapter 2, "The Well-Chosen Average".)
- Wilcox, R. R. (2010). Fundamentals of Modern Statistical Methods: Substantially Improving Power and Accuracy, 2nd ed. Springer. (On the median and trimmed means as robust measures of location.)