AI Tech Paper

Introduction

The potential of artificial intelligence structured on neuroscience

Predicting consumer behavior poses technical challenges because declarative methods are unable to capture the subconscious drivers of choice. Traditional neuroscientific instruments involve high costs and turnaround times that hinder their application at scale. Artificial intelligence resolves these operational constraints by delivering neuroscientific metrics with the speed and scalability that modern marketing requires. The predictive capacity of an algorithm is determined by the specificity of its training data. Models built on generic information lack the technical calibration needed to forecast the precise emotional and cognitive responses of real consumers.

This technical document sets out the architecture of Databrain, a predictive platform developed to overcome the operational constraints of traditional research and of general-purpose algorithmic models. The operating ecosystem is built on the technological integration of Neurons Inc., backed by one of the largest proprietary consumer neuroscience databases in the world, curated through more than a decade of empirical research. This corpus of behavioral data records levels of attention, cognitive processing and emotional response to commercial stimuli across digital, out-of-home and print formats. The accuracy of these predictive models derives from this quantitative asset, which processes more than 100 billion behavioral data points from more than 100,000 participants tested every year.

This document provides an exhaustive breakdown of the science, the technology and the validation protocols that underpin this predictive infrastructure. It details the theoretical framework, the architecture of the predictive vision models, the methodology behind each metric and the multi-level statistical validation process. Its central objective is to give market strategists and analysts a technical understanding of the system, establishing its viability as an auditing instrument for creative effectiveness.

SCIENTIFIC AND COMMERCIAL ARCHITECTURE

Neuropredictive engine adapted to LATAM

Predictive engine validated in behavioral neuroscience

Predictive modeling applied to conversion

Technology development led by Dr. Thomas Z. Ramsøy (PhD in Neurobiology, CEO of Neurons).

Strategic direction led by Darío G. Treco (MSc in Neuromarketing, CEO of Databrain).

• Data processing

  Databrain audits the raw biometric data supplied by the underlying technology to build guidelines for advertising optimization.

• Empirical execution

  Integrating this algorithmic intelligence makes it possible to scale the evaluation of visual assets through repeatable, measurable protocols.

• Market efficiency

  Every compositional change is grounded in verifiable metrics of attention and cognitive load, ensuring operational efficiency and maximizing return on media investment.

• Scientific authorship

  Author of four reference works in the behavioral industry.

• Institutional validation

  Models developed in collaboration with researchers from prestigious universities.

• Specialization and academic leadership

  Design of curricula and technical training in Behavioral Science and Neuromarketing applied to market efficiency.

The relevance of implicit responses to advertising efficiency

The physiology of the consumer decision

Structural limits of declarative research

Decision-making operates through a complex integration of conscious processing and automatic physiological responses. Empirical evidence accumulated over decades, endorsed by the Nobel Prize in Economics awarded to Daniel Kahneman in 2002, shows that subconscious variables act as fundamental predictors of economic behavior and commercial choice.

Most market research relies on explicit data-collection tools (surveys, interviews or focus groups), which document the rationalization that follows exposure to the stimulus. Since consumer decisions are intrinsically driven by underlying neurobiological processes, a technical audit of advertising performance requires instruments capable of recording the influences that operate before the subject verbalizes anything.

Predictive modeling through implicit metrics

Implicit measurement methodologies expose dimensions of behavior that conventional declarative metrics cannot reach. Biometric technologies, such as fast response tests (FRT), eye tracking and neuroimaging, make it possible to record automatic, instinctive reactions to graphic assets, brands or products. These physiological readings capture latent biases and preferences without social desirability interference. In this regard, the clinical evidence documented by Knutson et al. confirms that activation in specific neural circuits predicts the purchase decision more accurately than the consumer's own self-report.

Predictive modeling through implicit metrics

The activation of neural circuits associated with anticipatory emotions, for example the nucleus accumbens (NAcc), precedes and conditions purchase decisions during exposure to the product and to the price variable.

The temporal cascade of the response

1. 50-100 ms

   Attention: the orientation vector

   The primary threshold. A stimulus lacking early visual salience is excluded from subsequent neurological processing.

2. 300 ms

   Cognition: the decoding vector

   The process of semantic assimilation. It measures the alignment between the intent of the message and the load processed by the viewer.

3. 400 ms

   Emotions: the motivation vector

   Affective reactivity conditions the valence and bias of the advertising experience.

4. 1000 ms

   Memory: the retention vector

   Encoding of the asset in long-term neural structures to ensure future retrieval and the consolidation of brand equity.

Data collection and processing protocols

Extracting commercial guidelines demands an unshakable data infrastructure. The multi-layer statistical protocols in place ensure that every metric collected, both in physical laboratory settings and through digital panels at global scale, is representative, reliable and strictly replicable. This methodological rigor ensures that corporate decisions are supported exclusively by empirical evidence.

Robust sampling and demographic representativeness

The predictive validity of any algorithmic model depends directly on the integrity of its sample. Rigorous selection methods are applied to certify that the physiological and behavioral variables recorded are an accurate reflection of the target market segment.

Scientific sample sizing

The operating standard sets a minimum threshold of 30 individuals per cohort (segmented by variables such as age range or income level) for high-fidelity in-person biometric studies, and 100 individuals for digital panels. These parameters derive from mathematical calculations of statistical power and effect size, ensuring the rigor and replicability of the conclusions drawn.

Targeted demographic representation

For experimental designs that require comparative analysis between isolated segments, the methodology scales proportionally. A study contrasting three distinct demographic groups requires a minimum of $3 x 30$ participants in the laboratory and $3 x 100$ in online measurement. Recruitment is structured around the client's commercial requirements, controlling for gender parity and strictly delimiting segmentation variables.

Structural contingency (over-recruiting)

As a mitigation mechanism against possible anomalies in the readings, exclusions due to calibration or dropouts during the testing phase, the protocols include a systematic 10% over-recruitment, safeguarding the mathematical validity of the final sample.

Relationship between sample size and data reliability

To certify the reliability of ratings on semantic associations, the system measures how error fluctuates as the sample grows. The curve shows that, as sample volume increases, the average standard deviation drops sharply until it stabilizes. A volume of ≈ 150$ participants marks the technical threshold at which the model reaches minimal residual uncertainty, ensuring high reliability in the commercial prediction.

Auditing and quality assurance for predictive models

Multi-layer validation protocols

Physical laboratory protocols (eye tracking and EEG)

Protocols for digital panels at scale

The remote testing infrastructure integrates algorithmic control measures that exceed the industry's conventional data-collection standards:

The infrastructure requires meticulous 7-phase quality control to isolate confounding variables and guarantee the absolute purity of the biometric signal collected:

Biological screening: Strict restriction on prior consumption of alcohol, caffeine or nicotine by test subjects, to ensure that unaltered cognitive and physiological responses are captured.

Real-time performance monitoring: Continuous auditing of motor reaction times to assess user engagement. Atypical metrics (response latencies outside the operating range of 0.3 s - 2.5 s in the FRT test) trigger system alerts. If a subject exceeds 50% of interactions outside acceptable margins, they are disqualified and their data discarded.

Technical integrity and calibration: Hardware and software diagnostics are run before each session. The system automatically excludes participants who do not meet accuracy thresholds (e.g., eye-fixation accuracy <80% or EEG signal stabilization failures).

Methodological refinement and bias mitigation: Test design calibrated to neutralize documented cognitive biases. To counter the "first-word latency effect," the system inserts an anchor control term ("Great") at the start of sessions, mathematically normalizing subsequent reaction times.

Signal quality audit: Recordings with insufficient capture (below 65% of eye-tracking data) or excessive noise (EEG signals exceeding 3 standard deviations from the population mean) are systematically isolated and excluded from the algorithmic model.

Respondent training: Mandatory practice tasks to consolidate understanding of the interface before data collection begins, optimizing the accuracy of the final metrics.

Regulatory compliance and structural anonymization

Databrain's operations with global biometric data through the underlying technology are subject to the most demanding international regulations (the European Union's GDPR and the ethical standards of the Declaration of Helsinki).

• Complete separation of biometric data from personally identifiable variables through single-log encryption.

• Strict non-disclosure agreements (NDAs) for all personnel with server access.

• Cryptographic destruction of temporary project files once the processing and delivery cycle is complete.

Visual modeling and attention prediction

Attention as a critical conversion variable

Visual attention operates as the primary, indispensable filter in the consumer decision-making process. In environments saturated with stimuli, the ability to direct focus toward specific vectors determines the commercial viability of the message. Precise quantification of attention patterns makes it possible to optimize the composition of content, increasing media effectiveness and reducing wasted capital on graphic elements that fail to be decoded.

Eye trackers are an excellent tool for measuring consumer attention.

In this award-winning ad, the brand goes unnoticed

The data architecture that feeds the model

The accuracy of the attention model operated by Databrain derives structurally from the data architecture that feeds it. The algorithm rejects the use of generic saliency repositories and is calibrated exclusively on the underlying technology's proprietary consumer neuroscience database, which consolidates the following operating metrics: • More than one billion eye-tracking data points processed. • Audits of an inventory of more than 8,000 advertising pieces, with a data expansion rate of 20% year over year. • Strict sample parameters of 30 to 40 participants per asset analyzed. • Segmented evaluation across 15 consumption contexts, covering digital advertising, social media, packaging, print and out-of-home (OOH) formats.

Fidelity of the fundamental metric (ground-truth data)

The base data matrix (ground truth) used to train the artificial intelligence is extracted exclusively from laboratory-grade eye-tracking studies. Biometric collection uses high-end eye-tracking instrumentation (Tobii X2-30, Tobii Pro Nano and Tobii Glasses Pro 2 systems) to ensure absolute fidelity in recording fixation points, saccadic movement trajectories and attention retention times. This rigor in collecting raw information, combined with strict evaluation protocols, ensures that the visual projections delivered reach an accuracy above 95% when compared against physical eye-tracking tests in the laboratory.

Computer vision model architecture

Deep learning computer vision model

The attention model is an end-to-end convolutional autoencoder with a sophisticated encoder-decoder architecture. This structure integrates state-of-the-art computer vision components, such as ConvNeXt and ResNet50, to ensure balanced, high-precision output rendering.

A critical methodological decision was to train the model from scratch, exclusively on the underlying technology's proprietary database. This ecosystem contains massive volumes of eye-tracking data collected specifically on consumer behavior.

This approach ensures that the model's feature extraction is fully optimized to predict commercial attention, avoiding the biases inherent in generic public datasets. The training process was rigorously audited, assessing its algorithmic performance with statistical measures such as Kullback-Leibler divergence (KLD) and mean squared error (MSE) to guarantee a predictive accuracy level of 95%.

Functional diagram of the architecture of the predictive audiovisual attention model.

Biometric database

The artificial intelligence is built on consumer-oriented eye-tracking data drawn from ongoing biometric research. This corpus is one of the largest high-quality visual attention repositories in the world, compiling recordings from more than 20,000 participants worldwide. The sample includes individuals exposed to a wide range of real commercial stimuli, from social media and television advertising to reading print media.

Algorithmic training (machine learning)

From the physical eye-tracking collection, the system generated empirical heatmaps. On this volume of information, nearly 200 different machine learning architectures were trained and comparatively evaluated to isolate the best-performing model. To guarantee statistical viability and purity, the data corpus was randomly split in each training iteration: one portion was assigned to direct calibration of the network and a second portion strictly to validation testing.

High-fidelity attention projection

Empirical validation and exhaustive testing show that processing a graphic asset on the platform generates a predictive map with an accuracy above 95% within seconds. The output of the algorithm is statistically equivalent to the heatmap that would be obtained by physically collecting biometric data from approximately 150 subjects exposed to the same advertising piece for 5 seconds.

Evolution of the traditional saliency map

Predictive modeling for audiovisual formats

Attention prediction in dynamic content presents specific operational challenges. The system resolves this complexity through a specialized algorithmic model for video, whose architecture implements a 2-second sliding window that processes overlapping time segments to simulate human visual dynamics.

Early algorithmic iterations had a processing error: they predicted an instantaneous shift of attention whenever a new object appeared on screen. Human ocular physiology operates with an inherent mechanical latency; visual fixation lingers on the previous point of interest for a fraction of a second after a scene cut, before the saccade toward the new stimulus is executed. The sliding-window operation was algorithmically structured to replicate these neuromotor delays, giving the model strict behavioral realism. This emergent capability is the direct result of exposing the neural network to thousands of hours of real eye-tracking recordings provided by the underlying technology, confirming the superiority of the dataset in the final predictive performance.

Initial and final attention analysis

The algorithms project the topography of visual fixation during the first and last 2 seconds of a standard 5-second advertising exposure. Each time window (opening and closing) is processed by independent artificial intelligence networks. This analytical segmentation makes it possible to audit empirically whether critical conversion vectors, such as brand identity or the call to action (CTA), capture attention instantly or require prolonged exposure to be assimilated by the viewer.

Attention dispersion: the Focus metric

The Focus score is a coefficient mathematically derived from the attention heatmap. Grounded in the empirical research on "attention dispersion" by Teixeira, Wedel and Pieters, this indicator measures the topographic distribution of visual attention across the composition of the piece.

The algorithm identifies visual concentration: a high Focus index means that eye fixation will converge on a small number of specific areas, indicating high retention and a clear hierarchy. Conversely, a low score signals that gaze is dispersed across the content, revealing a design that fragments attention and increases cognitive friction.

Detection and areas of interest (AOI)

Strategic delimitation (what are AOIs?)

While global attention patterns provide a diagnosis of visual hierarchy, advertising effectiveness and conversion depend strictly on eye fixation converging on elements of direct commercial relevance. The Databrain platform provides granular auditing tools through Areas of Interest (AOIs). These delimit specific spatial regions within the composition, such as logos, products or calls to action, and aggregate the density of attention data over those vectors. They are the mathematical basis for any creative optimization.

Predictive model for automated detection

To scale analytical operations, the platform integrates a deep learning object detection model trained exclusively on advertising morphology.

In the Databrain interface, this workflow runs with maximum efficiency: from the top toolbar, the user triggers the auto-discovery command. After a confirmation dialog (which lets the user decide whether to keep or clear previous areas), the artificial intelligence maps the visual composition and automatically classifies the detected polygons into five critical parameters: Branding, Product, Headline, Body Text and CTA.

Unlike general-purpose AI, this specialized model understands the visual language of marketing and the saturation of its commercial environments, achieving algorithmic detection performance 40% more accurate than a widely used general-purpose commercial computer vision service.

Methodological exclusion of gaze maps (gaze mapping)

Unlike other commercial tools on the market, Databrain deliberately does not deliver gaze-path scores or gaze-mapping charts. This restriction reflects a strict methodological decision: scan paths show individual variance that is too severe to generate stable or reliable predictive projections. Incorporating this type of metric would mean compromising the standards of statistical accuracy and empirical rigor that sustain the platform's viability.

Analytical quantification: attention metrics (AOI scores)

From detection to quantification

The geometric delimitation of Areas of Interest (AOI) is only the isolation phase. The platform's auditing capability lies in extracting specific predictive telemetry for each polygon drawn. The algorithm processes the density of the heatmap and translates it into structured mathematical variables, eliminating subjective interpretation of graphic design and replacing it with commercial viability data.

Matrix of attention indicators

The Databrain platform processes and consolidates three fundamental indicators to audit the visual hierarchy and pull of each advertising element:

Total Attention (Share of Attention):

Determines the proportion of visual fixation captured by the designated area relative to the entire graphic piece. This metric quantifies the topographic dominance of the stimulus, establishing its real capacity to monopolize the consumer's visual processing against competing elements within the same design ecosystem.

Time Spent (Dwell Time / Time Spent):

Measures the absolute duration of eye fixation (expressed in seconds) on the polygon. This variable is the critical determinant of cognitive assimilation. It makes it possible to audit empirically whether the projected visual latency is biologically sufficient for the brain to decode the operative message (for example, the minimum reading time for a complex headline or the semantic recognition of a product).

Percentage Seen (Visibility / Seen %):

Establishes the statistical probability that the average viewer will fixate the fovea on the specific region during the exposure cycle. This predictive indicator assesses the risk of omission: it determines mathematically which elements will go unnoticed in the flow of human visual scanning, enabling structural correction of conversion vectors before capital is allocated to media.

Validation: a commitment to scientific accuracy

The multi-layer validation protocol

To guarantee scientific rigor, an exhaustive multi-layer validation protocol has been established to audit the accuracy of the predictive models against the gold standard of laboratory eye tracking. This framework rests on statistical "pillars of truth," each of which measures an independent vector of predictive quality.

Pearson correlation coefficient (CC): the "synchronization index"

The Pearson correlation coefficient measures the linear correlation between two variables (the pixel intensity matrices of the real human map versus the map projected by the AI). It represents the ratio of their covariance to the product of their standard deviations. Its value ranges from -1 (perfect negative correlation) to +1 (perfect positive correlation), with 0 indicating no correlation.

Operationally, this metric assesses how accurately the algorithm's projections fluctuate in strict synchrony with empirical human eye movements. If real consumers concentrate their attention intensely on a logo, the AI must predict an identical concentration on that same vector. A score of 1.0 would indicate absolute synchronization. The algorithmic model reaches a correlation of approximately 95% for total viewing time, showing that its analytical cadence is virtually indistinguishable from aggregate human behavior.

Kullback-Leibler divergence (KLD): the "allocation error"

Kullback-Leibler divergence is a metric from information theory that quantifies the mathematical distance between two probability distributions: the real distribution (human) and the predictive distribution (AI).

In terms of efficiency, the viewer's attention operates as limited capital or budget (100% of their attention). The different regions of the image (the logo, the face, the product, the background) are different financial assets in which that budget can be invested. The human distribution (the truth) represents market reality: real subjects might invest 60% of their attention capital in the face, 30% in the headline and 10% in the product. The AI prediction represents the strategy of an investment fund manager who anticipates how that attention budget will be allocated.

KLD is the "allocation error." It measures the mathematical gap between the algorithmic manager's strategy and the operational reality of the market.

• If the AI projects an investment of 50% in the face and 40% in the headline, the "distance" from reality is minimal. The predictive strategy is sound.

• If the AI projects an allocation of 90% to the background (a poor asset) and only 10% to the face, the "distance" is massive. The strategy has failed to align with the market.

The technical mandate requires the KLD score to remain close to 0. A low coefficient certifies that the AI understands not only which elements consumers look at, but the exact amount of attention capital they allocate to each asset compared with empirical reality.

Similarity measure (SIM): the "overlap test"

The similarity measure, also known as histogram intersection, calculates the sum of the minimum values in each bin (or pixel) of two normalized histograms (distributions). This is an evaluation of geographic accuracy. If the empirical human attention map were printed on a blue transparency and the AI-projected map on a yellow transparency, the SIM score measures what percentage of the image turns green when they are overlaid. In essence, it quantifies the physical intersection of both maps.

While the synchronization index confirms that intensity matches, the overlap test confirms that location is accurate. By validating the model against these three complementary metrics (synchronization, allocation fidelity and geographic accuracy) using a hybrid dataset composed of open-source benchmarks and proprietary commercial assets, the artificial intelligence used by Databrain establishes a solid empirical foundation. This moves predictive analytics from an inscrutable "black box" to a transparent, verified commercial intelligence tool.

A multidimensional approach to validity

Validating a predictive AI model requires going beyond a single accuracy metric. Based on peer-reviewed research (Ramsøy, JAR 2019), the technology integrated by Databrain has established its validity across multiple dimensions:

• Ecological validity: Predictions reflect real-world viewing behavior, not only isolated laboratory conditions.

• Construct validity: The model strictly measures what it claims to measure.

• External validity: Results generalize across different contexts, advertising formats and audiences.

• Internal validity: The algorithmic logic and mathematical structure of the model are sound.

Beyond validity, sensitivity and specificity have been evaluated to confirm that the model detects the attention signal accurately and avoids false positives. Reliability has been established using Monte Carlo simulation methods, which also underpin the recommended sample sizes and confirm high consistency across different groups.

In addition, clustering analyses have been carried out, including Cronbach's alpha and principal component analysis (PCA), to identify sets of highly interrelated metrics (for example, across purchase, trial and recommendation behaviors), ensuring that the scores are meaningfully distinct rather than redundant.

Finally, the metrics have been continuously calibrated with clients over the years, grounding them in real commercial usefulness and market efficiency, not only in statistical performance.

Attention model for images

The image model is compared against the MIT/Tuebingen Saliency Benchmark. This shows the performance of the visual attention prediction for images used by Databrain against the leaders in saliency map prediction, based on MIT's open-source dataset. Notably, the AI model used by Databrain is specialized in advertising materials, which is why it achieves higher accuracy scores than the MIT benchmark, which includes natural images.

The leaderboard can be consulted here.

Attention model for video

The video model is compared against the video saliency prediction leaderboard. This shows the performance of the audiovisual attention prediction used by Databrain against the leaders in saliency map prediction, based on different open-source datasets. Notably, the AI model used by Databrain is specialized in advertising materials, which is why it achieves higher accuracy scores than the MIT benchmark, which includes natural images.

The leaderboard can be consulted here. Mixed attention model (beta)

The new attention model for mixed-media ads is in beta. At present, there are no saliency prediction research models specialized in mixed-media stimuli. The KLD coefficient has not been calculated because of the limited test dataset for this model. The technology used by Databrain will continue to collect larger volumes of test data over the coming period.

For CC and SIM, a higher value indicates closer agreement with real eye-tracking data. KLD works the other way around: it measures the distance between the human and the predicted distributions, so a lower value is better.

Emotion prediction and quantification of the unconscious

Emotions vs. feelings

Although the terms "emotions" and "feelings" are often used interchangeably, it is essential to understand the distinction between them.

Emotions are automatic, unconscious reactions in the brain and body, triggered automatically by internal or external factors. Emotional responses to stimuli arise in less than half a second, which is why they are often described as a gut reaction.

Feelings, by contrast, are the conscious experiences of those emotional reactions. They arise only once unconscious emotional responses are brought into awareness. Therefore, feelings can never arise without prior emotional reactions. Databrain's metrics focus on emotions, which are stronger predictors of actual decisions and behavior, both at the individual level and at the market level.

In depth: the Fast Response Test (FRT) methodology

The scientific engine behind the Engagement score is a proprietary Fast Response Test (FRT). The FRT is a type of implicit association test grounded in decades of research in cognitive psychology, which shows that the time it takes to make a judgment is a reliable indicator of the strength of the underlying mental association.

In a typical FRT study, a participant is briefly exposed to a creative asset. Immediately after exposure, they are presented with a series of attribute words or phrases (for example, "Attractive," "Trustworthy," "Innovative"). Their task is to answer "Yes" or "No" as quickly and accurately as possible to indicate whether the word describes their perception of the asset.

The core principle is that a faster "Yes" to a positive attribute indicates a stronger, more automatic and more deeply rooted positive association. Conversely, a slower response or a "No" suggests a weaker or nonexistent association.

FRT calculation

The methodology combines three key elements:

1. A series of binary (yes/no) questions about an asset. 2. Response Time (RT): A precise measure of response latency recorded in milliseconds. RT is recorded as a proxy for certainty. 3. Group consensus: Group consensus is calculated as a proxy for market effects.

What is Response Time (RT)?

Response Time (RT), also known as reaction time or response latency, refers to the time it takes for behavioral responses to occur during a particular task (Donders, 1969; Luce, 1991). RT generally refers to the time elapsed between the presentation of the external stimulus and the appropriate response (Posner, 1978). Critically, RT is affected by emotional and cognitive processes, which allows it to be used as an index of unconscious emotion, motivation and cognitive processing.

Data weighting

Yes/No responses are weighted by response time to assess not only whether participants say yes or no, but how certain they are of their answer. Faster responses indicate greater certainty and are up-weighted, while slower responses indicate lower certainty and are down-weighted.

Studies have shown that group coherence is extremely accurate in predicting market effects. That is why the level of agreement across all yes/no responses is also weighted.

Data weighting produces a single FRT score, which effectively predicts the impact of the ad.

For example, if 50% answer "yes" with high variance in certainty, and 50% answer "no" with comparable certainty, the "no" carries more weight.

Metric calculation and prediction

The precise calculation of FRT-based metrics is designed to go beyond simple agreement and capture the true strength of an unconscious association. The final score is a sophisticated calculation that integrates three key data points: the binary 'yes' or 'no' answer, the response time measured in milliseconds, and the consensus across the entire group of participants.

In this model, faster response times for affirmative ('yes') answers receive greater weight, since this indicates a stronger, more automatic positive association.

By normalizing these individual response times and weighting them against group consensus, the methodology produces a highly reliable and nuanced score that accurately reflects the collective gut reaction to an asset.

To build the FRT-based predictive models, the technology integrated by Databrain draws on comprehensive empirical data (ground-truth data) from FRT studies with more than 25,000 participants. By training advanced machine learning models on this vast, specific dataset, the platform can predict these implicit responses directly from a visual asset with outstanding accuracy of more than 90% compared with real human response data.

Explainable AI in action: opening the black box

A structural requirement of the Databrain platform is the complete auditability and actionability of the data. To demonstrate empirically why an asset receives a specific metric, the platform implements the algorithmic technique "Grad-CAM (Gradient-weighted Class Activation Mapping)."

Grad-CAM is an analytical visualization method that produces a heatmap to isolate and highlight the specific topographic regions of an image that were most influential in the predictive model's mathematical decision. This Explainable Artificial Intelligence ("XAI") approach is applied with methodological variations depending on the indicator being audited.

For FRT-based models, such as Engagement, Grad-CAM heatmaps show exactly which graphic vectors drive the projected implicit response. In these maps, red zones delimit high-pull areas that contribute strongly to a positive, automatic emotional response, while green zones represent areas with lower or marginal emotional impact. This methodological systematization turns a theoretical Engagement score into visual intelligence that is tangible and strictly measurable for the market.

Data weighting

The path of the emotional response

1. Stimulus

   Neural processing

   The amygdala and thalamus process the stimulus before any conscious elaboration.

2. Unconscious

   Unconscious emotions

   Bodily responses are triggered, without the intervention of will.

3. Conscious

   Conscious emotions

   Global activity in the cortex, including the ventromedial prefrontal cortex, gives rise to conscious experience.

The three emotional metrics

• Engagement

  Provides a quantitative measure of the emotional appeal of a creative piece. It is empirically grounded in FRT responses to positive emotional association words such as attractive, happy, interesting and captivating.

• Trust

  Quantifies how credible, sincere and emotionally aligned with the viewer a creative piece is perceived to be. It reflects intuitive, unconscious perceptions of honesty and authenticity, derived from fast reactions to attributes such as trustworthy, honest and sincere. It introduces an analytical standard for evaluating brand credibility, moral alignment and emotional resonance: essential vectors for building brand equity and long-term loyalty.

• Avoidance

  Audits the degree of emotional rejection, disinterest or irritation that viewers experience toward a creative asset. It is the first AI-based metric specifically designed to quantify negative sentiment from a visceral, unconscious perspective. It is derived from empirical associations with concepts such as annoying and boring, and makes visible the operational risk that an ad repels, frustrates or fails to connect with its audience.

Cognitive demand: visual complexity index

What is a cognitive response?

Cognitive responses are the mental processes triggered by stimuli, such as advertisements. Understanding these responses is crucial to commercial success, because they reveal how consumers process information, structure their thinking and, potentially, behave. Creative pieces must balance simplicity and complexity to avoid both confusion and disinterest, ensuring that they capture and hold attention.

Cognitive demand and cognitive load

The Cognitive Demand score is a quantitative measure of the visual complexity and information load of an asset. A high score indicates that the stimulus is dense, cluttered or difficult to process, which can lead to viewer confusion, abandonment of the message and poor memory encoding.

The technical foundation of this metric is rooted in the robust mathematical concept of "Shannon entropy" from information theory (Bundesen, 1999). In this context, entropy works as a proxy for the amount of information or "surprise" in an image; higher entropy corresponds to a more complex and unpredictable visual signal that requires more cognitive resources to process. This mathematical basis ensures that the score is objective, reliable and consistently reflects the cognitive effort required to analyze a visual scene.

Managing cognitive load is essential to advertising efficiency. Cognitive load refers to the mental demand placed on working memory, which can only process limited information at any given time. Overloading ads with excessive information induces cognitive overload, resulting in loss of attention, reduced comprehension and negative brand perception.

Application to video: the threshold effect ("The Doorway Effect")

How the model was selected

When evaluating multiple scoring models for cognitive demand, the one with the greatest predictive capacity for comprehension and information processing was selected.

Application to the video format

In the context of video, the cognitive demand score becomes a high-impact analytical instrument for auditing narrative and pacing. The platform calculates this metric frame by frame, revealing fluctuations in cognitive load over the full duration of the piece.

Abrupt peaks usually correspond to scene changes or hard cuts. These moments are critical because they can trigger a cognitive phenomenon known as the "doorway effect" or conceptual closure: a temporary lapse in working memory that occurs when moving between distinct contexts.

By mathematically identifying these high-risk transitions, the platform provides actionable intelligence for optimizing video pacing, ensuring that critical information, such as the appearance of the brand or a key message, is not presented immediately after an abrupt cut, where the probability of omission and forgetting is highest.

A single ad can contain several of these changes, marked as abrupt shifts in the score. The faster the sequence of cuts, the greater the risk of conceptual closure and of key information being lost.

AI memory prediction

What is memory?

Declarative memory

Non-declarative memory

Declarative memory, also called explicit memory, involves the conscious recollection of facts and events, such as remembering what one did yesterday or the meaning of the word "neuroscience." It can be measured by assessing how much and what people remember through tasks such as free recall and cued recognition.

Non-declarative memory, or implicit memory, refers to unconscious memory processes that influence our behavior and skills, such as riding a bicycle or forming habits. It cannot be expressed verbally, since it is impossible to bring it consciously into attention. Traditional surveys are ineffective for assessing this type of memory; measuring it requires observing automatic patterns or behavioral variables.

The implicit brain

How is the existence of implicit associations demonstrated? Beyond the empirical findings that correlate implicit associations with human behavior, neuroscientific studies have physically demonstrated this effect.

For example, functional magnetic resonance imaging (fMRI) has recently been used to show that brand associations unconsciously trigger the brain's memory network, including the hippocampus. In the study, greater hippocampal activity when viewing corporate mascots, such as the Michelin Man, predicted a greater number of associations with those figures in a subsequent task.

In other words, even when we view brands or derived stimuli passively, or while performing another activity at the same time, our brain automatically activates the network of associations we have consolidated with those brands.

Ad memory

Brand memory

Ad memory refers to the algorithmic capacity to predict whether consumers will remember and recognize the ads they have seen. It is a crucial indicator for auditing the effectiveness of advertising campaigns, because it determines how well a creative asset manages to leave a residual impression in the minds of the market.

Brand memory assesses consumers' projected ability to recall an ad and associate it directly with the brand behind it. This process involves correctly decoding the commercial signature behind the piece, indicating how efficiently the ad managed to anchor the brand in consumers' cognitive structure.

Ideally, every ad should be structured to achieve top-of-mind (TOM) or first-mention memory, a cognitive state in which consumers retrieve the brand immediately without external cues. This level of recall certifies that a commercial asset has dominant pull and presence in the consumer's mind.

The Memory score predicts ad recall, providing a probability score on a scale of 0 to 100 that represents the empirical probability that a group of viewers will remember a commercial asset after brief exposure (<2 seconds) and a distraction period.

Methodology for predicting advertising memory

The memory artificial intelligence developed by Neurons and used by Databrain relies on a data-driven approach derived from the "Visual Memory Game" methodology developed by MIT. The original paper, together with subsequent research using this methodology, has provided a solid scientific foundation and initial validation for the metric.

The empirical study consists of exposing participants to a rapid sequence of images. Participants must indicate the exact moment they see an image for the second time. This rigorous methodology captures both reaction time and accuracy, thereby integrating cognitive engagement factors into the Memory score. The datasets used to train the model include a wide range of advertising materials, covering diverse industry use cases: from print formats and out-of-home (OOH) advertising to digital platforms such as websites and social media.

Neurons has collected data from more than 7,000 participants from the general population. This dataset is continuously enriched to ensure that the predictive model captures the most recent human behavior and the constant evolution of trends in the advertising market.

For video assets, each frame is scored individually using the image prediction model. This algorithmic model operates under the strict constraint of not using sequential video data or audio signals.

To further ensure the robustness of the Memory metric, the projected video scores have been correlated with raw empirical data (ground-truth data) from more than 100 videos. These videos were audited for ad recall by more than 1,000 participants using a Memory Recall Test. This methodology is grounded in highly reliable empirical approaches used to measure memory functions and is adapted from proven medical methods originally developed for the diagnosis of Alzheimer's disease.

The Memory score of the AI used by Databrain shows strong mathematical alignment with the results of human testing, suggesting that a sustained level of memory engagement throughout an ad is a powerful statistical predictor of ad recall and of the overall cognitive performance of the piece. This mathematical alignment between human telemetry and algorithmic projections shows that the Memory score is a reliable proxy for auditing advertising memorability in video content.

From prediction to suggestion: Databrain AI objectives

Objectives to guide your evaluation

Setting objectives on the Databrain platform is a critical step because it determines how the creative asset is evaluated and which benchmark data it is compared against. The selected objectives (Purpose, Industry and Use Case) ensure that the algorithmic projections are compared with assets of a similar nature, enabling the extraction of meaningful AI-generated commercial intelligence. This parameterization also ensures that the Impact Score is calculated in an algorithmically fair and contextually accurate way.

Defining the purpose

Industry parameters (industry benchmarks)

Creative assets are compared across 5 main industry categories and 12 specialized subcategories, providing access to analytical intelligence tailored to your exact market. Whether the asset operates in Fast-Moving Consumer Goods (FMCG), Finance, Automotive, Telecommunications, Tourism or Healthcare, ads are not measured against generic averages: they are contrasted against the appropriate competitive set. This translates into sharper benchmarks, more relevant impact scores and empirically grounded strategic decisions.

Read more about the industry categories here.

Use case parameters (use case benchmarks)

Parametric benchmarks are provided for 15 distinct use cases covering Digital Advertising (including display formats and social platforms such as Facebook, Instagram and YouTube), Traditional Advertising (print media, out-of-home - OOH, television), Product Assets (packaging) and Websites (corporate and e-commerce environments), plus a general "Any use case" category for cross-cutting comparisons. This methodological segmentation ensures that the creative asset is evaluated against the exact context for which it was designed (not against a generic statistical average), delivering the most accurate and actionable comparative insights in the market.

• Brand building

  Focuses on increasing brand awareness and recognition in the market, not on immediate sales traction. It builds a solid corporate image through narrative content, such as educational infographics or high-engagement videos. It cultivates consumer familiarity and trust, and lays the foundation for a sustainable commercial relationship.

• Conversion

  Capitalizes on the brand value built through previous efforts. These are strictly targeted initiatives, designed to turn awareness into a measurable action: buying a product, subscribing to a newsletter or following on social platforms. They are direct campaigns with a transactional purpose, featuring visuals and copy with a sense of urgency to prompt immediate responses.

From prediction to suggestion: Databrain AI benchmarks

Selecting the right benchmark

Databrain AI benchmarks (Powered by Neurons) are the recommended performance ranges for each key metric evaluated by the platform. They indicate the score an optimized creative asset should reach on metrics such as Focus, Engagement, Memory and others, providing a standardized industry measure against which to audit the projected performance of your creative. The benchmarks derive from a constantly growing Neurons dataset of more than 12,000 images and more than 10,000 videos (with more than 100,000 distinct Areas of Interest, or AOIs), collected from various regions and across multiple industries and use cases. This extensive volume of data ensures that the benchmarks are not only representative but are continuously refined as new telemetry comes in.

The calculation process

Here is the simplified process:

1. Assets are categorized: Creative pieces are grouped by factors such as industry, use case (e.g., social, display, OOH) and visual elements such as brand presence or product placement.

2. Metric scores are aggregated: Algorithmic predictions for metrics such as Focus, Engagement and Memory are consolidated across all assets within the same category.

3. Performance ranges are created: The full score distribution is divided into five performance segments (from very low to very high) to audit how assets typically perform.

4. Recommended benchmark ranges are set: The highest-performing segment (usually the top 20%) becomes the recommended benchmark that marketing professionals should target as their efficiency goal when optimizing creative.

These benchmarks are continuously validated against empirical performance data from real campaigns and decades of neuromarketing research, ensuring that they mathematically reflect what truly drives commercial results.

Size of the reference set

• 12,000+ images

  Static advertising assets analyzed.

• 10,000+ videos

  Audiovisual pieces incorporated into the set.

• 100,000+ AOIs

  Distinct areas of interest identified and classified.

Geographic coverage

Types of areas of interest

For images, five area types are classified: brand (branding), product, call to action (CTA), headline and body text (bodytext). For video, the typology focuses on the brand.

From prediction to suggestion: Impact Score

The Impact Score is a single quantitative score from 1 to 10 that projects how effective your ad is before it is launched in the market.

It distills key neuroscientific metrics (such as attention, Engagement, cognition and Memory) into a single indicator, parametrically calibrated according to your campaign objective: branding or conversion.

Instead of processing the data manually, you get an exact coefficient that reveals your ad's impact and dictates the next operational step. All backed by more than 20 years of empirical neuroscience research.

A score for clear direction

The Impact Score works as the "pre-flight check" for your commercial asset. A score of 7.0 or higher mathematically certifies that your asset is optimized, while a low coefficient represents a structural opportunity to correct the creative piece before your media budget is spent.

Objective-based prioritization

Each Impact Score is built on a selection of prioritized metrics, tailored to the objective (brand building vs. conversion) and the format (image vs. video) of the creative piece. These metrics were selected because they consistently align with success in real-market campaigns. Each variable is included because it has either solid theoretical support or a demonstrated empirical data correlation with transactional KPIs. Metrics that did not show a consistent impact (for example, attention on generic body text) were excluded from the algorithm to ensure diagnostic accuracy and commercial relevance.

Calculation methodology

To build this comprehensive coefficient, each prioritized element of your asset is first evaluated individually and assigned its own Impact Score. Each relevant metric receives a weighting between 1 and 10, charting precisely how the piece performs against benchmark expectations for its specific industry and use case.

This calculation is carried out by contrasting the asset's raw metric results against the recommended benchmark ranges. Assets that fall within the recommended range receive a performance bonus, while those that fall below incur a proportional mathematical penalty. By weighting the exact distance from the optimal benchmark range, this dynamic adjustment system ensures high-granularity scoring, preventing assets with different levels of traction from being grouped into the same generic categories.

Once these precise individual scores are consolidated, the overall Impact Score is calculated as the linear average of all available impact scores. If the design naturally lacks a specific element, such as an ad that does not feature a physical product, the system architecture intelligently omits that metric. The asset is not penalized for this absence; the algorithm averages only the detectable metrics to provide a fair and reliable analytical reflection of the elements actively at work in the piece.

How it is calculated

1. 01

   Measure the raw result

   The raw value of the asset for the audited metric is taken (for example, Focus = 55).

2. 02

   Compare against the benchmark

   It is contrasted with the specific range for the commercial objective and the category.

3. 03

   Place it in a segment

   It is placed in a rating band: low, medium or high.

4. 04

   Adjust and convert

   A bonus or penalty is applied according to mathematical proximity to the ideal range, and the result is converted to a score from 1 to 10.

From prediction to suggestion: recommendation engine

Selecting, adopting and leveraging LLMs

The recommendation engine, built on Neurons technology, integrates via API with state-of-the-art large language models (LLMs). Neurons selected these models after a blind evaluation with 200 marketers, insights managers and designers, who rated the usefulness of the responses. Their feedback identified the models best suited to deliver actionable creative insights.

As AI recommendations continue to be developed, the aim is always to use the highest-performing models available. This ensures continuous improvement. The current models already support multiple languages and process audio.

The method for training LLMs in marketing

The LLM is grounded in the scores and benchmarks produced by Neurons technology so that its insights are contextual and aligned with the asset's industry, use case and purpose. For greater relevance, AOI-based evaluations and brand kit information are provided, helping the model understand logos, products and key messages.

The LLM analyzes the asset together with the projected heatmap, benchmarks and brand context to generate summaries and recommendations tailored to the marketing scenario. Soon, the model will also add structured insight topics, improving the organization and interpretation of these suggestions.

Using LLMs for insights and thematic analysis

The Insights page turns the predictions into clear creative feedback. Instead of only showing scores and heatmaps, the LLMs explain what is happening in the design and why it matters. The model cross-references the asset with attention data, benchmarks and visual elements. It then organizes this analysis into structured topics: brand visibility, focal points, visual hierarchy and message clarity.

Each topic highlights what works well and what needs to improve, and offers practical recommendations so that professionals can optimize their creative quickly.

How client data is processed

The model uses the client's asset and the attention heatmap to analyze and generate the recommended insights. The client retains all rights to their data, and the AI-generated intelligence is part of that data. No data is used for algorithmic training purposes.

From insight to visual solution: Databrain generative AI

Visual recommendations translate written insights and recommendations into new, concrete visual examples: mockups that show how the proposed improvements to an ad would look in practice.

They are based on four inputs: the original creative piece; the attention, engagement, cognitive demand and memory scores; the diagnosed insights and recommendations; and an automated workflow that translates those recommendations into concrete generation and curation actions.

The closed loop

What sets this process apart from a general-purpose generative tool is that the output goes back into the predictive model. Variants are not simply generated and displayed: they are generated, measured again with the same engine that evaluated the original, and only those that project an improvement survive.

1. 01

   Translation

   The process starts from the text recommendations, grounded in the predictive metrics. These recommendations are converted into a structured format that image models interpret consistently.

2. 02

   Generation

   Multiple visual versions are produced from those guidelines. Each one is automatically checked to confirm that it reflects a visible change aligned with the intended recommendation.

3. 03

   Scoring

   The remaining versions are evaluated with the predictive engine. The original objectives and areas of interest are applied and a new Impact Score is calculated. Versions that do not meet the parametric criteria are discarded.

4. 04

   Curation

   From the set of versions that do demonstrate a projected improvement, a selection of the strongest and most visually distinct examples is shown.

Each run produces a curated set of examples that shortens the distance between diagnosis and decision: the professional does not receive a written suggestion to interpret, but concrete alternatives with their projected impact already calculated.

Scientific references

• Bagdziunaite, D., Nasan, K., Clement, J. & Ramsøy, T.Z. (2014). An added value of neuroscientific tools to understand consumers in-store behaviour. European Marketing Academy Conference, June 8.

• Bundesen, C. (1999). A new look at functions of attention. Visual Cognition, 6, 89-91.

• Bundesen, C., Habekost, T. & Kyllingsbæk, S. (2005). A neural theory of visual attention: bridging cognition and neurophysiology. Psychological Review, 112(2), 291-328.

• Christensen, M. S., Ramsøy, T. Z., Lund, T. E., Madsen, K. H., & Rowe, J. B. (2006). An fMRI study of the neural correlates of graded visual perception. NeuroImage, 31(4), 1711-25.

• Harel, J., Koch, C., & Perona, P. (2006). Graph-based visual saliency. Proceedings of Neural Information Processing Systems (NIPS).

• Hou, X., Harel, J., & Koch, C. (2011). Image signature: Highlighting sparse salient regions. IEEE Transactions on Pattern Analysis and Machine Intelligence.

• Itti, L. & Koch, C. (2000). A saliency-based search mechanism for overt and covert shifts of visual attention. Vision Research, 40(10-12), 1489-1506.

• Kyllingsbæk, S., Valla, C., Vanrie, J. & Bundesen, C. (2007). Effects of spatial separation between stimuli in whole report from brief visual displays. Perception and Psychophysics, 69(6), 1040-1050.

• Kyllingsbæk, S., Markussen, B. & Bundesen, C. (2012). Testing a poisson counter model for visual identification of briefly presented, mutually confusable single stimuli in pure accuracy tasks. Journal of Experimental Psychology: Human Perception and Performance, 38(3), 628-642.

• Milosavljevic, M., & Cerf, M. (2008). First attention then intention. International Journal of Advertising, 27(3), 381-398.

• Milosavljevic, M., Navalpakkam, V., Koch, C., & Rangel, A. (2012). Relative visual saliency differences induce sizable bias in consumer choice. Journal of Consumer Psychology, 22(1), 67-74.

• Plassmann, H., Ramsøy, T. Z., & Milosavljevic, M. (2012). Branding the brain: A critical review and outlook. Journal of Consumer Psychology, 22(1), 18-36.

• Ramsøy, T. Z. (2015). Introduction to neuromarketing & consumer neuroscience. Neurons Inc Publisher.

• Ramsøy, T. Z. (2019). Building a Foundation for Neuromarketing And Consumer Neuroscience Research. Journal of Advertising Research, 59(3), 281-294.

• Ramsøy, T. Z., Jacobsen, C., Friis-Olivarius, M., Bagdziunaite, D., & Skov, M. (2017). Predictive value of body posture and pupil dilation in assessing consumer preference and choice. Journal of Neuroscience, Psychology, and Economics, 10(2-3), 95-110.

• Ramsøy, T. Z., & Overgaard, M. (2004). Introspection and subliminal perception. Phenomenology and the Cognitive Sciences, 3, 1-23.

• Ramsøy, T. Z., & Skov, M. (2014). Brand preference affects the threshold for perceptual awareness. Journal of Consumer Behaviour, 13(1), 1-8.

• Trabulsi, J., Norouzi, K., Suurmets, S., Storm, M., Nieuwenhuis, I., & Ramsøy, T.Z. (2020). Optimal gaze filters for eye-tracking on small screens. Frontiers in Neuroscience.