(note: the following changes only impact Enterprise customers and existing Sales Hub Professional customers )
We’ve completely re-engineered and modernized the machine learning models that power lead scoring in your HubSpot portal, resulting in clearer presentation and 4x more accurate scores.
Okay, so what’s actually changing?
- A new property, “Likelihood to Close,” has been added to all your HubSpot contacts. Any contacts that have not yet closed as customers will have a value between 0 and 100, which represents their percent to close. This percentage is calculated using a slew of information about the contact, its associated company (including enriched HubSpot properties like employee size and industry), and engagement data from HubSpot CRM and HubSpot Sales Hub.
- Another property, “Contact Priority” has been added to all your contacts. This property groups contacts into four buckets, based on their “Likelihood to Close:” very high, high, medium, and low. Previously, this property used the categories, “Tier 1, Tier 2, Tier 3, and Tier 4”, and we’ve updated the labels for clarity based on user feedback.
- With the improved model behind “Likelihood to Close,” the existing properties “Predictive Lead Score” and “Lead Rating” will no longer update starting in the near future --- the specific date is still to be determined, and we’ll update this post as soon as it is. For the short term, this property will continue to exist in your portal with historical scores.
Why make a change?
When we first launched predictive lead scoring in 2015, it was HubSpot’s first application of machine learning. Since then, we’ve improved internal processes and adopted techniques and technologies that have been developed in the rapidly evolving field of machine learning research. As a result of the changes, the new property is more than four times more accurate than the original property.
In addition, the original predictive lead score required thousands of contacts with clean, accurate contact properties in order to produce predictive scores. In full transparency, many HubSpot customers lacked either the quantity or the quality of data to train a model that would accurately score contacts. The new model uses HubSpot Insights datasets to supplement data within a portal to make even more accurate predictions from the very first contact added to a portal.
What changes can we expect, from the “old” property?
- No predictive scoring dashboard. The newer technologies (see: neural networks) produce much more accurate scores, but because they aggregate data from so many places, they don’t provide the ability to pinpoint the effects of specific properties on specific contact scores. As a result, there’s no dashboard that breaks down the new score properties.
- Different numeric scales. Previously, the predictive lead scoring model was scaled around 50, so scores tended to cluster there. The new model (and associated “likelihood to close” property) use the actual probabilities as scores. With that in mind, great contacts will have higher scores than they would’ve under the previous model, and bad leads will have near-zero scores.
- Contacts that have already closed as customers are not scored. Once a contact closes as a customer, the ML model no longer predicts whether they’ll close as a customer, for obvious reasons.
The new predictive lead scoring properties are now available. Starting on September 5th, all new Marketing Hub Enterprise customers will only have the new and improved properties, and will not have the predictive scoring dashboard. On March 1st, the predictive dashboard will be removed from all HubSpot accounts.