Beyond Saliency: The use of AI models in mapping visibility and attention metrics

Whether it’s brand stretching, subtle optimization, or large-scale innovation, testing a product’s packaging before it hits the shelf or e-commerce platform is essential. This necessity calls for packaging screenings and packaging health check testing that can be time-demanding and costly.

AI-driven Predictive Eye Tracking solutions are designed to tackle these challenges. While speed is a common advantage among these tools, their true value lies in the accuracy of their predictions, which heavily depends on how their AI models are trained.

However, it’s important to recognize that even the most advanced AI-trained predictive eye tracking has its limitations. For more complex, context-rich situations, “real” Eye Tracking remains a gold standard, while well trained Predictive Eye Tracking is ideal for quick pack screenings and health checks.
 
Furthermore, adopting a mixed-method approach that incorporates methodologies like Max Diff, Surveys, and Click Tracking provides even more robust data and deeper insights to bring to the “research table.”

The industry is ready for a Predictive Eye Tracking saliency upgrade

The core of any AI Predictive Eye Tracking methodology lies in machine learning principles—or, simply put, in how the algorithm is trained.

Saliency-based AI models for Predictive Eye Tracking are designed to analyze visual elements such as contrast, color, shape, and spatial layout. These models excel at quickly identifying visually striking areas; however, saliency map is not enough to accurately predict metrics for packaging evaluation.

Real eye-tracking studies reveal how a group of participants viewed an image—identifying areas that drew the most attention and for how long, yielding critical KPIs such as Visibility and Attention. These KPIs are grounded in real respondent eye movements, precisely capturing where their gaze is directed. So how could real eye tracking principles be a guide mark in enchasing saliency trained AI models?

Our proprietary AI model is trained on more than 11.5 million gaze points derived from real Eye Tracking studies. It is context-driven and trained on an extensive dataset of real eye-tracking data specific to Pack stimuli. This real eye-tracking data is sourced from our “data library," built over the past years of conducting eye-tracking studies.

To ensure well-rounded and accurate output, our Predictive Eye Tracking methodology combines several AI and ML regression models with our proprietary prediction model. This hybrid approach allows us to base predictions on actual human visual behavior while integrating saliency models that account for the physical characteristics of stimuli—such as contrast, shape, and color and how these elements interact.

Validation of AI models: How to strike close to the ground truth

While saliency can predict attention to some degree, it often ignores other important factors like emotional impact, consumer intent, or cognitive biases, which can be critical for a comprehensive analysis.  As mentioned above, we built our model to incorporate these crucial factors using experience and knowledge gained in the previous 10 years.

Although our studies are performed using browser-based eye-tracker, we validated our AI model twofold - against browser-based eye-tracker and high resolution, extremly precise hardware eye-tracker, as the later one is considered as ground truth.

The difference between the predicted KPIs and the “ground truth” is called the error. Multiple training cycles are conducted for adjusting our model and reducing the error. This results in our generated KPIs matching the “ground truth” very closely.

Additionally, we constantly feed our algorithm with new data set upgrades and features to assure that our algorithm meets the highest industry standards.

How Predictive Eye Tracking elevates pack screenings and pack health checks

When you're in the early stages of creating or redesigning a pack, do you find yourself asking questions like: Which pack stands out with the highest preference compared to competitors? Or wondering how to narrow down multiple design versions to a few key options that should be further tested with real Eye Tracking and refined through iterations?

If so, the most comprehensive pack screening results—providing both qualitative depth and quantitative precision—are achieved by combining Surveys with Predictive Eye Tracking, MaxDiff, and Click Tracking.

With Predictive Eye Tracking, even ad hoc “health checks” after implementing pack design iterations can be carried out in a cost-effective, scalable, and timely way.

In conclusion

In a market where competition is relentless, leveraging innovative methodologies to refine packaging design is no longer optional—it's essential.

To truly capture consumer attention and optimize packaging, AI models must be trained beyond saliency, incorporating deeper insights that reflect in-context human behavior.

Integrating Predictive Eye Tracking in testing pack screening solutions alongside methods like MaxDiff, Surveys, and Click Tracking ensures a more comprehensive, yet cost-effective and timely, understanding of consumer behavior. However, for broader, real-world scenarios, “real” Eye Tracking remains a gold standard.