ZipRecruiter Uses Predictive Signals to Bring Greater Precision to Employer Acquisition

Design an employer LTV model built to price every employer accurately

The first step was deciding what to predict and for whom. Voyantis worked with ZipRecruiter's team to land on n-week revenue as the prediction target, the point at which employer behavior had revealed enough differentiation to price each employer accurately at acquisition.

To capture the variance in employer value, the model drew on thousands of data points that reflected hiring intent, organizational sophistication, and other indicators that play into the hiring lifecycle. Together, these insights provided a picture of each employer's value that no single conversion event could provide.

Unlike a standard LTV model that smooths variance in pursuit of average accuracy, Voyantis designed ZipRecruiter's model to amplify the spread of predicted employer values. This gave Google a signal rich enough to learn from the individual worth of each employer rather than an average.

Engineer signals for Google's unique learning logic

ZipRecruiter had learned firsthand that a predictive model alone won't drive better campaign performance. Google's algorithm is built for deterministic signals, actions that actually happened. Predicted employer value is inherently probabilistic, so getting it into a format Google could actually learn from required a dedicated layer of signal engineering.

Voyantis's AI decisioning platform handled the full signal engineering layer, including:

• Timing signal delivery to Google's optimization window, with predictions delivered as early as one hour post-registration and updated as behavioral data accumulated, ensuring consistency between early and mature signals.
• Holding signals where early behavioral data isn't strong enough to justify a bid, then revisiting them as more data accumulated. An employer who posts a job within the first hour reveals different early intent than one who only completes registration and goes quiet. The model allows confidence to grow rather than writing them off entirely.
• Capping outliers so high-value employers wouldn't distort the algorithm.

Keep the system calibrated as the business and auction evolved

Bidding changes who you acquire, platform logic shifts without documentation, and data pipelines fluctuate, any of which can degrade signal quality before it shows up in metrics. Voyantis maintained the system continuously, with every layer tracked automatically. When something looked unusual, automated monitoring flagged anomalies immediately and activated responses to protect signal continuity.

Before deploying updated models, Voyantis ran rigorous validation, measuring prediction accuracy and ranking performance to confirm the new model better differentiated high-value users, then deployed the update in a way that preserved Google's accumulated learning. The Voyantis platform continuously monitors model health, intervening deliberately to keep the signal reaching Google calibrated and getting smarter with every cycle.

For the ZipRecruiter team, that means a system that runs itself, so they can stay focused on growing the business rather than maintaining infrastructure.