When a lookalikes recommendation is created, your catalog is analysed. The process groups similar products together. If you were a winter sports retailer, products could be grouped by tags such as 'snowboard', 'female', 'burton', 'red', and so on. Each attribute is scored; the higher the confidence score, the more likely the tag is accurate.

Next, a contact’s purchase and browsing history are analysed. Products with a suitable similarity score are chosen. Each contact is then given a ranked product set unique to them.

Predictive recommendations automatically generate a fallback. For contacts where no prediction is possible, the fallback displays instead. Typically a best seller recommendation, which you can edit as needed.

If predictive recommendations aren't enabled on your account:

From the side panel, you can request the feature is enabled on your account for an additional monthly charge. Your data is reviewed first, then you'll be contacted to confirm the feature is ready to use, or to discuss any data dependencies that need resolving.

Predictive recommendations include a 30-day free trial.

Review the data dependencies for this recommendation type before you start.

This model works best when your product data is high quality and detailed.

Catalogs with a high number of products bhigh-qualityut low order volumes don't produce optimal results.

Order data is one indicator of contact preference. Using Web behavior tracking improves the accuracy of the machine learning model.

Lookalikes uses content-based filtering, which has a narrower scope than best next, which uses collaborative filtering. This makes it particularly effective for niche purchasers.