Get started with Optimizely Web Recommendations

  • Updated
Optimizely Web Recommendations is not currently included in the packaging for new customers. It is not an out-of-the-box solution and requires great effort to configure. If you are looking for personalization solutions to use with Optimizely Web Experimentation, our Customer Success Managers can help review our portfolio of tools to find the right fit for you.
This topic describes how to:
  • Show visitors content or products they may be interested in based on their browsing behaviors
  • Choose the best algorithm for your recommendations
  • Test and iterate on your recommendations
  • Measure the impact of your Recommendations campaign

Optimizely Web Recommendations makes it easy to add recommendations to any page on your site and measure the impact with Stats Engine. You can introduce visitors to products and content based on your chosen algorithm and measure results.

Here is what you can do with Optimizely Web Recommendations:

  • Experiment with multiple recommendations (layouts, algorithms, headers, and placement) at the same time
  • Fold recommendations into other visual changes that you test in your experiments
  • Create Optimizely Web Recommendations campaigns
  • Set traffic allocation for a campaign to show a certain percentage of visitors the original variation or the variation with recommendations
  • Target your campaigns to desktop or mobile visitors
  • Modify the algorithm, layout, placement, and header text of your recommendations
  • Measure the impact on your Results page and segment by mobile and desktop visitors
  • Automate recommendations with algorithms for hands-off ROI
  • Use Personalization to take manual control of messaging for bigger audiences or special promotions
  • Target recommendations to specific audiences, available with Personalization
  • Configure different recommendations (layout, algorithm, headers, placement) for specific audiences and deliver them with Personalization
  • Use the functionality that Personalization offers for managing, running, and analyzing campaigns

However, before you jump into building your first recommendations experiment, you will need to take care of a couple of tasks first. This article walks you through: 

  • The pre-setup process;

  • Creating a catalog and a recommender for your experiment;

  • Configuring the Recommendations algorithm and deciding where it will display;

  • Previewing and verifying your results.

Read on for more details.

Pre-setup

Before using Optimizely Web Recommendations, you must contact your Customer Success Manager. They will help with the initial setup, and with getting you set up with the services bundle you will need:

  • Optimizely Web Recommendations platform training

  • Optimizely Web Recommendations extension training

  • Optimizely Web Recommendations experiment strategy training

  • Design & build a recommendations experiment test plan

  • Build & launch a recommendation experiment or campaign

  • Custom results review

  • Project management oversight

  • Build & configure an extension for recommendations (optional)

These pieces of training should take about 30 hours to complete. After that, you will be ready to build your own recommendations, experiments, and campaigns.

Create a catalog

Your recommendation experiment will need a list of items to recommend to your visitors. These items are stored in a catalog. When you create a recommendations experiment, Optimizely will ask you to assign a catalog, so you will have to create it first.

What goes into a catalog? That depends on your business. If you are running a retail site, this might be all the items you sell. For B2B companies, this might be a list of blog posts or help articles.

To build a catalog, follow these steps:

  1. Navigate to Implementation > Catalogs and click the Create New Catalog button.

recommendations_1.png

  1. The New Catalog window appears. Give your catalog a name and enter a description.

  2. Add the events and related tags you want to include in this catalog in the Catalog Events box.

 

recommendations_2.png

Optimizely Web Recommendations uses tags to determine which fields from your catalog will be displayed. You should create a tag for each of those fields in this step.

  1. Once you have added all the events to your new catalog, click Create Catalog.

  2. To download a CSV copy of your catalog, or to archive it, click the … button for the catalog you are interested in. It will take at least 24 hours for Optimizely to assemble it for download.

recommendations_3.png

Optimizely runs a validation check when you create your catalog to ensure the items listed in it are still active. You may need to whitelist the user agent “Optimizely Recommender” to allow this validation check to occur; otherwise, Optimizely will assume none of the items in your catalog are accessible. For this reason, you should turn off validation completely if your site is behind a login.

If your site uses custom pricing for each customer, do not collect any field that may appear differently to different users viewing the same item. Otherwise, the catalog will show inconsistent information for these items.

If you have questions about this, contact customer support.

Select the appropriate menu item and follow the prompts.

Set up a recommender

In addition to a catalog, your recommendation experiment will need a recommender, which is the combination of logic, filters and user inputs that tells Optimizely what items from your catalog it should recommend.

You should already have set up your catalog in a previous step. To set up a recommender, follow these steps:

  1. Navigate to Implementation > Recommenders and click the Create new Recommender button.

  2. The New Recommender window will appear. Name your new recommender and type in a description.

Make sure the name you give your recommender is unique since all your recommenders—even the ones you created for other catalogs—will appear next to each other when selecting one for your experiment.

  1. Select the catalog you want this recommender to use from the Catalog drop-down menu. The recommender will not automatically use all the events in the catalog; instead, you will attach specific events to this recommender to use later.

Once you have selected a catalog for your recommender, you cannot change it to a different catalog. If you want to do that, you will have to start over by creating a new recommender.
 

  1. Select an algorithm for your new recommender to use from the Algorithm drop-down. Optimizely gives you four algorithms to choose from.

recommendations_4.png

  1. Choose the specific catalog events you want to use as inputs to your algorithm.

  2. If the algorithm supports boost events, add the ones you want in the Boost Events box.

  3. Optionally, you can add a filter to customize your recommender further. To learn how to create one, just click here.

You can change the algorithm after you have set up the recommender. However, doing so will remove all your events and filters, and you will have to start over.

  1. When you are done configuring your recommender, click Create Recommender.

  2. To create an extension to display your recommendations or to archive your recommender, click the … button next to the recommender's name. Select the appropriate menu item and follow the prompts.

 

recommendations_5.png

Create a filter

Optimizely gives you the ability to add powerful filters to your recommenders. When you include a filter, Optimizely will apply it to any catalog items returned by the algorithm. Items that do not meet the filter conditions are not included in the recommendation's results.

This section provides a description of how to use the Filters interface to build a new filter from within the New Recommender window. 

filter_1.png

  1. Decide whether you want your recommender’s default behavior to be exclusive or inclusive. “Default behavior,” in this case, describes what the recommender will do if you do not add any filtering conditions. For example, if you were to check the Exclude all recommendations by default box and then not include any conditions, your recommender would return no results: its default behavior was set to exclude all results, and it did exactly that. Likewise, if you were to leave the box unchecked without adding any conditions, you would get a list of every item in your catalog.

    Your filtering conditions will determine which results appear or are excluded from your recommendations; you will set those up in the following steps.

  2. Decide whether you want this filter to include specific items or to exclude them, and make your selection here.

  3. If you are building a filter with multiple conditions, specify whether you want it to apply to items that meet all the conditions you set out or if it should apply to items that meet any of them.

  4. Here, you can select the values you want the filter to evaluate. These values will reflect the names of the fields on your items in the catalog, which are acquired from the tags. Choose from any of the values listed, or add a custom value not included in that list.

filter_2.png

If you select custom, you can type in any literal fixed value.

filter_3.png

If you choose a tag instead, you can choose which item the tag refers to. The key item is the page the visitor is viewing, while the recommendation item is the item returned by your recommender.

filter_4.png

This only applies when you are using the co-buy or co-browse algorithms. The collaborative filtering and popular algorithms do not support the use of a key item, so this drop-down will not appear in those cases.

  1. Select the comparison operator your condition will use. Do you want these values to be equal to each other? Should one be larger than the other?

  2. This allows you to add another row of conditions to your filter.

filter_5.png

How this row is used by your filter depends on your selection in step 3. If you chose any, only one condition row must be evaluated as true to return a match. If you chose all, then both (or however many conditions you add) must be evaluated as true to return a match.

  1. The Add condition group button will add an entirely new condition group to your filter.

filter_6.png

The condition groups are evaluated to look for a match. If there are no matches, the recommender takes the default action specified in step 1.

When you are done building your filter, return to step 8 in the Set up a recommender process above.

Preview your results

Once a recommender is set up, you can preview the results it will generate before creating your experiment. If those results are not what you expected, you can refine the recommender until the results are correct.

Changes to filters or algorithms are not immediately visible. Your results may take up to 24 hours to update.

To preview your results, follow these steps:

  1. Navigate to Implementation > Recommenders and find the recommender you are interested in. Then click the Previewer button; the Recommender Previewer will appear.

 

recommendations_6.png

At the top, you will see a set of summary statistics for your site's traffic over the last seven days:

  • Partial recommendations, which are items that have fewer than ten recommendations, usually due to missing data

  • Full recommendations, which are items that have a full set of 20 recommendations each 

  • Total recommended items, which is the number of rows of recommendations calculated when the algorithm was last run

  1. In the Preview Item Recommendations box, enter an ItemID to preview the other items recommended to your visitors when they “view” that item. Note that this is useful for algorithms that make recommendations based on items the visitor has already viewed. For context-free algorithms—like Popular, for example—changing the ItemID will not change the output because the output is not tied to any specific item.

Recommendation algorithms

When you set up your recommendations, you can choose from one of four algorithms:

Algorithm name

Use case

Example

Co-browse

This algorithm tends to recommend items similar to those a visitor has recently viewed. For example, if a visitor has viewed several different tables during the current session, the algorithm will likely recommend other tables.

"Visitors who viewed this product also viewed these other products."

Co-buy

This algorithm works well for recommending complementary products to those a visitor has previously viewed. For example, if a visitor has viewed a table, this algorithm may recommend chairs to go with it.

"Visitors who bought this product also bought these other products."

Collaborative filtering

This algorithm serves recommendations based on the combined browsing history of the user and other similar users. This is often shown as “Recommended for You.”
Items recommended by this algorithm will differ for each person but will remain constant for the duration of any individual user’s viewing experience.

"You browsed similar products as this group of website visitors, and they tended to like them."

Popular

This algorithm helps showcase crowd favorites on the homepage. Use it to introduce best-selling items to new visitors.

"Items that other visitors viewed or bought most often."

Recommendations are generated every 24 hours so the catalog information may be up to 24 hours out of date.

Remember that the co-browse algorithm is sensitive to the order in which items have been viewed, while the co-buy algorithm is not; instead, it only asks if the items have been viewed by the same person.

Types of algorithms

A recommendations algorithm will always produce an ordered list of (item, relevancy_strength) pairs. These lists usually contain 20 results. Optimizely’s recommendation algorithms use the collective behavior of past visitors to arrive at their results. However, the input to each type of algorithm is different:
  • A visitor-based algorithm only cares about the visitor ID and will recommend based on each visitor’s specific viewing history. The results will be the same no matter which page the visitor is on; however, the recommendations will be generated specifically for each individual visitor. Examples are the recently-viewed and collaborative filtering algorithms. 
  • An item-based algorithm is only interested in the page the visitor is viewing. Results will be the same for each visitor but should vary from item to item. These are typically described as “similar items” or “people who liked this also liked.” Examples are the co-buy and co-browse algorithms.
  • A global algorithm does not care about the current visitor or page. These typically appear on high-level pages (for example, the home page) where they are described as “our most popular options” or “check out these best-sellers.” It can be a great place for new Optimizely Web Recommendation users to start, but since the results are not personalized, it may not be appropriate for all applications. Examples include the popularity algorithm, which displays items in descending order of views, clicks, purchases, or whichever event you use for your recommender.
Think of the input as a key to look up the output of a recommendation—the visitor-based algorithm will take a visitor ID as the input key, while an item-based algorithm will use the item ID. The global algorithm will not need a key at all.
 
Similarly, the output can be thought of as the target. In an item-based recommender, both the key and target will be item IDs; in a visitor-based recommender, the key would be a visitorID, and the target would be an item ID.
 
When you experiment and compare algorithms to each other, it only makes sense to compare apples to apples. It would be pointless to compare a co-browse algorithm to a popularity algorithm because they use completely different information. A more useful approach to experimentation would be adjusting the input events and the post filters.
 

Glossary of Recommendations terms

While you are probably familiar with most of the concepts and techniques used by Recommendations—like filters, algorithms, and unique IDs—there are a few that are not referenced by any other Optimizely feature and may be new to you:
  • Catalog: The collection of all items that can appear in your visitors' recommendations. 
  • Events: Optimizely uses visitor events, such as product page views, to construct the catalog. Additionally, events are used to generate the recommendations, serving as the signal used to score the different items in relation to the user or other items.
  • Item: Any unique product, piece of content, or other entity that can be recommended to your visitors, as well as all relevant metadata associated with it. For retail customers, items almost always correspond to specific products, but items can also be a website or URL in other contexts. Items are based on unique IDs; any catalog listings that share a unique ID are, by definition, the same item. Different pages may have different sets of tags and represent different item types.
  • Recommender: A combination of:
    • an algorithm,
    • the specific inputs to that algorithm, and
    • any business logic filters to be applied. 
Recommenders essentially take input from your visitors, run it through the algorithm and filters, and output a set of recommendations that are then displayed for your visitors.