Engagement Levels
Hook’s customer health score is called an engagement score. This is updated weekly for all accounts. This score calculates the likelihood of renewal for an account. A higher score means a higher likelihood to renew.
Hook’s engagement score refers to renewal likelihood, rather than an account's actual engagement with the product. It’s called an engagement score because engagement is a key factor for the score.
Depending on the value for the engagement score, an Engagement Level is given for each account. These Engagement Levels are:
Very High
High
Medium
Low
Very Low
An account is considered at risk of churn if it has a Medium, Low or Very Low Engagement Level.
Renewal Likelihood and Potential Renewal
Renewal Likelihood %
This represents the percentage likelihood that this account will renew in full, based on typical customers in the past with the same Engagement Level.
If an account has a Medium Engagement Level and historically we’ve seen accounts in that bucket renew in full at a rate of 83%, that will be the renewal likelihood.
Renewal likelihood is calculated by Engagement Level. All accounts with the same Engagement Level will have the same renewal likelihood.
In this example, all Medium accounts will have a 83% renewal likelihood.
Potential Renewal $
This value represents the typical spend at renewal if this account does renew. It is calculated by multiplying the current ARR by the typical percentage increase (or decrease) in spend by accounts with the same Engagement Level. This is based on historic data from similar customers.
Calculating Hook’s Engagement Score
Hook’s engagement score uses data from previous customers in the past two years that have churned or renewed. We measure their behaviour leading up to a renewal event, identifying the likelihood that each behaviour results in a renewal.
A renewal event occurs when an account reaches the end date of their subscription (contract), at which point a customer will either renew or churn.
We measure customer behaviour using various data points. This includes:
Metrics for insight, which are helpful for informing CSMs:
Support tickets
Time as a customer
Metrics that are actionable, where CSMs should speak to their customer. These turn into suggested actions:
Product usage
User activity
Using the patterns in data from previous customers, we build a machine learning model that predicts the likelihood that existing accounts will renew.
Hook’s model will include account downgrades that are less than 75% of the last subscription as churned accounts.
Model Accuracy
Hook model accuracy is measured using recall (churn accuracy) and precision (churn inaccuracy).
Measurement | Definition |
Recall | The amount of actual churn that the model correctly captures. This could be logo recall (number of accounts churned) or value recall (ARR churned). |
Precision | How often the model’s churn predictions are correct. |
When building a model for new customers or updating existing models, we calculate logo and value recall at intervals from 180 days before a renewal event. The accuracy of the model is then weighted across these different points in time.
Hook's engagement models typically achieve 75% accuracy in predicting customer churn. 70% is widely considered best in-class for our industry*.
Model accuracy is reviewed one month after the initial build. These models are then updated each quarter for our customers.
Models often over-predict churn, so Hook’s engagement score can be seen as cautious.
Related Article: Customer Distribution: See Your Forecast
Why Hook is different to other Customer Health Scores
Hook’s engagement score is entirely data driven.
We analyse raw customer data, then use machine learning to generate predictions for which customers will churn.
Most CS health scores are manually configured and use CSM and customer sentiment, making the scores subjective. Identifying the most important metrics is up to guesswork.
We leave that behind at Hook. Our engagement score uses science instead of gut feel to identify churn risks.
Our models are updated once a quarter, to ensure our health scores are performing at best-in class levels for churn prediction.
*Disclaimer:
The reported accuracy of approximately 75% reflects the model's performance during internal testing and may vary depending on specific use cases or data inputs. While we strive to improve model quality continuously, predictions generated by this system are not guaranteed. Users are encouraged to review outputs and apply human judgment where appropriate.