Twitter Impression Echoes
An examination of Twitter Analytics’ raw data reveals a surprising and somewhat mysterious artifact.
We always assumed a tweet’s best day for impressions is the first. We also assumed the number of impressions steadily declines thereafter. It turns out this may not be the case. Based on some initial, rudimentary analysis we have found a counterintuitive ‘echo’ on day five where tweets perform better than days two through four. There is also evidence of a less pronounced echo on day 12 and some arguable evidence there is a still smaller third echo around the 21st day.
Rather than seeing the data through the filter of somebody else’s analysis, we wanted to get as close to the source as Twitter will allow: this is Twitter Analytics which many may not have even noticed, let alone turned on. It is worth it, however, as it is a potential treasure trove of raw data available to anyone — even those who have not signed up as a Twitter developer. All you have to be prepared to do is a little busy work.
From the web app, we navigate to the home page, click the More menu item, then click the Analytics item and then click the View all Tweet activity link from any where on that page. From this next page, Twitter allows the download of the past five calendar months of tweet data by progressively selecting each month from the Last 28 Days drop-down list. For each month, we then click the Export data drop-down and select By Tweet from the resulting menu. This short video will walk you through the process
We go through these steps every few days for a number of our social clients. The resulting five CSV files are simply loaded into a spreadsheet for bulk storage. For one of our clients we collected 17,244 of these raw records spread over a population of 385 tweets. We made this data the focus of our analysis for this article.
The steps outlined above take about 10–15 per day per client so we don’t do it every day and on no specific schedule. For this dataset, there is an average of 45.0 impression counts per tweet spread over the maximum 150 day moving window covered by Twitter’s five months of downloadable data. This works out to an average of a little over three days between downloads. We tweet at least once a day for this client, so there are quite a few instances where day one data is not collected when the tweet goes out. This is the case whenever a tweet goes out just subsequent to one download session and a couple of days before the next. There are 214 tweets of the 385 for which we have a day one impression count. The 171 remaining tweets we only have impressions counts subsequent to day one.
What we know for sure is when an impression count for a given tweet is greater than the previous impression count. For the 214 for which we have day one data, we have a divisor that can be used for subsequent calculations for that particular tweet. But we didn’t want to filter out the 171 tweets where day one data was not present. Taking all this into account we wound up with 6,883 impression increments which were then fed into the subsequent analysis.
The subsequent impression count changes are still useful so long as we know the specific day in the life of the tweet on which that change in the impression count falls: in the range of days from two through 150. So long as we aggregate the incremental impression counts that fall into each day of the 150 day window, we have meaningful data for the entire population of tweets. We have distilled all of the data into 150 total impression counts: one for each of the days of the window.
For this dataset the aggregate impression count for days one and two is 35,307 and 7,853 respectively. We therefore calculated day two to be 22.2% of day one, which is the first data point in the visualization above. We omitted day one from the visualization because it is 100% — which doesn’t add anything other than making the behaviour of the subsequent days less easy to observe. The day three aggregate is 5,486 which is 15.5% of day one also as shown in the chart above. We went through similar calculations for all 150 days. However, for this visualization, we cut it off at 31 days as the rest of the values are essentially zero and are simply plastered against the x-axis.
Here’s the surprise: what we expected to see was a steady, power law decline in the number of impressions per day. What we saw instead was a peak at day five, which we’re calling an ‘echo’. Another less dramatic echo occurs on day 12. While it’s almost imperceptible, we think there may also be yet another echo around day 21.
In short — we’re not quite sure.
This particular client only has about 100 followers so any number of impressions much greater than that is purely as a result of subsequent engagement by other Twitter followers. A retweet, like or reply from the right influencer can send the impression count easily into the thousands. So our theory is the initial decline of the tweet impressions is stopped and turned round by a retweet or reply from a Twitter user with a large number of followers. Generally speaking, this would seem to take a couple of days and has the effect of ‘resetting the clock’ on the tweet after which the steady decline re-establishes itself. But this gives the tweet an added opportunity to be noticed by yet other Twitter users with significant follower counts.
To be candid, we’re not sure this behaviour can be used to any specific advantage. We’re glad to know it exists but how does that make it useful? Again, we’re not sure. However, it does confirm that structuring tweets so that they invite engagement by Twitter influencers is absolutely as important as we thought it was. Simply stated, it makes or breaks a tweet. As time permits we are going to do similar analysis on other clients’ datasets to see if the behaviour is repeated or if other patterns emerge.
Thank you very much for reading. May all of your tweets get a reply, a like or get retweeted!
©2020 Intellog Inc.
How about you? Have you noticed similar behaviour? Or other Twitter behaviours worthy of closer examination? Let us know! Do you have a lot of Twitter raw data you would like to analyze in a similar way? We would love to do this analysis for a Twitter user with significant follower count to see if that impacts the analysis. If you’re that kind of user and if we can help with any of that please get in touch!
Intellog Inc. is a Calgary, Alberta, Canada-based firm specializing in digital content creation, social media marketing and digital project development. Also, may thanks to our friend and colleague Tim Beck who read the draft and made many helpful suggestions.