A/B Testing Chatbots: How to Start (and Why You Must)
April 12, 2019 –
“Hey there 👋
I’m a Bottybot!
How can I help you?”
I don’t know which websites you’re going to visit today … but you’ll end up visiting at least one where you’ll hear a *Pop* sound and a bot will start “talking” to you.
… Offering you pre-sales support.
Or assisting you with your post-sales questions.
Or simply offering support.
Every day chatbots have millions of such conversations with users; bringing real, tangible business results such as more leads, more sales, and higher customer loyalty. And they’re pretty mainstream with a whopping 80% of businesses predicted to use them by 2020.
Because chatbots drive revenue, they can — just like any other revenue channel — be optimized for better results.
Optimizing Chatbots with A/B Testing (and other experiments)
Depending on how you use chatbots in your marketing, sales, and support strategy, running experiments on them can offer many benefits.
For example, chatbot experiments can help you identify:
- Pre-sales sequences that generate more and better leads
- Trial messaging that converts more leads into customers
- Onboarding experiences that convert better
- Customer success sequences that result in higher customer satisfaction (and loyalty)
- …. And support sequences that result in fewer tickets
In short: If you’re a business using chatbots, you can improve your ROI from the channel with A/B testing.
Quite a few chatbot solutions even come with native A/B testing functionalities that allow businesses to run experiments to find the best performing messaging, sequences, triggers and more.
But in order to run meaningful CRO experiments for chatbots, you must use the right optimization process.
A/B Testing Chatbots: The Process
Before you begin creating your chatbot experiments, first choose the metrics you want to improve.
For example, if you’re using a chatbot for marketing, your metric could be the number of leads who opt-in after a successful chatbot interaction.
Alternatively, if you’re using a chatbot for boosting sales, your metric could be the number of trial leads whose engagement score improves because of interacting with the chatbot.
Finally, if you’re using a chatbot for offering support, your metric could be the percentage decrease in the number of inbound tickets.
Whatever it is, once you’ve identified the metric (or metrics) to optimize, you’re ready to begin working on your chatbot experiment.
Here are three simple steps for setting up and running winning chatbot A/B tests:
Step #1:Hypothesis Crafting
Just like regular website or app experiments, chatbot experiments, too, begin with a clear hypothesis.
For example, when Magoosh, an online test preparation company, decided to run an onboarding experiment, it started off with a clear hypothesis:
If we send trial customers a welcome onboarding message when they first log into a Magoosh product, they will be more likely to purchase premium accounts in the future.
While Magoosh didn’t exactly test a chatbot, it did test if sending an automated welcome customer onboarding chat message could help with more conversions.
In your chatbot testing strategy, your hypothesis could become “Offering automated chatbot assistance to new trial signups would result in … “
You get the idea, right?
Tools for Writing Hypothesis for Your Experiments: These are five really cool CRO tools that will help you write a winning hypothesis for A/B testing your chatbots.
How to Create a Winning A/B Test Hypothesis: This webinar breaks down the process of writing a winning hypothesis into five simple steps. A must-watch if you’re only starting out with experiments.
Complex A/B Testing Hypothesis Generation: This is another excellent tutorial on writing hypothesis for your experiment. These hypothesizing tactics apply seamlessly to chatbot experiments.
Step #2:Designing the Experiments
Just as you would in a regular A/B testing or CRO experiment, in your second step, you need to “create” your chatbot experiments.
In this step, you need to translate your hypothesis to a “change” (or a set of changes) to test.
For example, if you hypothesized that a “more branded” chatbot will get better results for your marketing team, in this step, you’ll have to see what elements of your chatbox could be branded better. It could be your chatbot’s voice or tone or simply the visual interface.
While you’re at this step, do check out this guide from the awesome folks from Alma. It will be very helpful for designing your experiments. For instance, in this branding experiment, just visit the personality section of this chatbot testing guide and you’ll see some questions that will show you the branding items you could actually experiment with. See the screenshot below for inspiration:
Once you know what element/elements you’ll test (based on your hypothesis), determine the length of your chatbot experiment and the sample size.
Tools for Calculating the Duration and Sample Size of Your Experiments: Here are some of the best CRO tools to calculate the ideal sample size and duration for your chatbot experiments.
Convert’s A/B Test Duration Calculator: Just input your data into this calculator, and you’ll know how long your chatbot test or experiment should run. Convert’s A/B Test Duration Calculator: Just input your data into this calculator, and you’ll know how long your chatbot test or experiment should run.
Step #3:Learning from Experiments
Once your experiment is over and the data is in, it’s time to analyze your findings.
Usually, there are just three outcomes to any optimization experiment, including the ones you’ll run for your chatbots. These are:
- The control loses. Here, your hypothesis is validated and your change brings a positive impact on the numbers. An example of such a result would be getting 1000 opt-ins instead of 890 by changing your chatbot’s profile image from a cartoon to a mascot.
- The control wins. Here, your hypothesis needs to be rejected as your change brings a negative impact on the numbers. For example, the new mascot profile picture getting way lower signups than the regular cartoon picture.
- The test is inconclusive. These are usually the most common and often the most frustrating outcomes because you don’t get statistical significance to have a clear winner.
So once you’ve your test’s results, you need to go back to step #1 of your experimenting: the hypothesis step.
Either you can start a new experiment to test a new hypothesis or go with iterative testing, which means going back to a hypothesis that didn’t get validated (either because of a losing or an inconclusive test), improving it, and then re-running the test.
When doing iterative testing, make sure you spend time into understanding why your test failed in the first go.
Was it choosing a wrong test segment?
Was it a bad hypothesis all along?
Were your test logistics bad? The idea here is to learn all that you can from your winning, losing, and even your inconclusive chatbot experiments because that’s how you optimize — with continuous learning.
Wrapping it up …
If you’re more tech-savvy, you can take your chatbot experiments to a whole new level by testing with the content you feed your chatbot (or its “knowledge base”).
Or, you can also try a different learning algorithm.
Chatbots are here to stay and as machine learning matures, they will be up-front and centre, acting as the first touchpoint with large segments of your prospects.
It just makes sense to get on-board with A/B testing their performance.
Do you use chatbots to assist with your sales, marketing, or support work? If you do, are you running any experiments to see if your chatbots can do better? Tell us in the comments!