Discover what Attention Insight MCP can do for your workflow ?
Are there discounts available, or do I need to whisper the magic word?

Attention Insight Technology

Our AI platform generates eye-tracking heatmaps with up to 96% accuracy - without human participants. Trained on millions of real fixations, it learned to perceive designs and images the way people do.
60 sec to get a result
No participants needed
Up to 96 % accuracy
5.5M+ eye-tracking fixations

Eye tracking studies

Eye tracking is a process of measuring eye movement, determining where the person’s gaze is directed.

The result of an eye tracking study is an Attention Heatmap that shows how a group of participants viewed an image or video. Such eye movement data was needed to train our deep learning algorithms.

To carry out the eye tracking studies, we hired a neuromarketing lab which is a member of the International Neuromarketing Science and Business Association (NMSBA).

Data set

The Attention Insight algorithm is trained with approximately 5.5+ million fixations and 550+ million gaze points from eye tracking studies. The image datasets we use are both open-source and proprietary.

Data set statistics:

  • Participant attention duration: 4 seconds;
  • Average gender distribution: 58% women and 42% of men;
  • Average participant age distribution: varies from 7 years old to about 60+ years old. However, most participants fall into the 21–30 age bracket.
  • Participants are from the USA and Europe

Algorithm training

Our heatmaps are generated by a deep learning algorithm called Convolutional Neural Network (CNN). It is a computing system that has an architecture inspired by the biological brain and mimics how neuron layers work.

A CNN consists of multiple layers of nodes that are connected with different weight connections. These weights determine how much one node impacts the following node.

Before the training, a CNN has random weight connections resulting in inaccurate heatmaps. The difference between the generated heatmap and the “ground truth” (actual eye tracking heatmaps) is called the error. During the many training cycles, these weights between layers are adjusted so that the error is reduced. After the training, our generated heatmaps match the “ground truth” closely.

We regularly add new data sets to make sure that our algorithm meets industry standards and stays the most accurate algorithm in the market.

Accuracy validation

To measure the accuracy of our generated heatmaps, we submitted our results to the Massachusetts Institute of Technology (MIT) / Tuebingen saliency benchmark.

They sent us 300 test images, and we sent them back our model results on their testing data set of those 300 images.

After evaluating our results, MIT scientists concluded that our heatmaps match actual eye tracking heatmaps with 92.5% accuracy for general images. Across all types of designs, our heatmap accuracy is up to 96%

Real vs Predictive Eye Tracking

Predictive eye tracking uses advanced deep-learning models trained on millions of real eye-tracking data points. This allows designers to understand how viewers will visually engage with their content – instantly, affordably, and with reliability comparable to traditional eye-tracking studies.

Questions we sometime get on support
From more than 1,000 conversations per year, some of the questions we receive are about our technology.
Is it like Hotjar?

No. Attention Insight is a different type of tool than Hotjar – we do not use mouse-tracking data. Our platform is based on predictive eye-tracking, which relies on AI to simulate where users are likely to look, rather than tracking their actual mouse movements. However, you can combine our predictive insights with tools like Hotjar to get a more complete picture of user behavior.

No. Our platform does not use your webcam or record any video of you. All predictions are generated using AI models trained on pre-existing eye-tracking data – no live tracking or camera access is required.

No. We do not use your uploaded images or videos to train our algorithms. Our models are trained exclusively on dedicated eye-tracking datasets that were specifically collected for that purpose. Your data is used only to generate your results and is never added to any training set.

Yes. Our predictions are 94–96% accurate when compared to attention heatmaps from real eye-tracking studies. This high level of accuracy has been validated through extensive research and independent benchmarking.

No. Predictive heatmaps are a fast, scalable proxy for early-stage visual attention – particularly useful in the first 3–5 seconds of exposure. They show where people are likely to look, but they do not capture user intent, goals, task completion, comprehension, emotional response, or actions.

Use predictive heatmaps before launch to quickly catch visibility and hierarchy issues, validate layout decisions, and prioritize hypotheses. Then use usability testing, analytics, or A/B testing to validate real behavior and performance with real users and real traffic. Think of predictive heatmaps as a triage tool: they remove obvious visual issues early so deeper testing can focus on what matters.

It models initial visual attention – what people are most likely to notice in the first few seconds after seeing a design. The predictions reflect early visual processing, driven by factors like contrast, size, and visual hierarchy. They do not account for actions, reading, scrolling, or task-driven behavior – for that, use usability testing or analytics alongside predictive attention.

No. High predicted attention means an element is likely to be noticed, but it does not guarantee understanding, persuasion, or action. An element can attract attention and still be unclear, misunderstood, or create friction.

Predicted attention shows what is seen, not how it is interpreted. For this reason, attention insights should be evaluated alongside design objectives such as clarity, message comprehension, and ease of action, rather than treated as a standalone success metric.

No. Predictive eye tracking does not generate separate heatmaps for specific demographic groups. The predictions reflect general human visual attention patterns, not behavior segmented by age, gender, culture, or other demographic attributes. This ensures privacy, avoids demographic bias, and keeps the predictions focused on universal early visual processing.

In the first few seconds of viewing, attention is largely driven by shared human mechanisms such as contrast, size, and visual hierarchy. Demographic differences tend to influence later interpretation and decision-making, rather than what initially captures attention.

Yes – to a degree. The model is optimized for left-to-right (LTR) layouts, which dominate Western web design. If you’re designing for right-to-left (RTL) languages or culturally distinct visual flows, additional validation via localized testing is recommended.

While these differences don’t significantly affect early visual attention (which the model predicts), they can influence how users navigate and interpret a design after the first few seconds, once top-down processes like reading habits and cultural expectations come into play.

No. Predictions are generated by AI models trained on aggregated eye-tracking data and do not rely on personal data, demographics, or individual user information. No real users are analyzed, recorded, or tracked when generating predictions.

News & updates
Subscribe to our newsletter
Days
Hours
Minutes
Seconds
Subscribe to the FIGMA HERO monthly plan and get 40% off with code AT40 for next 12 months. Offer ends September 30 at 23:59 (UTC+2). How do I apply discount?