Quick Analysis of Your Customer Support Twitter Handle

Posted by Sinan Ozdemir on 10/3/17

You already know that a decision that is backed by some analysis and data is usually better than a decision that comes purely from the gut. So when you need to make a decision on ways of improving your customer success experience, you need some data to work with. Most organizations employ some twitter usage in their customer success, some more than others. That’s why it’s important to do some analysis on incoming tweets and how well they are replied to.

I used Python and some basic (and free) resources to do a quick and dirty analysis of a few companies’ public twitter feeds.

Let’s start with @Airbnbhelp, a very active twitter handle. By creating an app and using the free twitter API I can grab about 10 days worth of tweets, enough to get started.

First let’s look at a graph on the volume of tweets they receive by the day.

 

1

 

Wow, there’s a huge bump on the 20th of September of incoming tweets, If we take a look at some of the most common words/phrases on that day, we might get a sense for why that happened:

On September 20th:

The phrase/word `impacted` occurred 570 times

The phrase/word `activated` occurred 572 times

The phrase/word `earthquake` occurred 572 times

The phrase/word `program` occurred 669 times

The phrase/word `help` occurred 691 times

Wow! The tweets on that day are mostly in response to Airbnb’s disaster response program in response to the large earthquake in Central Mexico on the 19th!

Suppose now that we wish to see volume of tweets by the hour of the day. We can plot the volume of tweets that came in at different hours of the day (each hour is in military standard and in PST).

 

2

 

We can see a big bump of incoming tweets at 9am and the volume kind of tapers off around 5pm.

Let’s see if there’s a difference in response time during work hours (we will assume that’s 9am-5pm PST) and non work hours. To do this, I’ll grab the tweets that @AirbnbHelp sent out and count the number of times their tweets were in direct reply to a mention sent previously. Once I match up tweets and their responses, I will calculate the response time of that tweet in minutes. Then I can calculate @Airbnbhelp’s response time by the hour.

 

3

 

Finally, to summarize, I can calculate the response time of tweets that came in during work hours and compare that to the response time of tweets during non work hours.

Response time to tweets during work hours is 4.0 minutes

Response time to tweets during non-work hours is 8.0 minutes

Airbnb is definitely doing very well when it comes to response time. I ran this analysis on @NewEggService and got these results:

Response time to tweets during work hours is 46.0 minutes

Response time to tweets during non-work hours is 265.0 minutes

This analysis is by no means complete and should be reflective of the above companies' ability to handle customer service, but it does indicate that we can glean valuable insight simple from public twitter feeds.

The code on how I achieved these graphs can be found on our public Github here.

Topics: Artificial Intelligence, natural language processing, Customer Support, data visualization