Multivariate tests (MVT) let you change several elements on your site and determine which combinations of changes performs best.
Creating a MVT is similar to creating a standard A/B test except that you define multiple sections in a MVT. You can think of each section as a self-contained A/B test that modifies a single element, like a button or image. By creating multiple sections, you can test these elements at once.
Each white card represents a section, and you can create multiple variations within each section. As you build variations, Optimizely automatically generates all the combinations of these sections. Each combination pulls a variation from each of the sections you’ve already defined and rolls them together into a composite experience for your visitor. The combinations included in your MVT are visible in the Combinations tab.
The combination names are always made up of a series of capital letters. Each of these letters represents a variation from a particular section. For example, a MVT with three sections would have a combination named ABA. This combination would consist of the first variation of the first section (A), the second variation of the second section (B), and the first variation of the last section (A).
Below each combination name, you can also find a clear description of the different variations included in it. For example, in the image above, the combination labeled ABA is testing a combination of a specific headline (“Text A”), a particular text color (red), and a particular product shot (shoes).
Traffic allocation in MVTs: Full factorial vs partial factorial
Optimizely offers two different approaches to allocating traffic in multivariate experiments: full factorial or partial factorial.
In the full factorial approach, Optimizely tests every combination of variations against all the other potential combinations, allocating traffic on the basis of section variations. However, this affects the amount of traffic delivered to each combination.
Optimizely multiplies each of the variations' percentages together to determine the percentage of traffic allocated to a particular combination. For example, in a two-variation combination, if Variation 1 has a traffic allocation of 50% in section 1 and Variation 2 has an allocation of 25%, the total traffic allocation for that specific combination will be 12.5% (.50 x .25 = .125).
The advantage of the full factorial approach is that it is more rigorous. However, one drawback is that it requires higher levels of traffic to generate a significant result. Another is that you might end up testing specific combinations that may not make sense to test at all: for example, you would probably expect that testing a blue button against a blue background might not be a successful experiment.
In cases like this, you can use a partial factorial approach. Partial factorial lets you allocate traffic at the combination level, so you can completely exclude specific combinations from your experiment simply by setting their traffic allocations to zero. (By contrast, full factorial traffic allocation happens at the section level, making it impossible to prevent specific combinations from appearing.)
You can also set up your own test layouts, like Taguchi templates that mix variations in specific ways that ensure equal representation, but do not require all combinations to actually be generated. The advantage of partial factorial is that you can test far fewer combinations than full factorial, while still enjoying many of the same benefits.
Future releases will also enable auto allocation of traffic between combinations, so that you can find the most effective combination even more quickly.
Create a new multivariate test
The following example shows how to create a new multivariate test, with step-by-step instructions below:
Go to the Experiments dashboard.
Click Create New...
Select Multivariate Test from the dropdown. The New Multivariate Test modal window appears.
In the Name field, enter a name for your experiment.
In the Experiment Description box, optionally provide a description of this experiment.
Set up targeting to define where the experiment will run. From the Target By dropdown list, select either URL or Saved Pages:
If you choose URL, enter the address of the target page in the text box
If you choose Saved Pages, select the target page from the list of pages.
Add the audiences you want to include in your experiment.
You can combine multiple audiences in your experiment using AND and OR conditions. For example, you can target the experiment to visitors who qualify for the "Social Butterflies" audience AND the "Luxury Travelers" audience; a visitor must qualify for both to be included in the experiment. Or you can target the experiment to ("Social Butterflies" AND "Luxury Travelers") OR 5x Buyers. In this case, a visitor only needs to meet one set of qualifications or the other to enter the experiment.
Name each section and add as many variations as you need for each of them.
The number of possible combinations in a MVT is capped at 64. This works out to, at most, six sections with two variations each (2^6, or 2x2x2x2x2x2). You should also bear in mind that as the number of combinations in a MVT increases, the amount of traffic going to each one decreases, and results take longer to achieve.
Determine whether your experiment will use a full factorial or partial factorial approach, and to set the traffic allocations for each variation or combination. Click Traffic Allocation; the experiment's Traffic Allocation modal will appear.
Select either full factorial or partial factorial from the Multivariate Testing Mode dropdown.
Adjust the traffic allocation as needed.
Click Save to save your changes.
Once your experiment is set up, you can access your variations from either the Sections or Combinations tabs. The Sections tab displays a complete list of the sections you created for your experiment, along with all the variations contained within each one. Just click on the name of the variation or the Edit button to view and update the variation within the Visual Editor.
If you would prefer to simply view the actual outputted combinations themselves, you can do so from within the Combinations tab. Just click on the name of the combination you want to view to open it in the Visual Editor.
Results for an MVT work like a typical experiment, with data gathered for each of the combinations.
If you’re working with a large number of sections or variations in your MVT, each combination will likely take longer to reach significance due to the lower traffic volumes each one receives. With section rollups, you can collapse a MVT into a single A/B test on a single section. This way, you can get a sense for effect sizes and significance along that section, even when traffic is spread thin. See Section rollups in multivariate tests.
Multivariate testing FAQs
- Can I link MVTs to ideas within Program Management?
Yes! Optimizely supports MVT in program management.
- Are MVTs available in Full Stack?
Currently, MVT is only available for Optimizely Web.