Explainer: Under the hood of deal likelihood scores

Who's this for: Anyone forecasting with Gong

Where: Deals

Plan: Working with Gong Forecast requires a Forecast seat.

How accurate are we?

We use industry-standard AI-model performance metrics to measure the accuracy.

Model precision

We measure our model precision for positive predictions by asking, “Out of all deals we gave a high score, how many predictions were correct?”

The results show we have high precision, which indicates a low rate of false positives.

On average, Gong AI is 21% more precise than sales reps in predicting winning deals, as early as week 4 in the quarter. Our results are based on a test set of thousands of deals.

Naturally, no score is perfect. There may be activities on a deal that we don’t track or know about that affect the score positively or negatively.

About the AI model

We use a multi-model AI system that’s created as follows:

Step 2: Customized for your business

Next, the base model is customized for your business. By looking 2 years back at all the deals closed in your company specifically, we understand how variables that are unique to your business affect typical deal performance in your world.

In this step, we train the model on variable signals such as:

How long deals typically stay in each stage
Their typical level of activity
When and how much conversation is around pricing, legal, procurement, and so on

Step 3: Deals scored continuously

To ensure you get an accurate score every single day, the AI runs on both models every day. This ensures Gong AI is always up to date on the latest developments in each deal and also the model uses the past 2 years of information, including what happened yesterday.

Factors weighted dynamically every day

Our model doesn’t use a rule-based type of algorithm where we precondition the weighting, but rather a machine learning one where the weighting is part of the prediction and it changes for every deal. Our algorithms tune the weighting daily based on objective deal outcome data to achieve improved accuracy without bias. 50% of signals are based on conversation intelligence, and the remaining 50% are drawn from activity, contacts, timing, and historical data. The exact weighting given to any signal can change on a daily basis, per deal, depending on the data patterns learned by the AI model.

Reality-based signals

The deal signals we consider fall into the following categories:

Category	Examples
Conversations - Gong’s “special sauce”	Mentions of legal and pricing Red flags found in emails Competitors brought up Deal warnings Call interaction stats
Activity - deal-activity association	No next meeting scheduled Prospect ghosting No. of days since last call, email, or meeting
Contacts - enriched contact information	Number of contacts on deal, and whether that’s enough Who’s on the deal, and whether they have the power you need
Deal progression - what's behind or ahead of schedule	How fast the deal progresses Changes to the deal amount Changes to the close date, compared to won deals Deal age Proximity to close date
Historical performance - comparing each deal to how your past deals behaved	Time in stage Rep win rates Stage win rates Forecast category win rates

Score ranges indicate the likelihood of winning

To make it easy to search for and view deals that are or aren't doing well compared to others, we group the scores by three indicators:

Likelihood indicator	Score range
Low	0-35%
Fair	36-74%
High	75-100%

Menu

In this article

How accurate are we?

About the AI model

Reality-based signals

Score ranges indicate the likelihood of winning