August 5, 2013 by Benjamin Rey
Data Science blog post series
At Wisemetrics, we have been gathering a huge amount of social media data for over a year and thought it was time to share some insights we are getting from it. This post is part of a Data Science short blog posts series we hope you will find interesting.
We know it’s short
Facebook posts’ lifetime is known to be very short. Several studies have found that most engagement with a post happens within the first few hours (see for instance nice studies done by Edgerank Checker’s, Optim.al, Sotrender).
Having our hands on the right data, we’ve decided to take a peek at this lifetime thing. And yes, we found, just like others, that 75% of engagement occurs within the first 5 hours. But we’ve also looked at Impressions, and Reach.
How much shorter can it get?
On the graph below, we present median Engagement, Impression and Reach, over time with a confidence interval of +/-5% (to show variance among posts)
Impressions are even shorter than Engagement, with only 2 Hrs 30 min to reach 75% of its max, and Reach is even worse: 75% of your audience sees your message in less than 2 Hrs!
Can it get worse? Well if you care most about your fans, it takes only 1 Hr 50 mins for a post Reach to get to 75% of it’s potential (not shown on this graph).
There is quite a bit of variance between posts, but for short period of time all posts show the same behavior, and it takes a mere 30 minutes for a post to get 50% of its global Reach.
The idea behind looking at posts’ progress over time is to be able to predict, as soon as possible, if a post will fail or beat all expectations to adjust community management efforts (e.g.: rushing in order to publish a new post or waiting a bit more).
In terms of modeling, the log-log shape of posts progress over time is a nice discovery. However even though all posts do seem to share a common shape, variance is very large (as we could see with the large tubes from our initial graph). Using a linear log-log fitting, we thus couldn’t map each post into a “as-usual” or “killer-post” category just by looking at performance from the very first few minutes of a post.
Going a step further, through a machine-learning approach based on derivatives of the curve (speed of increase) to predict end-point, we’re finally getting OK results, but it’s still not quite convincing. It needs to be coupled with models predicting a post performance even before it is published. But that’s another story.
Many brands care for fans’ Reach, sometimes more than for Engagement. Optimizing the timing of your post is thus mandatory.
The best thing to do is to thoroughly analyse your posts history, as well as your peers & competitors, looking at dozens factors at a time, and predict the optimal timing, just for you (good news, it can be done through machine learning).
Next Data Science blog post coming in a few weeks
“Paid posts: will the thousands of likes & comments you get boost your future unpaid posts?”
bitly link for this post: http://bit.ly/15Alvdi